{"id": "showcase/01-convenience-store-night-scene", "group": "showcase", "slug": "01-convenience-store-night-scene", "title": "Convenience Store Night Scene", "prompt_words": 145, "prompt": "Create an ultra-realistic urban street group photo at a convenience store entrance at 10 PM summer night. 3-4 young people briefly chatting at the entrance, someone holding drinks, someone sitting on plastic outdoor chairs, someone standing looking at their phone. Bright white light streaming through the glass doors and windows, warm yellow street lights and distant car headlights outside. Characters wearing everyday clothes: T-shirts, shirts, shorts, jeans, sneakers. No internet celebrity styling. Faces and postures must look like real pedestrians, not overly polished. Environment must include real convenience store elements: freezer stickers, promotional posters, trash cans, entrance mats, glass reflections, shared bikes on roadside, water droplets from drink bottles on ground. The image should look like a very authentic life slice captured by a photographer in the city. Focus on testing natural multi-person interactions, night convenience store lighting, glass reflections, and ordinary people's vibe restoration.", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/01-convenience-store-night-scene"}
{"id": "showcase/02-raw-iphone-subway-station", "group": "showcase", "slug": "02-raw-iphone-subway-station", "title": "Raw Iphone Subway Station", "prompt_words": 38, "prompt": "Create a completely RAW quality, unprocessed, unedited image with full iPhone camera quality. A subway station in USA, a momentary blur. The subway is in motion. In front of the subway, there is an elderly woman and man.", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/02-raw-iphone-subway-station"}
{"id": "showcase/03-rice-grain-micro-typography", "group": "showcase", "slug": "03-rice-grain-micro-typography", "title": "Rice Grain Micro Typography", "prompt_words": 19, "prompt": "A massive pile of rice, and on one single grain of rice there is tiny text that reads \"wOw\"", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/03-rice-grain-micro-typography"}
{"id": "showcase/04-360-equirectangular-panorama", "group": "showcase", "slug": "04-360-equirectangular-panorama", "title": "360 Equirectangular Panorama", "prompt_words": 17, "prompt": "360 equirectangular image of a slot canyon in Arizona at noon, sunbeams falling through the narrow opening", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/04-360-equirectangular-panorama"}
{"id": "showcase/05-90s-point-and-shoot", "group": "showcase", "slug": "05-90s-point-and-shoot", "title": "90S Point And Shoot", "prompt_words": 5, "prompt": "90s + point-and-shoot camera quality", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/05-90s-point-and-shoot"}
{"id": "showcase/06-gta6-style-in-game-footage", "group": "showcase", "slug": "06-gta6-style-in-game-footage", "title": "Gta6 Style In Game Footage", "prompt_words": 47, "prompt": "GTA 6 in-game footage, very detailed, very realistic. Close-up shot taken from a stationary 4k monitor. (There's a slight blurriness in the image, as it feels like it was taken handheld). A wide, bright environment. Realistic details. The character is walking on the beach with / :dog.", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/06-gta6-style-in-game-footage"}
{"id": "showcase/07-100-pixel-art-items-grid", "group": "showcase", "slug": "07-100-pixel-art-items-grid", "title": "100 Pixel Art Items Grid", "prompt_words": 19, "prompt": "Create a single image containing a grid of 100 completely unique pixel art items, each with a meaningful label", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/07-100-pixel-art-items-grid"}
{"id": "showcase/08-live-streaming-screenshot", "group": "showcase", "slug": "08-live-streaming-screenshot", "title": "Live Streaming Screenshot", "prompt_words": 12, "prompt": "Generate a screenshot of a TikTok live stream featuring a woman streaming.", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/08-live-streaming-screenshot"}
{"id": "showcase/09-song-dynasty-social-feed", "group": "showcase", "slug": "09-song-dynasty-social-feed", "title": "Song Dynasty Social Feed", "prompt_words": 152, "prompt": "\"Song Dynasty People's Moments\"/\"SONG DYNASTY SOCIAL MEDIA FEED\", Ancient and modern time-travel humor fusion interface design style, The image simulates a mobile phone social media interface, but the content is entirely Song Dynasty scenes, The avatar is a portrait of a Song Dynasty literati, Username \"Su Dongpo SuShi_Official\", Post content \"Just arrived in Huangzhou, demoted but feeling okay. Made Dongpo pork myself today, tastes amazing, recipe attached:\", The attached image is a close-up of Dongpo pork in Gongbi painting style, Likes list \"Huang Tingjian, Qin Guan, Fo Yin etc. 126 people\", Comments section \"Wang Anshi: Hehe\" \"Sima Guang: Still the same taste\", Interface elements such as the like icon are replaced with Song Dynasty patterns, The status bar shows \"Great Song Mobile 5G\" and \"Third Year of Yuanfeng\", The color scheme is mobile phone dark mode paired with elegant Song Dynasty tones, A masterpiece of fun collision between history and social media", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/09-song-dynasty-social-feed"}
{"id": "showcase/10-wechat-chat-smartphone-photo", "group": "showcase", "slug": "10-wechat-chat-smartphone-photo", "title": "Wechat Chat Smartphone Photo", "prompt_words": 107, "prompt": "Generate an image, aspect ratio 3:4. A real-life smartphone photo, with the phone screen displaying a WeChat chat interface in Chinese dialogue, featuring chat bubbles and a Word document attachment. The interface has green and gray bubbles. The screen shows obvious fingerprints, smudges, and scratches. The glass has strong reflections and a direct light source creating glare. The frame is slightly tilted, handheld shot, natural ambient light, imperfect composition, strong sense of realism, casual everyday snapshot, high detail, 4K. The dialogue is between a boss and an employee: the employee sends a file to the boss, and the boss replies that they will take a look first.", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/10-wechat-chat-smartphone-photo"}
{"id": "showcase/11-pet-brand-collaboration", "group": "showcase", "slug": "11-pet-brand-collaboration", "title": "Pet Brand Collaboration", "prompt_words": 98, "prompt": "Theme: \"77 (cat's name) X KFC\" collaboration poster. Generate a collaboration poster featuring the same pet (absolutely consistent in appearance and coloring) with KFC brand identity (red-and-white color scheme, classic logo, restaurant scenes, etc.). Have the cat wear a KFC employee uniform, KFC employee hat, and name badge while standing at the counter, selling fried chicken, burgers, and combo meals, and interacting with chicken buckets, fries, soda, and other brand elements. The style should be lively and fun with a commercial co-branding feel, suitable for online promotion and event posters. Add appropriate Chinese copy freely for the poster.", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/11-pet-brand-collaboration"}
{"id": "showcase/12-world-time-clock-wall", "group": "showcase", "slug": "12-world-time-clock-wall", "title": "World Time Clock Wall", "prompt_words": 50, "prompt": "It's 10 AM in Los Angeles, 11 AM in Denver, 12 PM in Chicago, 1 PM in New York, 6 PM in London, and 2 AM in Tokyo. Render a wall with different analog clocks, each showing the correct time for its city, with the city label below each clock.", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/12-world-time-clock-wall"}
{"id": "showcase/13-silhouette-universe-poster", "group": "showcase", "slug": "13-silhouette-universe-poster", "title": "Silhouette Universe Poster", "prompt_words": 494, "prompt": "Automatically generate a high-aesthetic \"Silhouette Universe / Collector's Edition Narrative Poster\" based on [theme: the last lighthouse keeper]. Do not default to common containers such as bottles, hourglasses, glass domes, or pocket watches. Instead, let the AI choose the most symbolic, visually strong, and narratively suitable outer silhouette for the theme. The silhouette can be an artifact, building, gate, tower, archway, dome, stairwell, corridor, statue, profile, eye, hand, skull, wing, mask, mirror, throne, ring, crack, light curtain, shadow, geometric form, spatial cross-section, stage frame, abstract symbol, or any more creative contour that best represents the theme. The goal is not to simply place a world inside an object, but to make the entire themed universe grow naturally within, around, and through the silhouette itself. The silhouette should be elegant, recognizable, and compositionally dominant. Inside or along its boundary, generate a rich and layered narrative world strongly tied to the theme, including iconic scenes, key architecture or spaces, symbols and metaphors, traces of characters or civilizations, foreground-midground-background depth, and atmosphere with emotional tension. Elements such as doors, staircases, bridges, water, smoke, paths, light sources, ruins, mechanical structures, landscapes, abstract forms, creatures, or props should all feel unified and organically integrated rather than pasted together. The final composition should feel like a premium collector's poster with strong design value. The large shapes should feel stable, the main silhouette should be unmistakable, and the inner world should have depth, structure, and breathing room. Details should be rich but not crowded. You may add small human silhouettes, distant buildings, beams of light, doorways, bridges, steps, corridors, reflections, skylight, or far-background structures to enhance scale, story, and epic atmosphere. The overall mood should be quiet, grand, refined, and lingering rather than noisy or overloaded. Blend the feeling of a collector's edition film poster, high-end narrative visual design, dreamy watercolor texture, and fine printed paper. Emphasize paper grain, feathered edges, watercolor brush marks, gentle diffusion, atmospheric perspective, soft haze, selective volumetric light, light passing through mist, generous negative space, and restrained layout. The image should feel premium, poetic, majestic, sacred, nostalgic, quiet, and mythic. Let the AI choose an elevated color palette based on the theme, but keep it unified, restrained, tasteful, low-saturation, and premium. Avoid chaotic high saturation, cheap neon effects, or plastic digital color. Suitable palette families may include black-gold-gray, cool blue-gray, mist white-gray, brown-red with off-white, dark copper, aged paper tones, deep sea blue, twilight purple, or silver-gray, as long as they serve the theme. Final requirement: the first glance should give strong theme recognition and a memorable silhouette, the second glance should reveal a complete narrative world, and the third glance should still reward close inspection with depth and aftertaste. Avoid generic backgrounds, hard cut-and-paste compositions, templated fantasy assets, video-game promo art vibes, excessive cartooning, or realism that kills the artistic mood. If appropriate, subtle titles, numbering, signatures, or marks can be added as part of the poster design, but they must never overpower the image.", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/13-silhouette-universe-poster"}
{"id": "showcase/14-celebrity-in-real-life", "group": "showcase", "slug": "14-celebrity-in-real-life", "title": "Celebrity In Real Life", "prompt_words": 16, "prompt": "Sam Altman, Donald Trump, and Elon Musk working behind the counter of a busy movie theater", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/14-celebrity-in-real-life"}
{"id": "showcase/15-apple-park-keynote-crowd", "group": "showcase", "slug": "15-apple-park-keynote-crowd", "title": "Apple Park Keynote Crowd", "prompt_words": 23, "prompt": "Amateur iPhone photo at Apple Park during the iPhone 20 keynote, Tim Cook presenting on stage. Shot from the crowd at a distance", "images": ["full.jpg", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/15-apple-park-keynote-crowd"}
{"id": "showcase/16-direct-flash-night-portrait", "group": "showcase", "slug": "16-direct-flash-night-portrait", "title": "Direct Flash Night Portrait", "prompt_words": 117, "prompt": "35mm color film photography with harsh direct on-camera flash, specular highlights on skin and clothing, strong catchlights in eyes, high contrast flash illumination, authentic film grain and color shift, night street editorial style, close-up portrait, a woman in a black leather jacket leaning back against a shuttered shopfront at 1 AM, one hand pushing her hair back, chin slightly down, gaze just past the lens, hard shadow thrown straight behind her onto the wall, background falling away to black, wet asphalt at the bottom of the frame catching the flash, slight cyan drift in the shadows, no plastic skin, no digital over-sharpening, no airbrushing, no oily skin, no watermark, no text, authentic 35mm direct flash film look", "images": ["full-flare.jpg", "full-sunburst.jpg", "preview-flare.jpg", "preview-sunburst.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/16-direct-flash-night-portrait"}
{"id": "showcase/17-night-market-queue", "group": "showcase", "slug": "17-night-market-queue", "title": "Night Market Queue", "prompt_words": 81, "prompt": "Authentic night market food stall photograph at 11 PM, a queue of people along the front, steam rising off the griddle into bare bulbs strung overhead, the vendor working fast with both hands, trays of ingredients stacked around them, mixed light from hot bulbs above and a cold blue spill from the neighbouring stall, wet ground, worn scratched counter, real used cooking equipment, shallow depth of field, high ISO noise, no staging, no text, no signage, no brand names, no watermark", "images": ["full-flare.jpg", "full-sunburst.jpg", "preview-flare.jpg", "preview-sunburst.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/17-night-market-queue"}
{"id": "showcase/18-golden-hour-backlight", "group": "showcase", "slug": "18-golden-hour-backlight", "title": "Golden Hour Backlight", "prompt_words": 72, "prompt": "85mm portrait photograph backlit at golden hour with the sun directly behind the subject's head, hair glowing as a bright halo, face in soft open shade with bounced fill only, strong natural lens flare streaking across the frame, dust in the air catching the light, long shadows running toward the camera, real skin texture with visible pores and fine lines, no airbrushing, no plastic skin, no digital over-sharpening, no watermark, no text", "images": ["full-flare.jpg", "full-sunburst.jpg", "preview-flare.jpg", "preview-sunburst.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/18-golden-hour-backlight"}
{"id": "showcase/19-retro-racing-replay", "group": "showcase", "slug": "19-retro-racing-replay", "title": "Retro Racing Replay", "prompt_words": 109, "prompt": "A gameplay screenshot of a 1990s arcade racing game replay screen, 16:9. Low-polygon flat-shaded cars on a seaside coastal road seen from a chase camera, a low-poly sea to one side, dithered gradients on the horizon, visible polygon edges and low-resolution textures as a period console would render them. The HUD is fully rendered over the gameplay: a position indicator reading \"3/8\", a lap counter reading \"LAP 2/5\", a timer reading \"0'47\\\"12\", a best-lap reading \"0'45\\\"88\", and a speed readout reading \"241 km/h\" with a red-line bar. Chunky pixel font, all caps. Render only those five strings. No other text, no watermark, no real game logo or track name.", "images": ["full-flare.jpg", "full-sunburst.jpg", "preview-flare.jpg", "preview-sunburst.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/19-retro-racing-replay"}
{"id": "showcase/20-isometric-dungeon-key-art", "group": "showcase", "slug": "20-isometric-dungeon-key-art", "title": "Isometric Dungeon Key Art", "prompt_words": 131, "prompt": "A true isometric cutaway illustration of a dungeon three levels deep, seen at exactly 30 degrees with no perspective distortion, game key art style. The top level is a cold stone entrance open to the night and rain. The middle level holds a flooded cistern, a collapsed bridge and a rusted portcullis. The bottom level is a warm torch-lit vault with a stone altar and barred cells. A single hero carrying a lamp is on the spiral staircase between the middle and bottom levels. The cut walls show their thickness and the earth beside them, with roots and rubble. Warm torchlight inside against cold blue night above, volumetric light shafts through the entrance. One hero only. No text, no UI, no HUD, no watermark, no perspective, grid locked to 30 degrees.", "images": ["full-flare.jpg", "full-sunburst.jpg", "preview-flare.jpg", "preview-sunburst.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/showcase/20-isometric-dungeon-key-art"}
{"id": "cases/01-citrus-under-pressure", "group": "cases", "slug": "01-citrus-under-pressure", "title": "Citrus Under Pressure", "prompt_words": 479, "prompt": "A bright beverage advertisement photograph of a chilled can of orange sparkling soda, standing at a slight\ntilt on a smooth pale surface, shot on a 3:4 vertical canvas at close product distance.\n\nCOMPOSITION\nThe can is the largest and most central element, occupying about 55% of the frame height, its base resting\njust below the lower third. A single thin orange slice curls behind the can's upper third, its circle echoing\nthe can's rim. One narrow arc of liquid rises from the surface behind the can and sweeps around its right\nside at the can's mid-height, then falls out of frame. The arc passes BEHIND the can and never crosses the\nlabel. The composition is built on three circles — the slice, the rim, the liquid arc — held on one vertical\naxis. Empty space falls to the lower left. Nothing touches the frame edges.\n\nSETTING AND LIGHT\nBackground: a seamless warm cream backdrop with a soft apricot falloff toward the lower right, no props, no\nhorizon. Lighting: one large soft source behind the can at the upper left, acting as a rim light that passes\nthrough the orange slice and glows through the liquid arc; one broad white bounce card in front at low\nintensity to keep the label legible. Contact shadow: one tight soft shadow directly beneath the can base.\nNo coloured gels, no second light, no reflections of studio equipment.\n\nMATERIALS\nCold aluminium can with a satin brushed finish and a fine bead-blasted sheen along the rim; heavy condensation\nacross the cold metal, with droplets varying from fine mist to a few larger runs, all catching the back light.\nOrange slice: translucent flesh with visible radial segments and a pith ring, the back light glowing through\nit so the pulp reads as wet rather than as coloured glass. Liquid: thin, glossy, opaque orange with a defined\nmeniscus, heavy enough to hold a smooth curve, catching a bright highlight line along its top edge.\n\nLABEL AND TEXT\nThe label is white with a matte paper texture, wrapped cleanly and registered straight to the can's vertical\naxis. It carries exactly this text, once each: the brand word \"ZEST\" in a bold condensed sans on the upper\nthird, \"ORANGE\" beneath it in a smaller light sans, and \"330 ML\" at the bottom of the label. No other words,\nnumbers, marks or symbols appear anywhere in the frame. Brand text must be crisp, correctly kerned, unobstructed\nand readable at thumbnail size.\n\nCONSTRAINTS\nOne liquid arc only. No scattered fruit pieces, no splashes, no droplets scattered through the air. The can's\nsilhouette, the label's proportions and the label's registration must stay unchanged. Do not add a second can,\na glass, ice cubes, leaves or a straw. No invented logos, no badges, no awards, no \"new\" or \"best\" flashes,\nno watermark. Avoid plastic-looking CGI, avoid flat airless lighting, avoid over-sharpened edges.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/01-citrus-under-pressure"}
{"id": "cases/02-after-hours-jazz-poster", "group": "cases", "slug": "02-after-hours-jazz-poster", "title": "After Hours Jazz Poster", "prompt_words": 374, "prompt": "A vertical editorial concert poster, 2:3, printed artwork seen flat and straight on — not a photograph of a\nposter on a wall.\n\nCOMPOSITION\nDeep indigo ground with a subtle ink-darkening toward the outer edges. A single narrow shaft of warm gold\nstage light enters from the upper left and falls through the centre of the sheet, and the illuminated volume\nof that beam resolves into the abstract silhouette of a saxophone: bell low right, neck rising left, keys\nsuggested by a row of small dark interruptions along the beam's right edge. The silhouette is built only from\nlight and its absence — no outline, no drawn instrument. The headline sits above the beam, set large; the\nsupporting lines sit below it in a compact block with clear air around them; the date and venue occupy their\nown zone at the bottom margin. The beam never crosses any letterform.\n\nTEXT — render exactly these four strings, each appearing once\n\"AFTER HOURS\"\n\"JAZZ WEEKENDER\"\n\"18–20 SEPTEMBER\"\n\"RIVERSIDE HALL\"\nHierarchy: \"AFTER HOURS\" is the largest, set in a high-contrast condensed serif, tracked tight, on two lines\nif needed. \"JAZZ WEEKENDER\" is roughly one third that height in a neutral uppercase sans with wide letter\nspacing. \"18–20 SEPTEMBER\" and \"RIVERSIDE HALL\" are the smallest, set on one shared left-aligned column in the\nbottom margin, separated by a thin rule. All type sits on one left axis. No other text anywhere.\n\nPAPER AND FINISH\nHeavy uncoated poster stock with visible fibre grain and a faint deckled suggestion along the edges; ink sits\nslightly into the paper rather than on top of it. The gold light is printed ink luminance, not a glow effect:\nflat in the deep areas, softly graduated only where the beam passes.\n\nCONSTRAINTS\nNo QR code, no barcode, no sponsor logos, no decorative microtext, no fictional support acts, no ticket\nprices, no \"sold out\" flashes. Do not paraphrase, abbreviate, re-case or hyphenate the four strings. Do not\nrotate, outline, shadow or gradient-fill the type. Do not add a saxophone photograph, a player, a crowd or a\nvenue illustration. Keep the poster plane flat: no frame, no perspective, no wall shadow, no mockup, no tape,\nno torn edges beyond a faint deckle. No watermark, no signature.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/02-after-hours-jazz-poster"}
{"id": "cases/03-midnight-bathhouse", "group": "cases", "slug": "03-midnight-bathhouse", "title": "Midnight Bathhouse", "prompt_words": 400, "prompt": "A three-storey cutaway illustration of a small mountain bathhouse at night, drawn straight on as a section\nthrough the building, on a 3:2 landscape canvas.\n\nTHE BUILDING\nA compact timber inn, roughly three times as wide as it is tall, its left half cut away so all three floors\nare visible at once. Outside: deep blue snow night, a low driftscape against the base, bare larch trunks\nbehind, falling snow. Inside: warm amber lamplight on timber, paper screens, and steam, clearly separated from\nthe cold exterior. Floor slabs read as solid timber sections; the roof is snow-loaded with a visible eave.\nOne staircase connects all three floors in a legible run, and its landings align with the floor slabs.\n\nLOWER FLOOR — the bath\nA sunken stone pool with faint steam rising off the water surface only, not through the walls. Wet stone\napron, a low wooden bucket, a folded towel on a rail. The water line is visible against the pool wall.\n\nMIDDLE FLOOR — the tea room\nLow table with a teapot mid-pour, two cups, a paper lantern, a folded blanket, a window onto the snow. A paper\nscreen partially drawn so the room behind it stays readable.\n\nUPPER FLOOR — the guest room\nTwo futons, a small chest, a hanging lamp, a pair of boots set beside the door, snow built up on the window sill.\n\nFIGURES\nFive small figures, each doing something different and quiet: one stepping into the bath, one seated pouring\ntea, one asleep under a futon, one standing at an upper window looking out, one on the stairs carrying a\ntray. Figures are small relative to the building, anatomically plain, and never repeated as identical stamps.\n\nMATERIAL AND LIGHT\nTimber: warm grain, visible joinery, worn edges at thresholds. Paper screens: translucent, glowing from the\nlamps behind them. Stone: cold grey-green with damp sheen near the pool. Snow: matte, with soft blue shadow.\nOutside light is cool and low-contrast; inside light is warm and pooled around each lamp. The two light\nsources never mix — snow stays blue, interiors stay amber.\n\nCONSTRAINTS\nWater and steam never pass through a wall, floor or the roof. Exterior falling snow never appears indoors.\nEach floor's boundary, door and stair connection must be readable. Do not fill rooms with clutter; the space\nmust stay explorable along the stairs. Do not add text, signage, numbers or logos. No watermark.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/03-midnight-bathhouse"}
{"id": "cases/04-miniature-crew-on-a-lens", "group": "cases", "slug": "04-miniature-crew-on-a-lens", "title": "Miniature Crew On A Lens", "prompt_words": 456, "prompt": "A macro photograph of a miniature construction site built on top of a real camera lens lying on a dark\nwooden desk, shot at 3:2 with a tilt-shift effect so the plane of focus runs diagonally across the lens barrel.\n\nTHE LENS AS LANDSCAPE\nA large prime lens lies on its side, filling most of the frame: black anodised aluminium barrel with fine\nknurled focus and aperture rings, a metal mount ring at the near end, engraved white numerals along the barrel,\nand the front glass element facing the camera at the far end. The glass is deep blue-violet with concentric\ncoating rings and a curved reflection of the room's ceiling. The lens is the terrain — everything else is built\non it.\n\nTHE SITE\nA miniature crew is working along the top of the barrel. On the rubber ring: a climbing scaffold of thin\naluminium poles lashed with small orange ties. Mid-barrel: a lattice crane with a counterweight, its hook\nlowered toward the glass. Near the mount: a striped barrier line of cones and tape around a \"work area\", with\na tiny open toolbox, a coiled cable, a folded tarp and two traffic cones. On the front glass element: a\nsingle worker standing well out on the glass, casting a reflection beneath himself.\n\nFIGURES\nSeven figures in high-visibility yellow vests and white helmets. Each is doing something different: one\nclimbing the scaffold, one walking a beam with a tether, one at the crane controls, one kneeling to check a\nlevel, one carrying a plank over the shoulder, one standing on the glass, one sitting on the barrel edge with\na clipboard. All are anatomically plain, correctly scaled to each other, and never repeated as identical stamps.\n\nLIGHT AND DEPTH\nOne warm desk lamp off-frame to the upper left, its shade just visible as a soft amber glow at the frame's\nedge, throwing warm directional light along the barrel and long soft shadows to the right. A cold blue rim\nfrom a window behind separates the figures from the background. Strong tilt-shift depth of field: the strip\nacross the barrel top is sharp, the near knurling and the far background fall into soft focus. Background\nbokeh of warm amber point lights and one blue highlight.\n\nCONSTRAINTS\nThe lens must remain a real, correctly proportioned camera lens: barrel diameter, ring widths, mount ring,\nengraved numerals and the glass element all plausible and consistent. No text anywhere except the lens's own\nengraved markings, which are embossed metal and unreadable at this scale. No logos, no brand marks, no\nwatermark. Do not make the scene look like plastic toys: real metal, real rubber, real glass, real tiny\nclothing with fabric weave. No flame, no welding sparks, no dust cloud.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/04-miniature-crew-on-a-lens"}
{"id": "cases/05-exploded-ramen-bowl", "group": "cases", "slug": "05-exploded-ramen-bowl", "title": "Exploded Ramen Bowl", "prompt_words": 528, "prompt": "A commercial food photograph of an exploded ramen bowl, drawn on a 4:3 canvas as a vertical assembly diagram\nin three dimensions: the finished bowl at the bottom, and its components suspended above it in the order they\nare assembled, each labelled.\n\nCOMPOSITION\nDark slate ground filling the frame, one value, with a soft pool of light in the centre. The ceramic bowl sits\nin the lower third, centred, seen from a slightly raised three-quarter angle so its interior is visible. Directly\nabove the bowl, seven components float in a vertical stack with generous even gaps, largest at the bottom,\nsmallest at the top. Each component has one thin straight leader line running horizontally out to a label at\nthe left or right margin; no leader crosses another, no leader crosses a component, and every leader ends exactly\non the component it names.\n\nTHE COMPONENTS, bottom to top\n1. Amber broth held as a perfect disc of liquid suspended in mid-air, with a defined meniscus and a bright\n   highlight along its edge; faint steam rising from this disc only.\n2. A coil of fresh noodles, lifted and separated so individual strands are visible, glossy and pale gold.\n3. Three slices of braised pork belly, edge-seared, with visible fat-to-meat layering and a caramelised rim.\n4. Two halves of a marinated soft-boiled egg, yolks jammy and slightly translucent, cut faces toward camera.\n5. A sheet of toasted nori standing almost upright, matte black-green with visible fibre structure.\n6. A small fan of sliced scallions, fresh green and white rings, crisp edges.\n7. A tight cluster of chili oil droplets suspended together, each catching a small bright specular point.\n\nTEXT — render exactly these seven strings, one per label, each appearing once\n\"BROTH\" / \"NOODLES\" / \"CHASHU\" / \"SOFT EGG\" / \"NORI\" / \"SCALLION\" / \"CHILI OIL\"\nTypography: one neutral uppercase sans, one size only, set horizontally, left-aligned on the left margin for\ncomponents 1, 3, 5, 7 and right-aligned on the right margin for components 2, 4, 6, so the leaders stay short.\nNo other text anywhere in the frame.\n\nLIGHT AND MATERIAL\nOne large soft key from the upper left, one weak fill from the right, and one narrow back light that makes the\nbroth glow from within and catches the rim of the egg yolks. Ceramic: matte glaze with fine crazing and a soft\nrim highlight. Liquid: glossy, opaque amber with a viscous meniscus. Noodles: wet and reflective without looking\noiled plastic. Pork: crisp caramelised edge, fibrous interior. Egg: dense jammy yolk with a matte albumen.\nSteam rises only from the broth disc, thin and directional, never covering a component or a label.\n\nCONSTRAINTS\nExactly seven components and exactly seven labels, one each, in the order given. Components must not overlap,\ntouch, or drift off the vertical axis. Do not add chopsticks, a spoon, a napkin, a second bowl, scattered\ningredients, flying droplets, herbs or decorative garnishes. Do not add brackets, arrows, dashed boxes,\nmeasurement lines or dimension marks — the only graphics in the frame are the seven thin leaders. Do not\nrestyle the labels, do not add units or prices. No watermark, no logo, no border.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/05-exploded-ramen-bowl"}
{"id": "cases/06-anamorphic-night-crossing", "group": "cases", "slug": "06-anamorphic-night-crossing", "title": "Anamorphic Night Crossing", "prompt_words": 457, "prompt": "An anamorphic night exterior, 21:9, shot on a rain-slicked city street minutes after the rain stopped — a\ncinematic still, not a tourist photograph.\n\nCOMPOSITION\nWide frame, camera at chest height on the near pavement, the road running diagonally away to the right so the\nreflections carry the eye into depth. A single cyclist crosses left to right through the middle third, small in\nframe, moving out of the left of the shot, caught mid-pedal; a pedestrian with a dark umbrella walks away from\ncamera on the far pavement, well down the street. Foreground: the wet kerb and the out-of-focus edge of a parked\nvan intruding into the lower right corner, cropped by the frame. Empty wet road occupies the lower half. Nothing\nis centred.\n\nTHE STREET\nNamed elements, all present: wet asphalt, painted lane markings, standing puddles holding reflections, a bento\nshop window with a paper lantern, a vending machine glowing on the near wall, a bento shop sign with unreadable\nlettering, a traffic signal showing amber, overhead power lines crossing the sky, a metal drain grate, a\nsteam vent from a basement grate, wet bicycle tyres, a low bollard, a shuttered roller door, a stack of plastic\ncrates beside a doorway.\n\nLIGHT\nNight, blue hour, no sun. Sources: the amber lantern of the bento shop at the left, at eye level; the cold white\nof the vending machine on the near right; the amber traffic signal high in the frame; and a weak blue-grey\nambient sky. One dominant direction: the warm shop light comes from the left and slightly behind, so the cyclist\nis rimmed and the road reflections stretch toward camera. Every shadow and every reflection must agree with the\nwarm source being on the left.\n\nLENS CHARACTER\nTrue anamorphic rendering: one long horizontal blue lens flare streak across the frame at eye level, sourced\nfrom the vending machine; oval bokeh in the background lights, elongated horizontally; a slight horizontal\nsmear at the frame edges; fine film grain; a very shallow depth of field that keeps the puddles and road\nreflections sharp in the middle distance and softens the far street.\n\nCONSTRAINTS\nNo readable text anywhere: all signage, the vending machine front and the shop lantern must stay illegible —\noff-angle, cropped, or out of focus — and must never resolve into invented words. No legible vehicle plates.\nNo people facing the camera, nobody posing, no crowd. No lens flare except the one horizontal streak. No\nwatermark. Do not light the scene with a fill: keep the contrast deep and the shadows held. Do not add neon\nsignage in colours that contradict the amber-left / cold-right lighting logic. No rain falling, no spray from\ntyres — the street is wet, the weather has stopped.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/06-anamorphic-night-crossing"}
{"id": "cases/07-six-panel-manga-spread", "group": "cases", "slug": "07-six-panel-manga-spread", "title": "Six Panel Manga Spread", "prompt_words": 511, "prompt": "A six-panel manga spread, 3:2 landscape, black and white, telling one continuous action with no dialogue\nand no text of any kind. Two columns of six equal panels: three columns across, two rows down, with even\ngutters. Reading order is left to right, then top to bottom.\n\nSTYLE\nTraditional ink manga: confident black linework with variable weight, screentone for mid-tones and shadow,\ncross-hatching for the darkest areas, and large untouched white areas for light. High contrast, clean panels,\nno grey wash. Screentone dots stay legible at print size. Nothing is coloured.\n\nTHE ACTION — one girl pursuing a cat through a narrow alley and up to a roof, six beats\nPanel 1 — wide establishing shot: a narrow residential alley at dusk, deep in frame, laundry lines overhead,\n  a stack of crates against a wall; the girl enters from the left edge, small, mid-stride.\nPanel 2 — extreme close-up on the ground: her feet and the cat's paws over a puddle, water displaced, motion\n  indicated by two short speed lines only.\nPanel 3 — medium shot from behind: the cat slips around a corner ahead, only tail and hindquarters visible;\n  the girl's arm reaches into frame from the left, fingers open, missing it.\nPanel 4 — over-the-shoulder: the alley opens onto a narrow exterior stairwell rising to the right; the cat is\n  two steps up, paused, looking back; the girl's shoulder and hair fill the left edge.\nPanel 5 — low angle from below: the girl at the top of the stairs, mid-leap, cape of her jacket lifting,\n  the roofline and sky behind her, strong diagonal composition.\nPanel 6 — wide, calm: a rooftop with a water tank and a low parapet; the cat sits on the parapet with its back\n  to us; the girl stands four steps away, doubled over and out of breath, one hand on her knee.\n\nCONSISTENCY ACROSS ALL SIX PANELS\nSame character throughout: a girl of about sixteen, short dark bobbed hair with a blunt fringe, a bomber jacket\nwith ribbed cuffs over a striped shirt, dark trousers, scuffed canvas shoes. Her hairstyle, fringe, jacket,\nshirt stripes and shoe design are identical in every panel where she appears. The cat is a single dark tabby\nof the same size and markings in every panel. The alley is the same alley: brick walls, the same laundry lines,\nthe same crates.\n\nCAMERA\nEvery panel uses a different shot — wide, extreme close-up, medium, over-the-shoulder, low angle, wide — and no\npanel repeats another panel's framing. Panels 1 and 6 are both wide but differ in subject and altitude.\n\nCONSTRAINTS\nNo text anywhere: no dialogue, no captions, no signage, no sound-effect lettering, no speech balloons, no page\nnumbers, no watermark, no signature. Do not add a title strip or a border beyond the panel gutters. Do not\ncolour any element. Do not let the character's design change between panels, and do not mirror her asymmetrically\nbetween panels. Nothing may be cropped by a panel edge that should not be: hands, feet and the cat's tail stay\ninside their panels.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/07-six-panel-manga-spread"}
{"id": "cases/08-isometric-bookshop-cutaway", "group": "cases", "slug": "08-isometric-bookshop-cutaway", "title": "Isometric Bookshop Cutaway", "prompt_words": 558, "prompt": "A true isometric cutaway illustration of a three-storey independent bookshop in the evening, on a 1:1 canvas,\ndrawn in a single axonometric projection at 30 degrees on both horizontal axes, orthographic, with no\nperspective and no vanishing point. The front-right quarter of the building is removed so all three floors and\nthe roof are visible at once. The building is lit from inside; it is dusk outside.\n\nTHE BUILDING\nNarrow and tall, roughly twice as tall as it is wide. Brick side walls with visible coursing, a slate roof with\na glazed skylight and a small chimney, a painted timber shopfront on the ground floor with a deep doorway\nrecess, and a gutter running the full width of the roof edge. Floor slabs read as solid sections with visible\nthickness.\n\nGROUND FLOOR — the shop\nA long service counter with a till, a stack of books waiting to be shelved, a low display table with books\nfanned open face-up, two deep reading armchairs, a floor lamp, a potted rubber plant, a brass bell on the\ncounter, a stack of newspaper, a bicycle leaning inside the doorway, and a cat asleep on the armchair.\n\nMIDDLE FLOOR — the shelves\nSix full-height shelves in two rows along the side walls, a rolling wooden ladder hooked to a rail, a reading\nnook with a window seat and a cushion, an open atlas on a stand, a globe on a cabinet, a desk lamp, a stack of\noversize art books lying flat, a rolled map in a corner, and a wall clock.\n\nTOP FLOOR — the café\nA small espresso machine on a counter with two cups and a knock box, three round tables with chairs, a tray of\npastries under a glass dome, a floor-to-ceiling window with the evening sky beyond, hanging pendant lamps,\na wall of small framed prints, a chalkboard, a water jug, and a tall shelf of mugs.\n\nROOF\nThe glazed skylight glowing from below, the chimney, a small water tank, two pigeons, and a weather vane.\n\nFIGURES\nSix small figures, each doing something different: one browsing a shelf with a book half out, one reading in a\nground-floor armchair, one climbing the rolling ladder, one carrying a stack of books up the stairs, one seated\nat a café table with a cup, one standing at the top-floor window looking out. Figures are small relative to the\nbuilding and are never repeated as identical stamps.\n\nLIGHT AND MATERIAL\nEvening: cool blue dusk light from outside, warm amber lamplight pooled inside each floor, and the skylight\nglowing up onto the roof. The two light temperatures never mix — exterior blue, interior amber. Materials:\npainted timber, brick, glass, brass, worn leather on the armchairs, matte paper. Wood grain and brick texture\nare visible but never noisy.\n\nCONSTRAINTS\nTrue isometric only: parallel verticals stay parallel, both floor plans keep the same 30-degree axes, and no\nelement is drawn in perspective. Each floor's ceiling must not be drawn — the viewer looks down into each level.\nStaircases must connect visibly between floors and land on the slabs. Do not add text, signage, prices, numbers,\nlettering on spines, or a shop name; all book spines are abstract blocks of colour with no lettering. No\nwatermark. Do not fill the building with so much furniture that the floor plates stop reading as separate levels.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/08-isometric-bookshop-cutaway"}
{"id": "cases/09-lunar-courier-character-sheet", "group": "cases", "slug": "09-lunar-courier-character-sheet", "title": "Lunar Courier Character Sheet", "prompt_words": 687, "prompt": "A character design sheet for a lunar courier, drawn on a 3:2 landscape sheet of warm off-white design paper at a\nslightly larger scale than a technical reference: four zones, each with a clear border of empty space, hand-drawn\nproduction-art quality, flat cel shading with a single light source from the upper left.\n\nTHE CHARACTER — reproduce identically in every view\nA courier of about thirty, wiry build, medium height, close-cropped dark hair, a narrow face with a strong\njawline and a thin scar through the left eyebrow. Wearing a cream short cape that stops at the hip with a\nscuffed brass clasp at the throat, a slate-grey flight suit with a high collar and ribbed cuffs, worn carbon\nboots scuffed at the toe, and fingerless gloves.\n\nEquipment: a hard-shell backpack in cobalt blue with a visible seam, two side pockets and a rolled blanket\nstrapped on top, and — the asymmetric detail — a single orange tool box clipped to the character's RIGHT hip.\nNothing is worn on the left hip.\n\nTHE HIP SIDE — read this before drawing any view\nThe orange tool box is on the character's OWN RIGHT hip. A figure facing the viewer is mirrored, so the two\nviews must show it on opposite sides of the sheet:\n- FRONT or three-quarter-front view: the box appears on the **LEFT** side of the sheet.\n- BACK view: the box appears on the **RIGHT** side of the sheet.\n- SIDE view: the character faces the **viewer's right**, so that the right hip is the near side and the box is\n  visible in profile at the front of the hip.\nNever draw the box on the same side of the sheet in the front view and the back view. Getting this wrong is the\nsingle failure that ruins the sheet.\n\nZONE 1 — main three-quarter front view, upper LEFT of the sheet, the largest figure\nFull body, standing, weight on the left leg, turned about thirty degrees, head facing the viewer, arms relaxed.\nThe orange tool box reads on the left side of the sheet. Both hands are empty and open.\n\nZONE 2 — side view, upper RIGHT of the sheet\nFull body, facing the viewer's right, standing straight, arms at the sides. The backpack's profile is visible:\nits depth, the two side pockets, the rolled blanket on top. The orange tool box is visible on the near hip.\n\nZONE 3 — back view, lower centre-left of the sheet, roughly a third smaller than Zone 1\nFull body, seen from directly behind. The cape falls open at the back, the backpack covers the upper back with\nits straps crossing at the chest, and the rolled blanket sits on top of it. The orange tool box appears on the\nRIGHT side of the sheet.\n\nZONE 4 — equipment detail, lower right of the sheet, no figure\nThe backpack shown on its own, opened: the main compartment, the two side pockets, and the internal divider\nlaid out flat. Drawn at about half the size of the backpack in Zone 2 and consistent with it in every\nproportion, seam position and colour. Nothing else is drawn in this zone.\n\nSHEET PRESENTATION\nDrawn figures and equipment only, on flat design paper. No grid lines, no measurement rulers, no dimension\narrows, no callouts, no leader lines. Line work is confident and slightly varied in weight; shadows are a single\nflat tone per material, not rendered gradients. The paper has a faint tooth and a soft uneven edge where the\nsheet meets the background.\n\nCONSTRAINTS\nNo text anywhere: no labels, no names, no view captions, no colour swatches with codes, no watermark, no\nsignature. The character's face, hair, cape, suit, boots and both pieces of equipment must be identical across\nZones 1 to 3 — same proportions, same colours, same seam positions. The backpack in Zone 4 must match the\nbackpack as worn in Zones 1 to 3. Do not mirror the orange tool box between the front and back views. Do not add\na second character, a pose sheet, a turnaround strip, a colour palette strip, or a background environment.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/09-lunar-courier-character-sheet"}
{"id": "cases/10-one-room-warmer", "group": "cases", "slug": "10-one-room-warmer", "title": "One Room Warmer", "prompt_words": 694, "prompt": "Image 1: the room — the sole authority for the architecture, the camera and the existing furniture. It supplies\nno new styling, no new objects and no new light.\n\nTASK\nKeep this room exactly as it is and make it warmer: change only the soft furnishings and add one lamp, so the\nsame room reads as lived-in and inviting while remaining unmistakably the same room.\n\nPRESERVE EXACTLY — architecture, joinery and furniture\n- The window: same position, same size, same plain white frame, and the same glazing. It is a simple two-leaf\n  casement window with ONE horizontal glazing bar in each leaf, six panes in total. Do not convert it into a\n  multi-pane Georgian or sash window, do not add bars, do not add muntins, do not change the frame profile.\n- The door: same position, same swing, same plain FLAT door leaf with no panels and no mouldings, same simple\n  lever handle on the same side. Do not add raised panels, mouldings, glazing or a different handle.\n- The sofa: same position against the left wall, same size, same silhouette, same wooden feet. Its upholstery\n  stays the same grey woven fabric of the same mid-grey value — do not lighten it, do not re-colour it, do not\n  change it to linen or cream, do not change the weave.\n- Walls, skirting, ceiling line, both room corners, the socket on the right wall and the low scuff mark on the\n  left wall, all unchanged.\n- The camera: identical position, identical height, identical angle and identical lens. The result must align\n  with Image 1 edge to edge, with the ceiling line, both corners and the floor plane in the same places.\n- The pale grey carpet and the faint worn path in the middle of the room.\n\nCHANGE — only these five things\n- Add a large wool rug under the sofa and slightly in front of it: warm terracotta with a simple repeating\n  geometric border.\n- Replace the four grey cushions with four cushions in warm rust, ochre and cream linen, one slightly creased.\n  The sofa itself keeps its grey upholstery — only the loose cushions change.\n- Hang floor-length linen curtains in warm oatmeal on a slim black rod above the window, drawn open to both\n  sides so the window stays completely visible and no glass is covered.\n- Add one slim floor lamp with a cream linen shade standing in the near right corner, switched on.\n- Set a small stack of two books on the floor beside the sofa's near arm, on the rug.\n\nMATCH NATURALLY\n- The lamp adds a warm pool of light entering from the right; the window keeps its cool daylight from the left.\n  Both must be visible in the final frame and neither may cancel the other.\n- Every new object casts a soft-edged shadow, and each shadow agrees with both light sources: a soft shadow to\n  the right from the window, a second softer shadow away from the lamp.\n- Fabric must read as fabric: wool with visible fibre and a slightly flattened pile where the sofa legs press\n  into it, linen with a soft vertical fold and a hint of translucency at the curtain's leading edge.\n- The rug must follow the floor plane in the same perspective as the carpet beneath it.\n\nDO NOT\n- Do not restyle the joinery. The window stays a plain two-leaf casement, the door stays a plain flat leaf.\n- Do not re-colour, re-upholster or restyle the sofa in any way.\n- Do not move, resize or remove the window, the door, the walls, the skirting, the ceiling or the sofa.\n- Do not change the camera, the crop, the aspect ratio or the perspective. This is the same frame.\n- Do not add architectural elements: no second window, no built-in shelving, no new doorway, no wallpaper, no\n  picture rail, no coving.\n- Do not add people, pets, plants, artwork, a television, a coffee table, a sideboard or a fireplace.\n- Do not brighten or HDR the whole frame. The daylight stays flat and cool; only the lamp area gains warmth.\n- No text, no signage, no watermark.", "images": ["full.jpg", "output.png", "preview.jpg", "reference-1.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/10-one-room-warmer"}
{"id": "cases/11-product-into-a-scene", "group": "cases", "slug": "11-product-into-a-scene", "title": "Product Into A Scene", "prompt_words": 534, "prompt": "Image 1: the product — the sole authority for the dripper's form, its proportions, its cream glaze, the throwing\nline, the interior ribs, the drain hole, the unglazed foot and the shape of the side handle. It supplies no\nsetting, no camera and no lighting.\nImage 2: the scene — the sole authority for the space, the counter, the camera position, the scale of everything\nin frame and the direction and quality of every light. It supplies no product.\n\nTASK\nPlace the dripper from Image 1 onto the clear strip of counter along the front edge of Image 2, standing upright\non its foot at the left third of the frame, and follow the scene's own light exactly.\n\nPRESERVE EXACTLY\n- The dripper's geometry: cone angle, rim diameter, base ring, handle size and its angle, the wall thickness at\n  the rim, and the interior ribs and drain hole.\n- The dripper's material: cream glaze with the same gloss level, the same visible throwing line, the same\n  slightly uneven rim, the same unglazed clay foot.\n- The scene: the oak worktop with its grain, knife marks and water ring, the tiled splashback, the wall above\n  it, the enamel kettle, the folded linen cloth, the bowl of beans and their positions.\n- The camera: identical position, height, angle and framing. The result must align with Image 2 edge to edge.\n- The empty strip of counter along the front edge stays otherwise empty.\n\nMATCH NATURALLY\n- Light direction: the window is off-frame to the left, so the dripper is lit from the left. Its lit side is on\n  the left, its shadow falls to the right, and its right side falls into the same gentle gradient as the\n  splashback behind it.\n- Colour temperature: warm in the highlight, cool in the shadow, matching the rest of the scene. Do not add a\n  warm or cool cast of its own.\n- Contact shadow: one soft-edged shadow beneath the base, short and to the right of the object, denser where the\n  foot ring meets the counter and fading outward.\n- Reflection: a faint, low-contrast reflection of the base in the oiled oak immediately beneath it, the same\n  intensity as the reflection the kettle already casts.\n- Scale: the dripper is roughly the height of the enamel kettle's body and stands on the same plane, its base\n  resting on the counter surface at the same depth as the kettle's base. It must not float or sink.\n- Grain and sharpness: identical to the rest of the frame at the same depth.\n\nDO NOT\n- Do not redesign, restyle, recolour or reshape the dripper, and do not change its glaze or its handle.\n- Do not change the counter, the splashback, the existing three objects, the camera, the crop or the light\n  direction.\n- Do not add coffee, a filter, a mug, beans, a spoon, steam, a hand or a second product.\n- Do not paste the object flat: no outline, no cut-out edge, no drop shadow of even density, no missing contact\n  shadow, no flat colour fill.\n- Do not apply a shadow that disagrees with the light coming from the left.\n- No text, no logo, no watermark.", "images": ["full.jpg", "output.png", "preview.jpg", "reference-1.jpg", "reference-2.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/11-product-into-a-scene"}
{"id": "cases/12-a-small-delivery", "group": "cases", "slug": "12-a-small-delivery", "title": "A Small Delivery", "prompt_words": 600, "prompt": "A four-panel wordless comic, 3:2 landscape, arranged as two columns of two equal panels with even gutters,\nreading left to right then top to bottom. Ink line with watercolour wash, consistent throughout. No text\nof any kind.\n\nSTYLE\nLoose confident ink linework with a fine nib, watercolour washes layered underneath, edges softening slightly\noutside the lines. Cool grey-blue palette for the rain, warm amber for the lit interior. Paper texture visible\nin the washes. Whites are left as raw paper. Nothing is digitally sharp or vector-clean.\n\nTHE CHARACTER — identical in all four panels\nA small boxy postal robot, about knee height, with a rounded rectangular body in faded red paint with visible\nscratches and a dent in the left flank, a single round lamp eye set high on the front, a small brass slot on\nthe chest, short articulated legs with flat feet, and a worn canvas mail satchel on a strap across its body.\nIts proportions, paint wear and the dent must be in the same place in every panel.\n\nTHE STORY, four beats\nPanel 1 — the discovery. Medium shot, low angle at the robot's eye level. Rain falls across the frame. The\nrobot stands on a wet pavement looking down at a small paper crane lying crumpled in a puddle, one wing\nsoaked through and dark. A single street lamp above and behind it throws a cold cone of light; the crane sits\njust inside the light's edge.\nPanel 2 — the protection. Close-up, from just above. The robot has lifted the crane clear of the water and is\nholding it against its chest with one small articulated hand while it pulls a large used envelope out of the\nsatchel with the other. The envelope is already creased and its flap is torn. The robot's lamp eye is angled\ndown at the crane.\nPanel 3 — the journey. Wide shot, from behind and slightly to the side. The robot walks away from us up a\nsteep wet cobbled lane toward a single warm light at the far end. It is holding the envelope arched over the\ncrane like a small roof; the crane is hidden underneath, only its folded tip visible. Rain streaks the frame\ndiagonally. The lane is empty and dark.\nPanel 4 — the arrival. Medium close-up, from a low angle. The robot stands on a dry stone doorstep in front of\na warm amber doorway with a paper of light spilling out onto the wet step. It is holding the crane up with both\nhands, offering it. The crane is dry and its folds have opened back into shape. The rain has stopped and only\ndrips from the gutter remain.\n\nCONSISTENCY\nThe same robot, the same satchel, the same envelope once it appears, the same crane. Lighting is consistent\nthroughout: cold blue-grey for the rain in panels 1 to 3, warm amber for the doorway in panel 4. The paper\ncrane's condition progresses in one direction only — soaked in panel 1, dark and limp in panel 2 and 3, dry and\nopened back into shape in panel 4.\n\nCONSTRAINTS\nNo text anywhere: no dialogue, no captions, no sound-effect lettering, no signage, no page numbers, no\nwatermark, no signature. Do not add a title, a border beyond the gutters, or a gutter gutter label. Do not add\npeople, cats, cars, bins or other props. Do not let the robot's design change between panels, and do not mirror\nits dent or its satchel strap between panels. Do not make the crane into a bird — it stays a folded paper object\nin every panel.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/12-a-small-delivery"}
{"id": "cases/13-watercolor-coastal-town", "group": "cases", "slug": "13-watercolor-coastal-town", "title": "Watercolor Coastal Town", "prompt_words": 445, "prompt": "A loose watercolour painting of a small coastal town on a steep hillside, on a 3:2 landscape sheet of rough\ncold-pressed paper, painted from across the water at a low vantage point.\n\nCOMPOSITION\nThe town rises diagonally from the lower right to the upper left, filling the right two thirds of the sheet\nwith stacked whitewashed houses and terracotta roofs. To the left, a strip of harbour water occupies the lower\nthird and opens toward the upper left, where the paper is left almost entirely unpainted. In the immediate\nforeground at the lower right, a short flight of worn stone steps and a low wall anchor the composition and are\npainted darkest. A single small boat sits on the water at the middle left, very simply described, no rigging\ndetail. The sky is bare paper with one pale wash at the very top edge. No element touches the sheet's edges;\ngenerous margins are left on all four sides.\n\nWATERCOLOUR TECHNIQUE — this is the subject of the painting as much as the town is\nWet-in-wet for the sky and the water, so the two bleed into each other along a soft horizontal band. Wet-on-dry\nfor the roofs and the walls, so their edges stay crisp on the lit sides and soften on the shadowed ones. Where\ntwo washes meet, one is allowed to bloom into the other and the granulation of the pigment settles visibly in\nthat boundary. Roof terracotta granulates into the wall below it in places. Several edges are deliberately lost:\nthe far end of the town dissolves into the sky with no line at all. The paper's tooth shows through every wash,\nand a few small dry-brush skips appear where pigment was dragged across the grain. Whites are unpainted paper —\nthere is no white paint anywhere in the image.\n\nPALETTE — four pigments only\nA warm ochre, a burnt sienna for the roofs, a cool grey-blue for the water and distant hill, and one deep\nindigo for the darkest notes in the foreground steps and the shadowed sides of two or three houses. Colours are\nallowed to mix on the paper rather than on the palette, so the greys lean green in places and violet in others.\n\nCONSTRAINTS\nNo people, no animals, no vehicles, no boats beyond the single one, no cranes, no washing lines with legible\nlaundry, no signage. No text, no signature, no stamp, no watermark. Do not tighten the painting: no hard\noutlines around buildings, no evenly-filled shapes, no white body paint, no digital-looking gradients. Do not\npaint every window — most house fronts are left as a single wash with one or two dark window marks only.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/13-watercolor-coastal-town"}
{"id": "cases/14-1936-travel-poster", "group": "cases", "slug": "14-1936-travel-poster", "title": "1936 Travel Poster", "prompt_words": 512, "prompt": "A 1936 travel poster, 2:3 portrait, designed as a four-colour silkscreen print on aged poster stock and seen\nperfectly flat and straight on — this is the printed sheet, not a photograph of one on a wall.\n\nCOMPOSITION\nA steep mountain valley seen head-on, arranged in three flat bands. The lower band is the valley floor, a single\nflat dark green. Above it, a tall viaduct crosses the full width on five arches, drawn as one flat ochre shape\nwith the arches cut out as negative space. On the viaduct, a passenger train in profile running left to right:\nlocomotive, tender and four carriages, reduced to their simplest readable silhouettes with no interior detail\nand no visible windows, only a row of small rectangles along the carriages. Behind and above the viaduct, three\npeaks rise in two flat blues, the nearer one darker. The sky is the paper colour left bare, with two long\nhorizontal clouds as single flat shapes. Vertical composition: the train sits on the upper third, the valley\nfloor occupies the lower quarter, and the peaks fill the space between.\n\nTEXT — render exactly these two strings, each appearing once\n\"A LOWER LINE\"\n\"1936\"\nSet \"A LOWER LINE\" in a bold geometric sans, all capitals, generously letter-spaced, centred in the band of\nsky below the peaks; its width is about two thirds of the sheet's width. Set \"1936\" much smaller in the same\ntypeface, centred directly beneath it, at about one sixth of the title's height. No other text anywhere: no\nroute names, no station names, no prices, no small print, no printer's marks.\n\nPRINT CHARACTER — deliberately imperfect\nFour flat inks only: ochre, deep green, two blues, plus the paper. Each colour sits in its own solid shape with\nhard edges and no gradients, no shading and no halftones. Registration is very slightly out on one or two\nplates, so a thin sliver of paper shows along the left edge of the train's ochre and along the top edge of one\nblue peak — under one millimetre and never enough to look broken. Ink is slightly uneven within a shape where\nthe screen was thin, with fine horizontal streaks visible in the largest flat areas. The paper is aged: a warm\noff-white with light foxing at the corners, two soft handling creases running diagonally across the lower left,\na small water stain at the upper right that has very slightly lifted the blue, and a tiny nibbled loss at the\nbottom edge. Four pinholes, one in each corner, sit just inside the margin.\n\nCONSTRAINTS\nNo illustration style other than flat silkscreen: no painterly edges, no airbrush, no glow, no photo texture,\nno 3D shading, no drop shadows. Do not add people, birds, trees with individual leaves, rocks, rivers, road\nsigns or a border. Do not paraphrase, re-case, hyphenate or translate the two strings, and do not add a\nyear, a place name or a slogan. No mockup: no wall, no frame, no shadow behind the sheet, no perspective tilt,\nno tape or pins other than the four pinholes.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/14-1936-travel-poster"}
{"id": "cases/15-irezumi-flash-sheet", "group": "cases", "slug": "15-irezumi-flash-sheet", "title": "Irezumi Flash Sheet", "prompt_words": 528, "prompt": "A traditional Japanese tattoo flash sheet, 3:4 portrait, drawn in black sumi ink with grey wash and one accent\ncolour on a single aged sheet of off-white paper, flat and straight on.\n\nSHEET\nSix motifs arranged in a 3 by 2 grid with wide, even clear space between them and a generous margin on all four\nsides. Each motif is a finished, self-contained design that could be lifted off the sheet and tattooed as it\nstands. No boxes, no frames, no dividing rules, no shading behind the motifs. The paper is warm off-white with\na faint fibrous tooth, a soft crease across the lower third, and a very slight tea-coloured tone at the edges.\n\nTHE SIX MOTIFS, top row left to right then bottom row left to right\n1. A koi carp swimming upward and to the left, body curved into a strong S, fins spread and filled with fine\n   black lines, a few scales picked out individually near the shoulder and the rest left as graded grey wash.\n2. A single open peony, petals layered from a pale outer ring to a dense dark centre, with two leaves behind it\n   and one stem cut off square at the bottom.\n3. A breaking wave drawn with clawed white crests in the traditional manner, the interior filled with a fine\n   pattern of parallel curved lines and graded wash, the spray flicked outward at the top right.\n4. A coiled dragon head and shoulder only, seen from the front-left, jaws open, whiskers sweeping back, one\n   claw raised; the body is implied and cropped by the drawing's own edge, not by the sheet.\n5. A cherry blossom branch entering from the lower right with five open blossoms at different stages and three\n   closed buds, the branch drawn in a single confident stroke with two small knots.\n6. A hannya mask, front on, horns curving back, eyes narrowed, mouth open to show pointed teeth, the cheeks\n   graded from pale grey at the temples to near-black at the jaw.\n\nINK AND WASH\nFine confident black linework with clear variation in weight: heavy on the outer contours, hair-fine on scales,\nwhiskers, petal veins and water lines. Mid-tones are built from even black wash in distinct flat gradations,\nnot smooth gradients — you can see the edges where one wash value meets the next. The only colour on the sheet\nis a single vermilion accent, used sparingly and identically in each motif: one small area per design, never\nmore than about a tenth of it. Whites are untouched paper and are used as a designed element, not left over.\n\nCONSTRAINTS\nNo text anywhere: no labels, no Japanese characters, no signature, no seal, no stamp, no numbers, no brand mark,\nno watermark. Do not add background: no clouds, no water, no flowers behind the motifs. Do not merge two motifs\ninto one design and do not let one motif run into its neighbour's space. Do not use more than one accent colour,\ndo not fill large areas with it, and do not outline the motifs with a coloured line. Do not render as a 3D\nillustration, a tattoo on skin, or a photograph of a sheet.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/15-irezumi-flash-sheet"}
{"id": "cases/16-editorial-beauty-portrait", "group": "cases", "slug": "16-editorial-beauty-portrait", "title": "Editorial Beauty Portrait", "prompt_words": 648, "prompt": "A beauty editorial photograph, 2:3 portrait, a head-and-shoulders portrait with a skincare product standing in\nthe foreground on the same focal plane, shot in a studio against a deep plum seamless.\n\nTHE FACE\nA woman of about thirty-five, head and shoulders in frame, turned about forty degrees to her left so three\nquarters of her face is toward the camera, chin very slightly lowered, eyes looking just off the lens to the\nleft. Expression settled and unperformed — mouth closed and relaxed, no smile for the camera. Dark hair pulled\nback off the face with a few loose strands at the temple and one at the jaw. Bare shoulders and a high-necked\nfine-rib knit in off-white, the collar visible at the lower edge of the frame.\n\nSKIN — the point of the picture\nReal skin at close range: visible pores across the nose and inner cheeks, fine lines at the outer eye and\nbetween the brows, a faint natural redness across the cheeks and chin, a few small freckles on the temple and\nthe bridge of the nose, one small mole below the left jaw, fine down on the jawline catching the light, and a\nslight shine on the forehead, nose tip and upper lip — healthy, not greasy, and nowhere matte-flat. Skin tone is\neven in tone but not in texture. No retouching, no smoothing, no pore-filling, no whitened sclera, no reshaped\nfeatures. The under-eye area keeps its real depth and colour.\n\nPRODUCT\nAn amber glass dropper bottle, 30 ml, standing upright on a low dark stone plinth in the lower left foreground,\nat the same focus plane as the face so both are sharp. Matte black dropper cap with a fine ribbed skirt, a\npipette bulb visible through the glass shoulder, the liquid inside a clear pale gold. A rectangular off-white\npaper label wraps the front of the bottle, slightly proud at its edge, with a fine fibre texture. The bottle is\nabout a third the height of the face in frame and sits closer to the camera, lower in the picture.\n\nEXACT VISIBLE TEXT — render exactly these two strings on the label, once each\n\"SERUM No.3\"\n\"30 ML\"\n\"SERUM No.3\" is the larger, in a fine uppercase sans with wide letter spacing, sitting on the upper third of the\nlabel. \"30 ML\" is much smaller, centred at the bottom of the label. Both must be crisp and correctly kerned. No\nother text, numbers, symbols or marks appear anywhere in the frame.\n\nLIGHT AND CAMERA\nOne large soft source from the front left at about forty-five degrees, close to the subject so the falloff is\ngentle; one narrow strip light from behind on the right to rim the cheek, the jaw and the shoulder; one black\nflag on the right to deepen the shadow side. The shadow side of the face keeps detail — it is deep, not blocked.\nNo fill from the front. Camera at eye level, 85 mm equivalent, head and shoulders, the top of the head inside\nthe frame with headroom, the collar cut by the lower edge. Shallow depth of field, but the eyes, the lips and\nthe label are all inside it.\n\nCOLOUR AND FINISH\nDeep plum seamless background, one value, no gradient and no visible sweep edge. Palette is warm neutral skin\nagainst cool plum, with the amber glass the only saturated element. Natural colour, slight film grain, no\ncoloured gels, no glow, no bloom.\n\nCONSTRAINTS\nNo other products, no cotton pads, no flowers, no water droplets on the face, no jewellery, no visible hands or\nfingers anywhere in the frame, no second figure. No invented claims, ingredient names, badges, awards, price\ntags, barcodes or certification marks. Do not add text anywhere except the two strings above. Do not smooth,\nlift, slim, widen or symmetrise the face. Do not make the skin plastic, waxy or uniformly matte. No watermark.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/16-editorial-beauty-portrait"}
{"id": "cases/17-fashion-lookbook-frame", "group": "cases", "slug": "17-fashion-lookbook-frame", "title": "Fashion Lookbook Frame", "prompt_words": 562, "prompt": "A full-body fashion editorial frame, 3:4 portrait, shot in a stripped-back concrete studio in front of a single\nlarge window, daylight only.\n\nTHE FRAME\nFull body, head to feet, both feet inside the frame with clear floor beneath them and headroom above the head.\nCamera at chest height, 50 mm equivalent, straight on, no tilt and no wide-angle distortion. The figure stands\njust left of centre, turned about thirty degrees, leaving a wide column of empty concrete wall and window light\nto the right. The floor is visible across the lower quarter. Nothing is cropped.\n\nTHE SUBJECT AND THE OUTFIT\nA woman in her late twenties, about 175 cm, standing with her weight on the left leg, the right leg relaxed and\nslightly forward, one hand pushed into the coat pocket so no fingers are visible anywhere in the frame, the\nother arm hanging relaxed at her side with the hand mostly hidden behind the coat's edge. Chin level, gaze\ndirected off to the left of the lens, expression neutral and unperformed, mouth closed.\n\nOutfit, in this order from the outside in: an oversized double-breasted wool coat in camel, cut long to mid-calf,\nwith wide notch lapels, horn buttons and a visible brushed nap; beneath it a fine-gauge ribbed knit in chalk\nwhite with a high rolled neck; wide-leg wool trousers in charcoal with a single centre crease; low leather\nderby shoes in dark brown, laced. The coat is worn open, hanging straight from the shoulders, so the knit and\nthe trouser waist are both visible. One shoulder of the coat sits slightly forward, as it would if a real\ngarment were worn open.\n\nSTUDIO AND LIGHT\nA bare concrete studio: board-formed walls with visible horizontal formwork lines and a scatter of small tie\nholes, a polished concrete floor with a faint sheen, one large steel-framed window occupying the right third of\nthe back wall. Daylight only, entering through that window from the right at a low morning angle, so the\nsubject is lit from her right side: the lit side of the coat shows the wool's nap and the shadowed side holds\ndetail without going black. Long soft-edged shadow of the figure falls to the left across the floor. No\nartificial light, no fill, no reflector, no coloured light.\n\nMATERIAL\nWool coat: brushed nap readable at this distance, edges of the lapel slightly lighter where the fibre catches\nthe light, one or two fine creases across the sleeve at the elbow. Rib knit: fine vertical ribs, a matte\nsurface, a soft fold at the neck. Trousers: smooth worsted wool with a crisp centre crease and a slight break\nover the shoe. Leather shoes: scuffed toe, one clean highlight. Concrete: cold grey, matte, small aggregate\nvisible in the floor's polished sheen.\n\nCONSTRAINTS\nNo visible hands, no fingers, no jewellery, no bag, no hat, no scarf, no second figure, no props, no furniture,\nno plants. No text, no signage, no logo on the clothing, no watermark. Do not change the outfit's proportions,\ndo not shorten the coat, and do not turn it into a different garment. Do not add a background wall other than\nthe concrete described, and do not add a cyclorama or a studio backdrop paper. Do not pose the subject: no\nhand on hip, no arm raised, no walk, no jump. Do not retouch the face or slim the body.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/17-fashion-lookbook-frame"}
{"id": "cases/18-brutalist-library-at-dusk", "group": "cases", "slug": "18-brutalist-library-at-dusk", "title": "Brutalist Library At Dusk", "prompt_words": 492, "prompt": "A wide architectural photograph of a brutalist public library at dusk, 21:9, shot from across a wet plaza.\n\nTHE BUILDING\nA long, low concrete volume, roughly five times as wide as it is tall, sitting heavily on a raised podium.\nIts upper mass cantilevers out over a recessed ground floor, so a deep shadowed undercroft runs the full width\nbeneath it. The concrete is board-formed, with horizontal formwork lines running the entire length at even\nspacing, a regular grid of small circular tie holes, and long vertical rain streaks staining the surface below\nthe parapet. A continuous band of clerestory glazing sits directly under the roof slab along the whole facade,\nglowing warm from the interior lights inside. At the left end, a broad flight of concrete steps rises to the\nentrance, flanked by two low concrete planters holding bare winter shrubs. The roof slab reads as a single\nheavy horizontal line against the sky.\n\nTHE PLAZA\nWet dark paving across the lower half of the frame, holding long reflections of the glowing window band and of\nthe street lamps. Two shallow puddles with visible edges. A row of low concrete bollards runs across the middle\ndistance. A single steel handrail beside the steps. One street lamp at the far right, out of focus. In the\nforeground, the corner of a concrete planter crops into the lower right of the frame, out of focus.\n\nTHE FIGURE\nOne person, small in the frame, walking away from the camera toward the entrance at the left, wearing a long\ndark coat and carrying a closed umbrella. Their reflection stretches down into the wet paving beneath them.\nThey are the only person in the frame and they occupy roughly one fifteenth of the frame height.\n\nLIGHT\nBlue hour, after sunset, no sun. The dominant light comes from the building itself: the warm amber clerestory\nband is the brightest thing in the picture and it is the source of every warm reflection on the wet plaza. Cool\nblue-grey ambient sky light fills the concrete's upper surfaces and the far end of the plaza. Two street lamps\nadd small cold pools of light at the right. The two temperatures never blend — the concrete outside the window\nband stays cold, the interior glow stays warm. Long soft shadows fall away from the building toward the camera.\n\nCONSTRAINTS\nNo readable signage, no lettering on the building, no logos, no vehicle number plates, no advertising. Any\ntext-like mark must stay illegible: off-angle, out of focus or too distant to resolve, and it must never resolve\ninto invented words. No cars, no buses, no bicycles, no litter, no crowds and no second figure. No lens flare.\nDo not straighten, tilt or perspective-correct the frame: keep the camera level and the verticals true. Do not\nmake the building glossy or reflective; concrete is matte everywhere except where it is wet. No watermark. Do\nnot add neon, coloured gels, mist or light beams.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/18-brutalist-library-at-dusk"}
{"id": "cases/19-coffee-brand-touchpoint-board", "group": "cases", "slug": "19-coffee-brand-touchpoint-board", "title": "Coffee Brand Touchpoint Board", "prompt_words": 693, "prompt": "A brand identity touchpoint board for a specialty coffee roaster, 1:1 square, presented as a studio flat lay\nphotographed from directly above: nine physical items arranged in a 3 by 3 grid on a warm greige paper\nbackdrop, evenly spaced, each item fully inside its own cell with clear air around it.\n\nTHE BRAND — identical on every item\nThe wordmark is the words \"NORTH HALL\" set in a single high-contrast serif, all capitals, with the two words on\none line separated by a single space, tracked tight. It appears on seven of the nine items and must be\nletter-for-letter identical on all of them: same typeface, same weight, same letterforms, same proportions,\nsame spacing between the two words, and the same clear space around it equal to the height of the letter H on\nevery side. A small circular emblem — a thin-stroke arch over a horizontal line — sits above the wordmark on\npackaging only, and is the same size relative to the wordmark wherever it appears.\n\nTHE NINE ITEMS, reading left to right then top to bottom\n1. A 250 g kraft coffee bag, front on, with a cream label panel carrying the wordmark in the upper third and\n   \"ETHIOPIA GUJI\" in a small wide-spaced sans beneath it, plus a row of three small filled squares at the\n   bottom of the label indicating roast level.\n2. The same bag seen from behind, filling less of its cell, showing a plain kraft back with a small cream\n   patch carrying two lines of tiny type and a debossed fold line.\n3. A stone-coloured takeaway cup standing with a printed kraft sleeve around its middle; the wordmark runs\n   across the sleeve once, horizontally, at the sleeve's vertical centre.\n4. A neat stack of thick cotton-rag business cards, the top card square to the camera, carrying the emblem and\n   the wordmark centred, with \"SPECIALTY COFFEE\" in small wide-tracked capitals beneath it.\n5. A slim loyalty card in the same stone colour, laying flat, with the wordmark small at the top left and a row\n   of ten empty circular stamp positions across its lower half, none of them filled.\n6. A type specimen strip: the wordmark set large in the centre of a narrow cream card, and directly beneath it\n   the same wordmark at one third of the size, showing that the letterforms do not change with size.\n7. Five colour chips in a row — deep espresso brown, clay terracotta, bone cream, slate grey, and one muted\n   sage — each a small square with a narrow margin between them, no text on any chip.\n8. A single die-cut vinyl sticker, cream on kraft, carrying only the emblem at a larger size, with no wordmark.\n9. A folded kraft napkin with the wordmark stamped once in the lower right corner in slightly uneven ink, so\n   the stamp's edges are a little soft and one corner of the impression is fainter than the rest.\n\nMATERIAL AND LIGHT\nOne large soft source from the upper left, close to the set, so shadows fall down and to the right and are soft\nand short. Materials must be distinguishable: matte uncoated kraft with visible fibre, thick cotton-rag card\nstock with a deckled textured edge, matt laminated paper on the sleeve, and a slightly glossy vinyl sticker\ncatching one soft highlight. Slight top-down perspective flattening with no keystone distortion.\n\nCOLOUR\nThe palette is limited to the five chips: deep espresso, clay terracotta, bone cream, slate grey and muted sage.\nKraft brown is the substrate, not a palette colour.\n\nCONSTRAINTS\nThe wordmark must be spelled exactly \"NORTH HALL\" everywhere it appears, with no variation, no alternate\nspelling, no translated version and no second logo. Render only these five strings and no others: \"NORTH HALL\",\n\"ETHIOPIA GUJI\", \"SPECIALTY COFFEE\", \"ROASTED IN ROTTERDAM\", \"250 G\". Do not add taglines, tasting notes,\nweights, prices, awards, certifications, recycling symbols, barcodes, QR codes or any other text. Do not add\na tenth item, a plant, a cup of coffee, beans, a hand or a prop. Do not repeat an item in two cells. Do not\nrotate the items: everything is square to the camera. No watermark.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/19-coffee-brand-touchpoint-board"}
{"id": "cases/20-fintech-dashboard", "group": "cases", "slug": "20-fintech-dashboard", "title": "Fintech Dashboard", "prompt_words": 767, "prompt": "A fintech operations dashboard, 16:9, screen only — a flat rectangular interface shown straight on with no\ndevice, no desk, no hand and no perspective. This is a UI mock with real committed data, not a photograph.\n\nLAYOUT — fixed regions, no overlap\n- Top bar, 9% of the frame height: product name at the left, the date-range control and one account chip at the\n  right, and a single hairline rule closing the bar across the full width.\n- Left rail, 15% of the frame width: six navigation items stacked with even spacing. The first is active, marked\n  by a 3 px accent bar on its left edge and a tinted row background.\n- Main area, to the right of the rail, split into three stacked panels with even gaps: KPI row 24% of the main\n  height, charts row 36%, table panel 40%.\n- KPI row: four equal tiles side by side, each with its label above its value, its value above its delta, and\n  one small sparkline at the tile's right edge.\n- Charts row: a line chart at 62% of the row width and a stacked bar chart at 38%, with one gutter between them.\n- Table panel: a six-column table with a header row, five data rows and a footer line.\n- Every panel sits on a flat ground with a 1 px border. No drop shadows anywhere.\n\nDATA — this must be internally consistent, and the arithmetic must close\n- Net volume: $4.82M, up 12.4%, sparkline rising left to right.\n- Active accounts: 18,204, up 3.1%, sparkline rising left to right.\n- Auth rate: 97.6%, down 0.4%, sparkline falling left to right.\n- Settlement time: 1.9 days, down 0.3 days, sparkline falling left to right.\n  Each delta's sign must agree with the direction its own sparkline moves. All four sparklines appear.\n- The line chart has exactly two series, each with a legend entry, and no legend entry without a series.\n- The five table amounts below sum to exactly $50,671.50, which is the figure stated in the table footer.\n\nEXACT VISIBLE TEXT — render exactly these strings\n- Product name: \"MERIDIAN PAY\"\n- Date range: \"JAN-JUN 2026\"\n- Account chip: \"Acme Ltd\"\n- Navigation, top to bottom: \"Overview\" / \"Payments\" / \"Payouts\" / \"Customers\" / \"Risk\" / \"Settings\"\n- KPI tile 1: \"NET VOLUME\" then \"$4.82M\" then \"+12.4%\"\n- KPI tile 2: \"ACTIVE ACCOUNTS\" then \"18,204\" then \"+3.1%\"\n- KPI tile 3: \"AUTH RATE\" then \"97.6%\" then \"-0.4%\"\n- KPI tile 4: \"SETTLEMENT TIME\" then \"1.9 days\" then \"-0.3 days\"\n- Line chart title: \"Volume and refunds\"; its legend entries: \"Gross volume\" and \"Refunds\"\n- Line chart x-axis ticks, left to right: \"JAN\" \"FEB\" \"MAR\" \"APR\" \"MAY\" \"JUN\"\n- Line chart y-axis: four labelled ticks ascending bottom to top: \"0\", \"1M\", \"2M\", \"3M\"\n- Bar chart title: \"Payouts by rail\"; its legend entries: \"ACH\", \"Card\", \"Wire\"\n- Table headers, in order: \"TRANSACTION ID\" \"MERCHANT\" \"RAIL\" \"AMOUNT\" \"STATUS\" \"SETTLED\"\n- Table rows, in order:\n  \"TXN-8842\" \"Kestrel Supply\" \"ACH\" \"$12,400.00\" \"Settled\" \"04 Jun\"\n  \"TXN-8843\" \"Northgate Ltd\" \"Card\" \"$8,150.50\" \"Settled\" \"05 Jun\"\n  \"TXN-8844\" \"Albion Works\" \"Wire\" \"$21,900.00\" \"Pending\" \"06 Jun\"\n  \"TXN-8845\" \"Vantage Foods\" \"ACH\" \"$5,320.75\" \"Settled\" \"06 Jun\"\n  \"TXN-8846\" \"Ridgeway Co\" \"Card\" \"$2,900.25\" \"Failed\" \"07 Jun\"\n- Table footer: \"Showing 1-5 of 312\" at the left and \"Subtotal $50,671.50\" at the right\n\nTYPOGRAPHY\nOne UI sans throughout, five sizes only: page-level product name, KPI value, panel title, table and axis text,\nand small labels. Numerals are tabular. The AMOUNT column is right-aligned; all other columns are left-aligned.\nKPI values are right-aligned within their tiles. All text is horizontal.\n\nRENDERING\nFlat interface rendering: no photographic lighting, no reflections, no glass, no bevels, no gradients. Ground\n\"#FBFAF8\", body text \"#1C1B19\", one accent \"#3D6B4F\" used only for the active nav bar, the positive deltas and\nthe primary chart series; negative deltas in a muted \"#B4442E\". Hairline borders in \"#E3DFD8\". A 1 px grid of\nfaint horizontal rules in the table only.\n\nCONSTRAINTS\nRender only the strings listed above. Do not invent a metric, a merchant, a transaction, a menu item, a column,\na tooltip, a notification badge or a footnote. Do not use lorem ipsum, \"TBD\", \"xx,xxx\" or placeholder dashes.\nEvery chart must have both labelled axes and legend entries that match the series actually drawn. No two table\nrows may share a transaction ID and no value may repeat verbatim in two rows of the same column. No browser\nchrome, no OS status bar, no notch, no cursor, no selection highlight, no watermark. Do not round or reformat\nany figure, and do not change a single digit.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/20-fintech-dashboard"}
{"id": "cases/21-mobile-onboarding-flow", "group": "cases", "slug": "21-mobile-onboarding-flow", "title": "Mobile Onboarding Flow", "prompt_words": 655, "prompt": "A mobile app onboarding set, 9:16 portrait, presented as three phone screens stacked vertically on a soft warm\ngreige background with even gaps between them, shown straight on with no perspective and no hand.\n\nTHE THREE SCREENS, top to bottom\n- Each screen is a rounded rectangle at the same size and the same width, occupying about 30% of the frame\n  height each, their left and right edges aligned to one vertical axis so the stack reads as a single column.\n- Each screen shows its content only: no device bezel, no notch, no camera cutout, no status bar, no home\n  indicator, no browser chrome. The rounded corners are the only indication that these are screens.\n\nSCREEN 1 — welcome\nA small thin-stroke emblem centred in the upper third: an arch over a horizontal line. Below it, a two-line\nheadline centred. Below that, one short supporting line centred in a lighter weight. At the bottom, a single\nfull-width primary button with rounded ends, filled in the accent colour, with its label centred.\n\nSCREEN 2 — notifications\nA simple line illustration centred in the upper third: a rounded bell with a single small dot badge at its upper\nright, drawn with the same thin stroke as the emblem on screen 1. Below it, a two-line headline centred and one\nsupporting line. At the bottom, the same full-width primary button, and directly above it one text-only\nsecondary link centred in the accent colour.\n\nSCREEN 3 — complete\nA centred mark in the upper third: a thin-stroke circle with a simple check inside it. Below it, a short\nheadline of four words centred, one supporting line, and a small three-row summary list — each row a label at\nthe left and a value at the right, separated by faint rules. At the bottom, the same full-width primary button.\n\nEXACT VISIBLE TEXT — render exactly these strings, once each, in the screen stated\n- Screen 1 headline line 1: \"Money, in one place\"\n- Screen 1 headline line 2: \"No spreadsheets\"\n- Screen 1 supporting line: \"Track every account and payout in a single view.\"\n- Screen 1 button: \"Get started\"\n- Screen 2 headline line 1: \"Know the moment\"\n- Screen 2 headline line 2: \"money moves\"\n- Screen 2 supporting line: \"Get a push when a payment settles or fails.\"\n- Screen 2 button: \"Allow notifications\"\n- Screen 2 secondary link: \"Not now\"\n- Screen 3 headline: \"You are all set\"\n- Screen 3 supporting line: \"Your first sync is already running.\"\n- Screen 3 summary row labels, top to bottom: \"Accounts\", \"Currency\", \"Sync\"\n- Screen 3 summary row values, top to bottom: \"3 linked\", \"EUR\", \"Live\"\n- Screen 3 button: \"Open dashboard\"\nNo other text anywhere — no page dots, no step counters, no \"1 of 3\", no version numbers, no legal line, no\nwatermark.\n\nSTYLE AND MATERIAL\nA light, warm, quiet interface. Ground is a warm off-white; body text is near-black; one accent colour, a muted\ngreen, used only for the three primary buttons, the secondary link and the small bell badge. Type is one\ngeometric sans throughout, in exactly four sizes: headline, supporting line, button label, and the small\nsummary rows. Rounded rectangles with generous internal padding. Buttons are flat fills with no gradient, no\nshadow and no border. Thin line illustrations use a single stroke weight throughout and no fills except the\nbell badge.\n\nCONSTRAINTS\nNo dark mode, no glassmorphism, no neon, no gradient meshes, no photographic imagery, no product screenshots\ninside the screens, no charts or graphs on any screen. Do not add a fourth screen, a floating card, a tooltip,\na notification banner, an avatar or a logo other than the emblem described. Do not rotate, tilt or perspective\nthe screens. Do not paraphrase, re-case or re-punctuate any string, and do not add a full stop that is not\nwritten above. No watermark, no signature, no frame around the whole set.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/21-mobile-onboarding-flow"}
{"id": "cases/22-station-observation-deck", "group": "cases", "slug": "22-station-observation-deck", "title": "Station Observation Deck", "prompt_words": 658, "prompt": "An interior photograph from inside an orbital station's observation deck, 21:9, looking out through a huge\ncurved window at the limb of a planet. Shot on a wide lens from the back of the deck so the window frame, the\ninterior and the view all read at once.\n\nTHE WINDOW\nThe window is the far wall: a single curved pane about four metres tall and twelve wide, held in a deep\nstructural frame. The frame is heavy and layered — an outer structural rib, then a primary seal, then an inner\ntrim panel — and every layer is visible in section where it meets the glass. Running horizontally across the\nglass at waist height is one thin defroster wire, and at its lower edge a narrow metal sill. The glass itself\nis not perfectly clear: a faint double reflection of the interior lights sits across its upper left, and where\nthe deck's warm light is brightest the glass picks up a soft haze.\n\nTHE VIEW\nA planet limb fills the lower two thirds of the window, seen from low orbit: a curved horizon running from the\nlower left to the middle right, cloud bands layered in soft horizontal streaks, and a deep blue ocean under\nbroken cloud toward the terminator. Above the horizon, a thin bright arc of atmosphere — a single luminous\nline of pale blue-white — separates the planet from the black of space. Above that, space is almost black with\na very sparse scatter of stars, no more than thirty visible in the whole pane. In the upper right of the\nwindow, one small airless moon, a plain grey disc with visible craters along its lower edge, no larger than a\nthumbnail at this distance.\n\nTHE DECK\nThe interior is compact and functional. In the foreground, low and slightly out of focus, the top rail of a\nsteel handrail crosses the lower left. Along the left wall, four equipment lockers with louvred vents and\nplain recessed latches. On the ceiling, two continuous strips of warm light running the length of the deck,\ntheir housings visible. The floor is open steel grating. Against the right wall, a narrow upholstered bench\nwith a folded grey blanket and one steel water bottle. In a recessed tray beside the window, a shallow\nhydroponic unit with four small green plants under a single small grow lamp. On the sill by the window, one\nceramic mug, a pair of work gloves and a hand-held tablet lying face down.\n\nTHE FIGURE\nOne person, standing at the centre of the window with their back to camera, small in the frame — about one\nninth of the frame height — silhouetted against the planet, wearing a soft grey flight suit with the hood down,\none hand resting on the sill. No other person anywhere.\n\nLIGHT\nTwo sources only. Inside: the warm amber ceiling strips, which light the deck, the lockers, the bench and the\nfigure's shoulders from above and behind, and put a warm reflection across the upper left of the glass.\nOutside: the cold blue-white light of the planet, which rims the window frame from beyond and lights nothing\ninside the deck. The two temperatures never mix: interiors stay amber, the planet stays cold, and the figure is\nread as a silhouette because the outside is brighter than the inside.\n\nCONSTRAINTS\nNo text anywhere: no signage, no labels on the lockers, no warning placards, no numbers, no stencilled\nlettering, no readable screens, no watermark. Do not add a second figure, a spaceship, a satellite, a docking\nbay, a spacecraft passing by, or a laser. No lens flare, no light beams, no god rays. Do not make space purple,\ndo not add nebulae or coloured stars, and do not fill the window with debris. Keep the verticals of the window\nframe true and the horizon a genuine curve, not a straight line. No HUD overlays, no targeting reticles, no\ngraphics layered over the view.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/22-station-observation-deck"}
{"id": "cases/23-dragon-on-a-ruined-tower", "group": "cases", "slug": "23-dragon-on-a-ruined-tower", "title": "Dragon On A Ruined Tower", "prompt_words": 644, "prompt": "A wide fantasy landscape painting, 16:9, of a dragon perched on the broken crown of a ruined stone tower at\ngolden hour, seen from a distance across a valley so that scale reads before detail does.\n\nTHE TOWER\nA single round tower of dark weathered stone, standing on a rocky spur in the middle distance, its upper third\ncollapsed: the crown is broken open on the near side, the wall's cross-section showing rubble core and double\nskins of squared masonry. Two merlons survive on the far side; the rest is a stepped ruin. A spiral of narrow\nwindows runs up the shaft, dark and empty. Ivy climbs the lower third. The tower is roughly one sixth of the\nframe height, and it is the only building in the picture.\n\nTHE DRAGON\nPerched on the broken crown, gripping the shattered stone with four clawed feet, tail curled around the\nremaining merlon. It is long and lean rather than bulky, with a serpentine neck and a narrow head, and its\nnear wing is half-spread and half-folded, drooping down the outside of the tower so its tip hangs in open air.\nThe far wing is folded along the back. Its head is turned in profile to the left, looking out down the valley.\nIts total body length is about twice the tower's width, so it reads as large without dominating the frame.\n\nTHE VALLEY\nBelow and beyond the tower, a broad valley recedes in five or six distinct layers of depth: a foreground ridge\nin near-black silhouette at the lower right; the spur and the tower; a middle distance of wooded slopes in\nmuted green; a wide river catching the light as a pale ribbon; a far range of hills in cool blue-grey; and a\nfinal band of distant peaks reduced almost to haze. Three separate layers of low mist lie in the folds between\nridges, each flat-bottomed and each paler than the layer behind it. A single narrow road winds through the\nmiddle distance with two tiny switchbacks; no vehicles and no travellers are visible on it.\n\nLIGHT — this is the whole picture\nThe sun is out of frame to the left and very low, just above the horizon line behind the far hills. It throws\nwarm amber light almost horizontally: every left-facing surface catches it, every right-facing surface falls\ninto cool shadow, and the shadows run long toward the right. The dragon is rim-lit along the top of its neck,\nthe leading edge of its wing and the whole length of its tail. The sun's disc itself is not visible; only its\nglow behind the haze of the far range. The mist layers catch the light so each one is warm at its top edge and\ncool underneath. The sky above the hills moves from warm amber at the horizon through pale gold to a cool\nblue-grey at the top of the frame.\n\nWING MEMBRANE\nWhere the near wing is spread against the light, the membrane between the fingers is translucent: the low sun\npasses through it, showing the vein structure as darker lines, and the sky behind glows through the thinner\nareas. Where the membrane overlaps the body, it is opaque and in shadow. This contrast is the brightest and\nmost delicate thing in the picture and must be clearly readable.\n\nCONSTRAINTS\nOne dragon only, one tower only. No rider, no saddle, no harness, no army, no village, no castle, no second\nruin. No text, no banners with lettering, no heraldry, no signature, no watermark. Do not put the sun's disc in\nframe, do not add lens flare, do not add fire or smoke from the dragon. Do not make the dragon metallic or\narmoured; it is a scaled animal with leathery wings. Do not flatten the valley into two layers, and do not\nplace the horizon at the centre of the frame.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/23-dragon-on-a-ruined-tower"}
{"id": "cases/24-wizards-study", "group": "cases", "slug": "24-wizards-study", "title": "Wizards Study", "prompt_words": 915, "prompt": "A dense interior painting of a wizard's study, 3:2, seen from standing height at the doorway looking in, lit by\nlamplight and a fire. This is a maximalist picture: every surface is occupied and the eye should find something\nnew on the fourth look. Forty-four separate objects are listed below and all forty-four must appear.\n\nTHE ROOM\nA stone-walled study about four metres square, one small deep-set leaded window high on the left wall with the\nnight beyond it, a low arched fireplace on the right wall, a timber ceiling with exposed joists, and a floor of\nworn dark boards partly covered by a threadbare patterned rug.\n\nTHE DESK — twelve objects\n1. A long oak desk across the middle of the room, its surface scarred and ink-stained.\n2. A brass desk lamp with a green glass shade, lit, throwing a warm pool onto the desk.\n3. An open leather grimoire on a wooden cradle, its pages yellowed, one red silk ribbon.\n4. A second closed book with brass corner pieces and a clasp, lying flat.\n5. A stack of three smaller books at different angles.\n6. A crystal ball on a brass ring stand, one bright point of light inside it.\n7. A stone mortar with green residue in it and its pestle laid across the rim.\n8. A rack of nine small glass vials, each holding a different coloured liquid.\n9. A brass astrolabe with visible graduated rings.\n10. A scroll in an open leather tube, with a corner of a star chart unrolled beside it.\n11. A quill standing in a brass inkwell, and two loose quills lying beside it.\n12. A pair of brass dividers resting on a wooden ruler.\n\nTHE LEFT SHELF WALL — nine objects\n13. A tall open-fronted shelf unit leaning against the left wall.\n14. A mounted armillary sphere on a wooden base.\n15. The horned skull of a large animal, hung on the wall above the shelf.\n16. A glass jar of dried beetles.\n17. A glass jar of pale root slices.\n18. A brass balance scale with its pan and four stacked weights.\n19. A magnifying lens on a turned handle.\n20. A small hourglass, half run through.\n21. A chipped ceramic bowl holding three large seed pods.\n\nTHE FIREPLACE AND HEARTH — eight objects\n22. The low arched fireplace with a live fire banked low, glowing embers visible.\n23. A black iron cauldron on a three-legged trivet, steam rising from it.\n24. A long iron poker and a pair of tongs leaning against the stone.\n25. A stack of split firewood in a wicker basket beside the hearth.\n26. A string of dried red peppers hanging on the wall beside the chimney.\n27. A brass candelabra with three candles, all lit.\n28. A single tall wax candle standing in a holder on the mantel.\n29. A broken mantel clock, its face open, its hands stopped.\n\nTHE CEILING AND CORNERS — seven objects\n30. A rack of dried herbs hanging head-down from a ceiling joist.\n31. A bundle of dried roots suspended beside them.\n32. A wicker basket of mushrooms on the floor beneath the herbs.\n33. A birdcage on a stand with a small sleeping songbird inside.\n34. An owl perched on a wall bracket, awake and watching.\n35. Cobwebs filling the upper right corner.\n36. Dust suspended in the shaft of light falling from the high window.\n\nTHE FLOOR AND FURNITURE — eight objects\n37. A tall-backed wooden chair with one elbow worn through, pulled back from the desk.\n38. A wingback armchair beside the fire with a patchwork blanket on it.\n39. A grey cat asleep in the armchair.\n40. A small iron-bound chest against the wall, its lid closed.\n41. A padlocked strongbox sitting on top of the chest.\n42. A lute leaning in the far corner beside the window.\n43. A spilt pile of white salt on the boards, swept halfway into a drift.\n44. A chalk diagram, half erased, drawn on the boards beside the desk.\n\nWALLS AND FURNISHINGS\nA framed chart of an unfamiliar coastline hangs on the wall behind the desk. A heavy tapestry, rolled and\nleaning in the corner, its colours faded. A rack of long wooden spoons and a small teapot on a side table.\n\nLIGHT AND PALETTE\nThree sources only: the lamp with its warm green-tinted pool on the desk; the fire, low and orange, from the\nright; and one cold shaft of moonlight through the high window, striking the floor and the board wall on the\nleft. Each source keeps its own colour and its own shadow direction: lamp from above the desk, fire from the\nright at knee height, moon from upper left. The palette is warm — amber, oak, brass, deep greens and blacks —\nwith the cold moon shaft and the crystal ball's point as the only cool notes.\n\nCONSTRAINTS\nAll forty-four listed objects must be present, each visible and identifiable, and no object may appear twice.\nNo text anywhere: no legible lettering on any book page, spine, chart, clock face or label, and no runes,\nsigils or symbols anywhere. Do not add a second person; the room is empty of people. Do not add a second cat,\na second owl or a second bird. Do not add hanging chains, floating objects, spell effects, glowing runes or any\nmagical light. No watermark. Keep the room believable: every object rests on a surface or hangs from a fixing,\nand nothing floats.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/24-wizards-study"}
{"id": "cases/25-five-zone-explainer", "group": "cases", "slug": "25-five-zone-explainer", "title": "Five Zone Explainer", "prompt_words": 790, "prompt": "A fixed-zone explainer infographic, 4:3 landscape, titled \"How a Hydroelectric Dam Works\", drawn as flat vector\nart with no photograph, no 3D render and no perspective beyond the cutaway.\n\nZONE 1 — TITLE BAND, top 12% of the frame height\nThe title \"HOW A HYDROELECTRIC DAM WORKS\" set left aligned in a bold condensed sans, all capitals, with the\nsubtitle \"From falling water to your light switch\" beneath it in a lighter weight at roughly a third of the\ntitle's size. A single hairline rule closes the band across the full width.\n\nZONE 2 — MAIN DIAGRAM, left 62% of the width, from below the title band down to the process strip\nA cutaway cross-section of a concrete gravity dam seen from the side, with the reservoir behind it on the left\nand the river valley below on the right. Within this zone:\n- The reservoir: a large body of water filling the left third, its surface a flat pale blue, with a shoreline\n  of rock and pine trees along its upper edge.\n- The dam wall: a massive grey concrete wedge, widest at the base, with visible horizontal construction\n  joints. Its crest carries a low parapet and a gantry crane on rails.\n- The intake: a rectangular opening in the dam's upstream face just below the waterline, with a trash rack of\n  vertical bars across it and a steel gate above it.\n- The penstock: a thick steel pipe running down through the body of the dam at a steep angle, drawn in\n  section so its wall thickness reads, ending at the turbine below.\n- The powerhouse: a rectangular hall cut into the base of the dam on the downstream side, containing a\n  vertical turbine and, directly above it on the same shaft, a cylindrical generator.\n- The draft tube: a curved widening passage carrying water away from the turbine out into the river.\n- The switchyard: two transformers and a row of insulators on a flat platform beside the powerhouse, with\n  three transmission pylons stepping away to the right.\n- The spillway: a separate overflow channel at the dam crest on the right, dry, with a stepped chute.\nWater direction is shown by five small white arrows: into the intake, down the penstock, through the turbine,\nout of the draft tube and away downstream.\n\nZONE 3 — CALLOUT COLUMN, right 34% of the width, beside the diagram\nThree stacked panels of equal height, each with a coloured circular number badge at its top left, one line of\nbold label text, and two lines of smaller explanatory text beneath. A thin leader line runs from each badge to\nits element in the diagram, and no leader crosses another.\n\nZONE 4 — PROCESS STRIP, bottom 16% of the frame height, full width beneath both columns\nFour equal boxes in a row, each with a coloured square icon at its left and one short label to the right of it,\njoined by three right-pointing chevrons between the boxes.\n\nZONE 5 — LEGEND, bottom right corner, inside the process strip's row\nThree legend entries stacked vertically, each with a short coloured line sample at the left and one short label\nbeside it.\n\nEXACT VISIBLE TEXT — render exactly these strings, once each\n- Title: \"HOW A HYDROELECTRIC DAM WORKS\"\n- Subtitle: \"From falling water to your light switch\"\n- Callout 1 label: \"RESERVOIR\" and its text: \"Water is stored at height, holding potential energy.\"\n- Callout 2 label: \"PENSTOCK\" and its text: \"The pipe converts that height into pressure and speed.\"\n- Callout 3 label: \"TURBINE\" and its text: \"Moving water spins the turbine and the generator above it.\"\n- Process boxes, left to right: \"Store\" / \"Release\" / \"Spin\" / \"Generate\"\n- Legend entries, top to bottom: \"Water\" / \"Concrete\" / \"Electricity\"\nNo other text anywhere: no units, no measurements, no capacity figures, no equations, no logos, no watermark.\n\nSTYLE\nFlat vector illustration: solid fills, uniform hairline outlines, no gradients, no drop shadows, no texture.\nBackground flat white. Palette: concrete grey, water blue, a warm amber for electricity, and one accent colour\nused only on the three number badges and the three chevrons. One neutral sans throughout, in exactly four\nsizes: title, subtitle, labels, and explanatory text.\n\nCONSTRAINTS\nExactly three callouts, exactly four process boxes, exactly three legend entries, exactly five direction\narrows, exactly three pylons. Every leader line must end on the element it names and must not cross another\nleader, a callout panel or any part of the diagram. Do not add a second dam, a fish ladder, a boat, a person,\na bird, a road or a town. Do not put text on any part of the diagram itself. Do not use 3D shading, isometric\nprojection or perspective.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/25-five-zone-explainer"}
{"id": "cases/26-sixteen-pose-contact-sheet", "group": "cases", "slug": "26-sixteen-pose-contact-sheet", "title": "Sixteen Pose Contact Sheet", "prompt_words": 675, "prompt": "A character pose contact sheet, 1:1 square, sixteen panels in a 4 by 4 grid with even gutters and a thin dark\nouter border. Each panel shows the same person in a different standing pose, photographed on a plain mid-grey\nseamless with flat even lighting. This is a reference sheet, not a beauty photograph.\n\nTHE CAMERA — identical in all sixteen panels\nOne camera, one position, never varied: eye level at 60% of the figure's height, straight on, 85 mm equivalent,\nno wide-angle distortion, no tilt, no zoom change, no perspective change between panels. Because the camera and\nthe distance never move, the ground line falls at exactly the same height in all sixteen panels and the top of\nthe head falls at exactly the same height in all sixteen panels. The figure is centred horizontally in every\npanel with even margins left and right, and the feet are inside the panel in all sixteen.\n\nTHE CHARACTER — identical in all sixteen panels\nA woman of about thirty, athletic build, medium height. Short dark hair cut to the jaw with a blunt fringe, a\nthin scar on her LEFT forearm about four centimetres long, and a small mole on her RIGHT cheekbone. She wears\na plain charcoal cotton t-shirt with a crew neck, olive cargo trousers with a single pocket flap on the LEFT\nthigh only, and tan canvas boots laced to the ankle. Her clothing, hair, scar and mole must be identical in\nevery panel, and the pocket flap must stay on the left thigh in all sixteen — it must never appear on the right.\n\nTHE SIXTEEN POSES, reading left to right then top to bottom\n1. Neutral stance, feet shoulder-width apart, arms hanging at her sides.\n2. Weight on the left leg, right knee slightly bent and relaxed, arms at her sides.\n3. Feet together, hands clasped in front of her at waist height.\n4. Feet apart, arms folded across her chest.\n5. Left hand on her hip, right arm hanging down.\n6. Right hand on her hip, left arm hanging down.\n7. Both hands on her hips, elbows out, feet apart.\n8. Left arm raised forward, pointing at chest height, right arm down.\n9. Right arm raised straight up above her head, left arm down.\n10. Both arms raised out to the sides at shoulder height, palms down.\n11. Leaning forward from the hips, both hands resting on her thighs.\n12. Standing turned a quarter away, looking back over her right shoulder.\n13. Half crouch, knees bent, weight low, both arms held forward.\n14. Standing up on the balls of both feet, heels off the ground, arms swung back.\n15. Right foot forward a full stride, torso twisted to the left, arms loose.\n16. Standing straight, both hands clasped behind her back.\n\nPANEL NUMBERS\nA small numeral in the bottom left corner of each panel, reading 1 to 16 in order, all the same size and in the\nsame position within their panels. These numerals are the only text on the sheet.\n\nLIGHTING\nOne large soft source directly in front of the figure and slightly above, plus a soft fill from both sides, so\nthe lighting is even and flat and identical in all sixteen panels. One short soft shadow directly beneath each\nfigure, the same shape and length everywhere. No dramatic lighting, no coloured light, no rim light, no change\nof exposure between panels.\n\nCONSTRAINTS\nAll sixteen panels are present and all sixteen poses are different. No sitting, kneeling, lying, jumping or\nairborne pose: both feet are on the ground in every panel. Do not hide the feet, do not crop the head, and do\nnot let the figure leave the panel. Do not mirror the character's asymmetric details between panels — the scar\nstays on the left forearm and the mole stays on the right cheekbone in all sixteen. Do not add props, weapons,\nbags, hats, furniture or a background other than the plain seamless. No text anywhere except the sixteen panel\nnumerals, and no number larger than 16.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/26-sixteen-pose-contact-sheet"}
{"id": "cases/27-16bit-town-at-dusk", "group": "cases", "slug": "27-16bit-town-at-dusk", "title": "16Bit Town At Dusk", "prompt_words": 651, "prompt": "Pixel art in the style of a 16-bit console role-playing game, 16:9 landscape, showing a small walled town at\ndusk seen from a slight raised angle, as a single game screen.\n\nTHE PIXEL GRID — this is the whole style\nThe image is built from large, visible square pixels at one consistent size across the entire frame, roughly the\nscale of a 320 by 180 pixel image enlarged for display. Every edge in the picture is a hard staircase of whole\npixels: there is no anti-aliasing, no sub-pixel smoothness, no soft blending and no blurred edge anywhere.\nDiagonal edges step in visible blocks. Curves are chunky. Nothing in the image is smaller than one pixel, and\nnothing has a detail finer than the pixel grid. Colour transitions happen by dithering — ordered checkerboard\nand scattered-dot patterns of two palette colours — never by a smooth gradient.\n\nPALETTE — exactly twenty-four colours and no more\nA dusk palette of twenty-four fixed colours: four blues for the sky, two deep purples for the far hills, three\ngreens for foliage, two browns for wood and earth, three greys for stone and the road, four warm ambers and\noranges for lit windows and lanterns, one off-white for highlights, one near-black for outlines and one muted\nred for roof tiles. Every colour in the image is one of these twenty-four; no colour is blended, faded or\ntinted outside the palette.\n\nTHE TOWN\nA low stone wall with a square gatehouse runs across the middle of the frame, the gate standing open. Behind it,\nfifteen or so small buildings with steeply pitched red-tiled roofs crowd together along two narrow streets;\nwindows are small and most of them are lit warm amber. A square tower with a single clock face stands above the\nroofs at the left, and a windmill turns slowly on a low hill at the right. A cobbled road enters through the\ngate and runs toward the viewer, widening in the foreground. A stone well with a wooden frame stands in the\nsmall square just inside the gate.\n\nTHE LANDSCAPE\nBehind the town, three stepped bands of hills recede in increasingly dark purples, undulating rather than\nsmooth. On the far left, a river runs down from the hills and passes under a two-arch stone bridge, its surface\ndrawn as horizontal bands of two blues with dithered highlights, and it catches a few dither-speckled glints of\nthe dying light. Three conical pine trees stand on the near bank and four on the far bank, all the same size\nand drawn from the same four or five pixels of shape.\n\nTHE SKY\nLate dusk: the sky is a vertical gradient built entirely from dithered bands of the four blues, from a deeper\nblue at the top to a warm orange band at the horizon, with no smooth blend anywhere. Eight small stars are\nplaced in the upper third, each a single off-white pixel with no glow. One low cloud sits in the middle of the\nsky, a flat two-colour shape with a dithered lower edge.\n\nTHE FIGURE\nA single small sprite of a traveller stands on the cobbled road just inside the gate, facing away from the\nviewer. The sprite is about twelve pixels tall, drawn from the same palette, with a visible dark outline around\nits whole silhouette. It is the only figure in the scene and casts no shadow.\n\nCONSTRAINTS\nThe whole image must read as pixel art, not as a photograph or painting with a pixel filter applied over it:\nno photographic texture, no painted brushwork, no smooth gradients, no 3D rendering, no depth of field. No text\nanywhere: no signposts, no shop signs, no town name, no window lettering, no watermark. Do not add a second\nfigure, animals, carts, vehicles, laundry, banners or smoke. Do not add modern elements. Keep the horizon\nstraight and place it in the upper third of the frame.", "images": ["full.jpg", "output.png", "preview.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/27-16bit-town-at-dusk"}
{"id": "cases/28-product-on-a-new-background", "group": "cases", "slug": "28-product-on-a-new-background", "title": "Product On A New Background", "prompt_words": 583, "prompt": "Image 1: the product — the sole authority for the jar's form, its proportions, its glass thickness and base, its\nbronze knurled lid, the level and colour of the honey inside, the paper label, and every character printed on\nit. It supplies no setting, no camera and no lighting.\n\nTASK\nKeep the jar exactly as it is and re-photograph it on a weathered oak board in a kitchen, keeping the label\nsquare to the camera and fully legible.\n\nPRESERVE EXACTLY\n- The jar's geometry: squat proportions, wide shoulder, short neck, the thickness of the glass at the base, and\n  the lid's diameter relative to the body.\n- The lid: dark bronze, the fine knurled band around its edge, and its position squarely on the neck.\n- The contents: the same deep amber honey at the same level, with the same cloudy depth and the same\n  glass-to-honey edge visible in the shoulder.\n- The label: the same off-white uncoated paper, the same wrap and registration to the jar's vertical axis, the\n  same uneven cut edge, and the same text in the same places — \"HEATHER HONEY\" on the upper third in\n  wide-tracked uppercase serif, one thin horizontal rule below it, and \"340 G\" small and centred at the bottom.\n  Both strings must render once, crisply, and unobstructed.\n- The jar's silhouette. It must not be restyled, made taller, made into plastic, or given a different lid.\n\nTHE NEW SETTING\nA worn oak board, its surface scarred and slightly uneven, running across the lower third of the frame. To the\nright of the jar, a folded unbleached linen cloth. To the left of the jar, lying flat on the board, one small\nsprig of dried heather. Behind, out of focus, a plain warm plaster wall.\n\nMATCH NATURALLY\n- Light direction: one window off-frame to the LEFT, so the jar is lit from the left. Its lit side is on the\n  left, its shadow falls to the right, and the linen cloth's fold shadows agree with the same direction.\n- Contact shadow: one soft-edged shadow beneath the base, short and to the right, denser where the base meets\n  the board and fading outward. The jar must sit on the board, not float above it.\n- Reflection: a faint reflection of the jar's base and lower body in the polished parts of the oak immediately\n  beneath it, weaker than the reflection of the jar in Image 1.\n- Glass behaviour: the same internal refraction and the same dark edge where the glass meets the surface below.\n- Colour temperature: the morning window light is warm and the shadows are cool, applied to the jar, the cloth\n  and the heather without changing the honey's own amber.\n- Scale: the jar remains the largest object in the frame, and the cloth and heather are both clearly smaller\n  than it.\n\nDO NOT\n- Do not redesign, restyle, recolour or reshape the jar, the lid or the label, and do not change a single\n  character on the label.\n- Do not add a second jar, a spoon, a dipper, toast, a knife, flowers, honeycomb, a bee, a cup or a hand.\n- Do not paste the jar flat: no outline, no cut-out edge, no uniform drop shadow, no missing contact shadow,\n  no flat colour fill, no ghost halo around the silhouette.\n- Do not apply a shadow or reflection that disagrees with the light coming from the left.\n- No text anywhere except the two label strings, no logo, no watermark.", "images": ["full.jpg", "output.png", "preview.jpg", "reference-1.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/28-product-on-a-new-background"}
{"id": "cases/29-photo-to-watercolor", "group": "cases", "slug": "29-photo-to-watercolor", "title": "Photo To Watercolor", "prompt_words": 772, "prompt": "Image 1: the photograph — the sole authority for the composition, the pose, the framing, the direction of the\nlight and the identity of the person. It supplies no paint, no brushwork and no paper.\n\nTASK\nRepaint Image 1 as an original watercolour on rough cold-pressed paper, at the same composition and the same\nframing, so that it reads as a painted study made from the photograph rather than a photograph with a filter\nover it.\n\nTHE PERSON — preserve her exactly as she appears in Image 1\n- Her hair is worn LOOSE and down, falling just past her jaw with a slight wave, with several strands lying\n  across her cheek and one behind the ear. Her hair is NOT tied up: do not add a bun, a twist, a knot, a\n  ponytail or a hair clip, and do not lift it off her neck.\n- Her clothing is exactly as in the photograph: a pale sage linen shirt with the collar open, sleeves rolled to\n  the elbow, tucked into DARK charcoal trousers. The trousers are dark. Do not lighten them, do not make them\n  cream, beige or stone, and do not change them to a skirt.\n- Both arms hang relaxed at her sides and BOTH HANDS ARE EMPTY AND VISIBLE, held slightly away from the body.\n  Do not put a hand in a pocket, do not cross her arms, and do not hide either hand.\n- The only bag in the picture is the thin dark leather STRAP crossing her right shoulder, visible where it\n  passes over the shirt. There is no bag body visible, no tote bag, no canvas bag and no second strap.\n- Her face: the same structure, the same narrow jaw, the same straight nose and eyebrow shape, and the same\n  quiet expression with the mouth closed. Her gaze is directed off to the left of the frame.\n\nPRESERVE — the setting and the light\n- The wall runs from the near left into the distance at the same angle, and it is RED-BROWN BRICK with visible\n  courses and mortar lines, with moss at its base. It is brick, not stone, not rough rubble and not concrete.\n- The planting along the wall's base, the gravel path receding behind her, and the tops of the trees above the\n  wall line in the same place.\n- The sun comes from behind and to her left exactly as in Image 1: the rim of her shoulder, the edge of her hair\n  and the fold of her sleeve are the brightest edges, and the long shadows fall on the gravel toward the right.\n- The gravel is neutral grey and the brick is warm red-brown.\n\nCHANGE — the medium, completely\n- Paint, not pixels. Loose wet-in-wet washes laid in broad passages, pigment blooming where two washes meet,\n  granulation settling visibly at those boundaries, and the paper's tooth showing through every wash.\n- Several edges deliberately lost: the far end of the wall and the distant treetops dissolve into the paper with\n  no line at all, and the gravel path fades out rather than being drawn.\n- Whites are unpainted paper. There is no white body paint anywhere, and the brightest values come from areas\n  where nothing was applied.\n- A few dry-brush skips across the grain, and one or two places where the paper was left to dry before the next\n  wash, so the boundary is crisp and irregular rather than soft.\n- Palette limited to five pigments: warm ochre, burnt sienna, cool grey-blue, soft sage for the shirt, and one\n  deep indigo for the darkest notes in her hair and the wall shadow.\n- Figures are suggested, not detailed: her face is a small number of carefully placed shapes with the features\n  read from value rather than from line, and her hands are simple masses with no drawn fingers.\n\nDO NOT\n- Do not restyle her hair, her clothing or her pose in any way. Do not add a bag, a hat, a scarf or sunglasses.\n- Do not make it photographic: no sharp photographic detail, no photographic grain, no lens blur, no depth of\n  field, no noise and no digital gradient.\n- Do not smooth her face into a generic pretty face, and do not change her apparent age.\n- Do not change the direction of her gaze or the direction of the light.\n- Do not add people, animals, a gate, a doorway, a bench, a bicycle or flowers in her hands.\n- Do not sign it, date it, add a stamp or a watermark, and do not paint a white border around the picture.", "images": ["full.jpg", "output.png", "preview.jpg", "reference-1.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/29-photo-to-watercolor"}
{"id": "cases/30-character-across-scenes", "group": "cases", "slug": "30-character-across-scenes", "title": "Character Across Scenes", "prompt_words": 792, "prompt": "Image 1: the person — the sole authority for who she is: her face, her age, her hair and the braid, her glasses,\nher freckles, the scar on her jaw, and the clothing she wears. It supplies no pose, no place and no light.\nImage 2: the pose — the sole authority for how she is arranged: the seated posture, the turn of the body, the\nposition of the limbs, the position of the head, and the framing of the shot. It supplies no identity and no\nenvironment.\nImage 3: the place and the light — the sole authority for the space, the camera position within it, the scale of\neverything, the atmosphere and the direction and quality of every light. It supplies no person.\n\nTASK\nPaint the botanist from Image 1 seated in the pose from Image 2 on the low stone abutment of the footbridge in\nImage 3, lit by that clearing's dawn light.\n\nCONFLICT POLICY — read this before composing\n- Image 1 wins on identity and on the appearance of her clothing, hair, glasses and the braid. Her freckles,\n  the scar and the braid's side must not change.\n- Image 2 wins on the pose and on where the frame is cut, so both feet stay inside the picture.\n- Image 3 wins on everything about the world: the depth of the fog, the wetness of the ground, the direction of\n  the light and the colour of the atmosphere.\n- Where Image 1 and Image 3 disagree about colour, Image 3 wins on the light falling on her and Image 1 wins on\n  the hue of the garment itself: a sage shirt stays sage, but its shading is derived from the clearing.\n- Where Image 2 and Image 3 disagree about scale, Image 3 wins, so the bridge and the trunks keep their size\n  and the figure is scaled to them.\n\nPRESERVE EXACTLY\n- Her face: the same structure, the same freckles across the nose and both cheekbones, the same round wire\n  glasses, and the braid over the same shoulder ending at the same length.\n- Her clothing: the same dark green canvas jacket with its corduroy collar, worn open at the throat over the\n  same grey shirt, and the same dark trousers.\n- The pose: seated, body turned a quarter to the right, leaning forward with both forearms on the thighs, left\n  foot flat on the ground, right foot slightly back, head turned to look off to the right.\n- The place: the clearing's pine trunks receding into fog, the needle floor, the ferns, the shallow stream, the\n  timber footbridge on its two low stone abutments, and the fog thickest at about eight metres' depth.\n- The light: one low sun outside the frame to the upper left, filtered through fog so it is diffuse and\n  volumetric rather than hard, with everything near the light band warm and everything nearer and further away\n  cool blue-grey.\n\nMATCH NATURALLY\n- She sits ON the stone abutment: her weight compresses the moss there, and her shadow falls on the stone\n  beneath her, agreeing with the light coming from the upper left.\n- Her shadow joins the long trunk shadows already running toward the lower right; she casts no shadow of her own\n  in a different direction.\n- Fog behaves with depth: she sits behind the nearest fog and in front of the thickest band, so her silhouette\n  is slightly softened at its edges and the ground between her and the camera is a little lighter than the\n  ground beyond her.\n- Her jacket's damp-weather behaviour matches the clearing: canvas takes on a faint sheen where the fog settles\n  and a slight darkening at the shoulders.\n- Colour temperature: her rim light is warm pale gold because the sun is behind and above her left shoulder;\n  everything facing the camera is lit only by the cool ambient fog, so her face is cool and her edges are warm.\n- Grain, focus and contrast identical to the surrounding scene.\n\nDO NOT\n- Do not change her identity, her age, her build, or the side the braid falls on.\n- Do not change the pose, add a hand on a rail, straighten her spine, or stand her up.\n- Do not change the bridge, the stream, the trunks, the fog's depth, the camera height or the direction of the\n  light.\n- Do not add a second person, a dog, a backpack, a tripod, a notebook, a campfire, a tent or litter.\n- Do not introduce hard sun shafts or god rays; the light is diffuse in fog.\n- Do not add colour accents that are not in Image 3, and do not warm the whole frame.\n- No text, no signage, no watermark.", "images": ["full.jpg", "output.png", "preview.jpg", "reference-1.jpg", "reference-2.jpg", "reference-3.jpg"], "source": "https://github.com/callirra-ai/gpt-image-2-5-prompt-atlas/tree/main/cases/30-character-across-scenes"}