alvdansen/illustration-1.0-qwen-image

Illustration 1.0 — Qwen-Image

A style LoRA for Qwen-Image trained on 244 curated illustration and anime reference images across diverse visual styles. Produces images spanning cute anime character design, European bande dessinée, indie risograph prints, storybook watercolor, retro shoujo manga, and graphic novel illustration.

Published alongside its straight-baseline twin as the paired comparison in Forgetting on Purpose: Generalization as the Quality Criterion for Small-Dataset LoRA Fine-Tuning — Alvdansen Labs, May 2026. Read the paper · Source on GitHub.

Style Influences

• Japanese anime and manga illustration — clean cel shading, expressive character design, shoujo and slice-of-life aesthetics
• European graphic novel and bande dessinée — ink crosshatching, atmospheric landscapes, ligne claire
• Indie illustration and zine culture — risograph print textures, limited palettes, grainy halftone
• Children's book and storybook illustration — watercolor washes, warm palettes, charming character proportions
• Retro anime aesthetics — 80s/90s anime film stills, VHS grain, bold color blocking
• Concept art and character design — flat color fills, turnaround sheets, fashion illustration

Usage

No trigger word — this is a style LoRA. Describe what you want and add style cues like "ink and watercolor", "clean cel shading", "risograph print" to steer the output.

Recommended Inference Settings

Sampler: euler
Scheduler: simple
CFG: 3.5
Steps: 45 (30–60 works well)
LoRA strength: 0.8–1.0

Sample Generations

	Prompt
	close-up, a girl with short messy silver hair and round glasses, wearing a chunky knit turtleneck sweater, one hand tucking hair behind her ear, looking slightly past the viewer with half-closed eyes, anime style
	a girl with short hair in a bomber jacket leaning against a wall, clean cel shading, bold graphic composition, 90s ranma era anime, film grain
	a wanderer approaching a stone gate in the desert, european graphic novel, detailed ink hatching, warm sand tones, moebius style
	a fox sleeping in a hollowed-out log, children's book watercolor, soft wet-on-wet washes, autumn leaf palette
	portrait, a girl with star-shaped hair clips, bold graphic shapes, limited three-color palette, screen print flatness, harajuku fashion illustration
	a cat wearing a tiny cape perched on a fence post, indie risograph print, two-color teal and coral, grainy paper texture
	an astronaut sitting on a rocky surface with a small robot, retro watercolor, warm olive and cream tones, hand-painted feel
	wide shot, a girl on a bicycle coasting downhill, lofi film still, clean cel shading, golden hour warmth, anime style
	wide shot, a lone figure on a cliff overlooking the sea with seagulls, bande dessinée, fine ink hatching, muted blue-gray
	portrait, a girl with flowers growing from her hair, risograph print, three-color pink blue and cream, grainy texture
	a boy fixing a radio on a cluttered workbench, warm tungsten light, retro gouache illustration, ochre and burnt sienna palette
	close-up, a girl with braids and paint-stained fingers holding a sketchbook to her chest, soft cel shading, 90s anime, faded VHS warmth
	wide shot, a girl asleep in a hammock strung between two bookshelves, cozy interior light, retro watercolor, cream and dusty gold tones, hand-painted feel

Training Details

• Base model: Qwen-Image (FP8 quantized, text encoder FP8)
• Training steps: 59,000, across a 5-phase chained dataset schedule (four disjoint subsets trained sequentially, then the full combined dataset reintroduced for consolidation)
• Rank/Alpha: 42/42
• Learning rate: 5e-5
• Optimizer: AdamW 8-bit
• Caption dropout: 0.35
• EMA: enabled (decay 0.99)
• Noise scheduler: flowmatch
• Precision: bf16 with qfloat8 quantization
• Dataset: 244 curated images
• Trainer: ai-toolkit by Ostris
• Hardware: NVIDIA RTX 6000 Ada (A6000, 48 GB VRAM)

Full configuration in Appendix A of the paper.