Upload 1391 files

README.md

@@ -42,8 +42,11 @@ This release is a paired, deterministic comparison of eight Krea 2 Turbo checkpo
 1. **BF16** is the highest-fidelity reference because it is the unquantized published checkpoint.
 2. **INT8 ConvRot** narrowly leads the preregistered LPIPS-Alex endpoint. **GGUF Q8_0** is statistically tied with it on that endpoint (Holm-adjusted permutation p=0.9067) and leads INT8 on DISTS, DINO similarity, and reconstructed-weight SNR.
 3. **GGUF Q8_0** is the highest-fidelity of the three added formats.
-4. **GGUF Q4_K_M** is a storage-quality compromise: 6.972 GiB and much closer to BF16 than INT4 ConvRot, but slow on this Ada GPU under ComfyUI-GGUF dequantized execution.
-5. **INT4 ConvRot W4A4** is the smallest checkpoint and fastest of the three additions on this RTX 4060 Ti, but has the largest measured fidelity loss.
+4. **MXFP8** ranks fourth in fidelity (LPIPS 0.071229) and halves checkpoint size versus BF16, but its native SM 10.0 fast path was unavailable on this Ada GPU, resulting in 31.929 s fallback sampling.
+5. **FP8 Scaled** ranks fifth (LPIPS 0.093701) at 12.239 GiB and was the fastest of the original five formats at 19.503 s on this GPU.
+6. **GGUF Q4_K_M** is a storage-quality compromise: 6.972 GiB and much closer to BF16 than INT4 ConvRot, but slow on this Ada GPU under ComfyUI-GGUF dequantized execution.
+7. **NVFP4** is 7.147 GiB with moderate fidelity loss (LPIPS 0.205124). Like MXFP8, its native fast multiplication requires SM 10.0; the recorded 24.925 s sampling time uses the fallback path.
+8. **INT4 ConvRot W4A4** is the smallest checkpoint and fastest of the three additions on this RTX 4060 Ti, but has the largest measured fidelity loss.
 
 No weighted composite score is used. See [the full technical report](TECHNICAL_REPORT.md) and [`tables/decision_table.csv`](tables/decision_table.csv).
 

checksums/SHA256SUMS

@@ -3,7 +3,7 @@
 afbd7232877f7fb929d8e7a2b3b396a41179a3106ee6c1c99f90aa470062f969  CITATION.cff
 fe888a3acbf08a05c3e976f76aa635f2a7034d09589955117c890f3f3a05c0a6  LICENSE-CODE
 10232ffb753de2527c89b1f4a69bb1479e846892c3b3e26a8550866d0f0fe440  LICENSE-DATA
-f1469ecd0c5758c1436a286af0134c11c47a670bfaf06ce389e9f728c4bd98e3  README.md
+1746cf85f90d495021261f7207ab63ee725f0597422739a80bc0ab02ab492ff4  README.md
 7860f02cff561fc62a93cabc452ea76a20452063e8207f97fecedb1a74fbadb2  TECHNICAL_REPORT.md
 f020293711e8ad7eab2758442e5479b39648cd6c725fd076403f66a029c60b5c  comparison_sheets/complete.json
 d71bfb4a773b3998636d68998ad885b52907b285a703cb2348976f4b613ba4de  comparison_sheets/contact_sheet_replicate0.png
@@ -1350,13 +1350,15 @@ ac62c40c4d2c67eea2f9d75d6a5b5ca49adee92159566f0244f1914079ee46a3  reproduction/b
 1b7617cd23d22cca2eacfabd42702c4e9cdaa38d52ad5544b397c4a50eab6fd8  reproduction/benchmark/requirements-analysis.txt
 ac7ecf2619f7e8c8b0c6e1a7b80358bc31bfa1422eeae178e9a52b7063076572  reproduction/benchmark/run_benchmark.ps1
 eafc0cd5d02d6ba78f230bd1143d82e709ff18581687fa9bc98ef82067e296c2  reproduction/benchmark/run_benchmark.sh
-5af229237f69e422eeba4af6072f8a8f3785302c275b5b2cd4642c6e3012d360  reproduction/benchmark/verify_results.py
+3134c540c001e898bbb7592faa12a13664f12a6d645a65e4a854bf9d67e6680f  reproduction/benchmark/verify_results.py
 45424545b70d1b53f9f4e1e401029f24c794b7ad39821df97794b0b7b8fa79e8  reproduction/benchmark/weight_audit.py
 2b9556cb65fa4778117dd8b4ecb0680a2f2d8998c67f1b8418f4748b9008b3b1  reproduction/benchmark/workflows/krea2_benchmark_api.json
+b2482808ae8ad7dc723bbee028d76650909449bf03b19118f046a3e9f0721ea7  reproduction/benchmark/workflows/krea2_benchmark_api_gguf.json
 4315ea5adc5f369d91aa1b654dba0ff641068de79ac5d121d201ee7c211fc38e  reproduction/benchmark/workflows/krea2_benchmark_interactive.json
+e77eff0f644162b32fdbb19e0787b1381a007b41897a4bacb9b47c18791b6a4d  reproduction/benchmark/workflows/krea2_benchmark_interactive_gguf.json
 150a133f4ff5508ca9828fd436fd3db854fbc7218c8fcd9217395eabdb51e3b3  reproduction/download_models.py
 ac43fe8c1625b5754ab3d829591758a50607320f4f694c32784e5ce026c2bc59  reproduction/requirements-release.txt
-8229ef539c836356036804a00ef32015c1b71c01af091c6b31d2051dc554a50f  scripts/prepare_release.py
+77b872302f3de58416d29954cd6b3b863f5388776795a60410bfa4ba3e44724b  scripts/prepare_release.py
 15a332d8bd9031c64539d7bd64df1f55619987ae9e07d06540585fc302d8ac82  scripts/upload_to_huggingface.py
 55c4f6d373d3fe4e7fde9e7c8a88c142de15a0dbaedc8f7be91f3cd997a77ce9  scripts/validate_release.py
 99eb77d4339d532e6fc11cfd50cd55b4f8db54f58f1ddcd93cf74361a9252012  tables/decision_table.csv

reproduction/benchmark/verify_results.py

@@ -134,6 +134,12 @@ def verify(config: dict[str, Any], results_dir: Path) -> dict[str, Any]:
             require(phrase in report, f"Technical report is missing required section/claim: {phrase}")
 
     workflows = [Path(config["_config_path"]).parent / "workflows" / "krea2_benchmark_api.json", Path(config["_config_path"]).parent / "workflows" / "krea2_benchmark_interactive.json"]
+    has_gguf = any(fmt.startswith("gguf") for fmt in config.get("formats", {}))
+    if has_gguf:
+        workflows.extend([
+            Path(config["_config_path"]).parent / "workflows" / "krea2_benchmark_api_gguf.json",
+            Path(config["_config_path"]).parent / "workflows" / "krea2_benchmark_interactive_gguf.json"
+        ])
     for workflow in workflows:
         require(workflow.is_file(), f"Workflow is missing: {workflow.name}")
         if workflow.is_file():

reproduction/benchmark/workflows/krea2_benchmark_api_gguf.json

@@ -0,0 +1,81 @@
+{
+  "1": {
+    "class_type": "UnetLoaderGGUF",
+    "inputs": {
+      "unet_name": "krea2_turbo-Q8_0.gguf"
+    }
+  },
+  "2": {
+    "class_type": "CLIPLoader",
+    "inputs": {
+      "clip_name": "qwen3vl_4b_bf16.safetensors",
+      "type": "krea2",
+      "device": "default"
+    }
+  },
+  "3": {
+    "class_type": "VAELoader",
+    "inputs": {
+      "vae_name": "qwen_image_vae.safetensors"
+    }
+  },
+  "4": {
+    "class_type": "CLIPTextEncode",
+    "inputs": {
+      "clip": ["2", 0],
+      "text": "benchmark prompt"
+    }
+  },
+  "5": {
+    "class_type": "ConditioningZeroOut",
+    "inputs": {
+      "conditioning": ["4", 0]
+    }
+  },
+  "6": {
+    "class_type": "EmptyLatentImage",
+    "inputs": {
+      "width": 1024,
+      "height": 1024,
+      "batch_size": 1
+    }
+  },
+  "7": {
+    "class_type": "KreaBenchmarkSampler",
+    "inputs": {
+      "model": ["1", 0],
+      "seed": 0,
+      "steps": 8,
+      "cfg": 1.0,
+      "sampler_name": "euler",
+      "scheduler": "simple",
+      "positive": ["4", 0],
+      "negative": ["5", 0],
+      "latent_image": ["6", 0],
+      "denoise": 1.0,
+      "run_id": "benchmark_run",
+      "capture_steps": true
+    }
+  },
+  "8": {
+    "class_type": "KreaBenchmarkVAEDecode",
+    "inputs": {
+      "samples": ["7", 0],
+      "vae": ["3", 0],
+      "run_id": "benchmark_run"
+    }
+  },
+  "9": {
+    "class_type": "KreaBenchmarkSave",
+    "inputs": {
+      "images": ["8", 0],
+      "samples": ["7", 0],
+      "run_id": "benchmark_run",
+      "format_id": "gguf_q8_0",
+      "prompt_id": "benchmark",
+      "prompt_text": "benchmark prompt",
+      "seed": 0,
+      "scored": true
+    }
+  }
+}

reproduction/benchmark/workflows/krea2_benchmark_interactive_gguf.json

@@ -0,0 +1,152 @@
+{
+  "id": "41d49ac8-187d-4f8c-b192-b77c0b79c517",
+  "revision": 0,
+  "last_node_id": 9,
+  "last_link_id": 10,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "UnetLoaderGGUF",
+      "pos": [40, 40],
+      "size": [340, 82],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [{"name": "MODEL", "type": "MODEL", "links": [1]}],
+      "properties": {"Node name for S&R": "UnetLoaderGGUF"},
+      "widgets_values": ["krea2_turbo-Q8_0.gguf"]
+    },
+    {
+      "id": 2,
+      "type": "CLIPLoader",
+      "pos": [40, 170],
+      "size": [340, 110],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [{"name": "CLIP", "type": "CLIP", "links": [2]}],
+      "properties": {"Node name for S&R": "CLIPLoader"},
+      "widgets_values": ["qwen3vl_4b_bf16.safetensors", "krea2", "default"]
+    },
+    {
+      "id": 3,
+      "type": "VAELoader",
+      "pos": [40, 330],
+      "size": [340, 62],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [{"name": "VAE", "type": "VAE", "links": [8]}],
+      "properties": {"Node name for S&R": "VAELoader"},
+      "widgets_values": ["qwen_image_vae.safetensors"]
+    },
+    {
+      "id": 4,
+      "type": "CLIPTextEncode",
+      "pos": [440, 150],
+      "size": [440, 220],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [{"name": "clip", "type": "CLIP", "link": 2}],
+      "outputs": [{"name": "CONDITIONING", "type": "CONDITIONING", "links": [3, 4]}],
+      "properties": {"Node name for S&R": "CLIPTextEncode"},
+      "widgets_values": ["Ultra-realistic close-up editorial portrait of an elderly woman by a north-facing window, natural unretouched skin with visible pores and fine wrinkles, individual silver-gray hairs and eyelashes, wet catchlights in both eyes, a dark indigo wool coat with clearly visible fabric weave, soft neutral background bokeh."]
+    },
+    {
+      "id": 5,
+      "type": "ConditioningZeroOut",
+      "pos": [930, 280],
+      "size": [220, 46],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [{"name": "conditioning", "type": "CONDITIONING", "link": 4}],
+      "outputs": [{"name": "CONDITIONING", "type": "CONDITIONING", "links": [5]}],
+      "properties": {"Node name for S&R": "ConditioningZeroOut"},
+      "widgets_values": []
+    },
+    {
+      "id": 6,
+      "type": "EmptyLatentImage",
+      "pos": [440, 430],
+      "size": [310, 106],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [{"name": "LATENT", "type": "LATENT", "links": [6]}],
+      "properties": {"Node name for S&R": "EmptyLatentImage"},
+      "widgets_values": [1024, 1024, 1]
+    },
+    {
+      "id": 7,
+      "type": "KreaBenchmarkSampler",
+      "pos": [1210, 80],
+      "size": [390, 420],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {"name": "model", "type": "MODEL", "link": 1},
+        {"name": "positive", "type": "CONDITIONING", "link": 3},
+        {"name": "negative", "type": "CONDITIONING", "link": 5},
+        {"name": "latent_image", "type": "LATENT", "link": 6}
+      ],
+      "outputs": [{"name": "LATENT", "type": "LATENT", "links": [7, 10]}],
+      "properties": {"Node name for S&R": "KreaBenchmarkSampler"},
+      "widgets_values": [1732874201, 8, 1.0, "euler", "simple", 1.0, "interactive_run", true]
+    },
+    {
+      "id": 8,
+      "type": "KreaBenchmarkVAEDecode",
+      "pos": [1670, 110],
+      "size": [300, 100],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {"name": "samples", "type": "LATENT", "link": 7},
+        {"name": "vae", "type": "VAE", "link": 8}
+      ],
+      "outputs": [{"name": "IMAGE", "type": "IMAGE", "links": [9]}],
+      "properties": {"Node name for S&R": "KreaBenchmarkVAEDecode"},
+      "widgets_values": ["interactive_run"]
+    },
+    {
+      "id": 9,
+      "type": "KreaBenchmarkSave",
+      "pos": [2040, 100],
+      "size": [410, 310],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {"name": "images", "type": "IMAGE", "link": 9},
+        {"name": "samples", "type": "LATENT", "link": 10}
+      ],
+      "outputs": [],
+      "properties": {"Node name for S&R": "KreaBenchmarkSave"},
+      "widgets_values": ["interactive_run", "gguf_q8_0", "interactive", "Ultra-realistic close-up editorial portrait of an elderly woman by a north-facing window.", 1732874201, true]
+    }
+  ],
+  "links": [
+    [1, 1, 0, 7, 0, "MODEL"],
+    [2, 2, 0, 4, 0, "CLIP"],
+    [3, 4, 0, 7, 1, "CONDITIONING"],
+    [4, 4, 0, 5, 0, "CONDITIONING"],
+    [5, 5, 0, 7, 2, "CONDITIONING"],
+    [6, 6, 0, 7, 3, "LATENT"],
+    [7, 7, 0, 8, 0, "LATENT"],
+    [8, 3, 0, 8, 1, "VAE"],
+    [9, 8, 0, 9, 0, "IMAGE"],
+    [10, 7, 0, 9, 1, "LATENT"]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {"frontendVersion": "1.44.19"},
+  "version": 0.4
+}

scripts/prepare_release.py

@@ -383,8 +383,11 @@ This release is a paired, deterministic comparison of eight Krea 2 Turbo checkpo
 1. **BF16** is the highest-fidelity reference because it is the unquantized published checkpoint.
 2. **INT8 ConvRot** narrowly leads the preregistered LPIPS-Alex endpoint. **GGUF Q8_0** is statistically tied with it on that endpoint (Holm-adjusted permutation p=0.9067) and leads INT8 on DISTS, DINO similarity, and reconstructed-weight SNR.
 3. **GGUF Q8_0** is the highest-fidelity of the three added formats.
-4. **GGUF Q4_K_M** is a storage-quality compromise: 6.972 GiB and much closer to BF16 than INT4 ConvRot, but slow on this Ada GPU under ComfyUI-GGUF dequantized execution.
-5. **INT4 ConvRot W4A4** is the smallest checkpoint and fastest of the three additions on this RTX 4060 Ti, but has the largest measured fidelity loss.
+4. **MXFP8** ranks fourth in fidelity (LPIPS 0.071229) and halves checkpoint size versus BF16, but its native SM 10.0 fast path was unavailable on this Ada GPU, resulting in 31.929 s fallback sampling.
+5. **FP8 Scaled** ranks fifth (LPIPS 0.093701) at 12.239 GiB and was the fastest of the original five formats at 19.503 s on this GPU.
+6. **GGUF Q4_K_M** is a storage-quality compromise: 6.972 GiB and much closer to BF16 than INT4 ConvRot, but slow on this Ada GPU under ComfyUI-GGUF dequantized execution.
+7. **NVFP4** is 7.147 GiB with moderate fidelity loss (LPIPS 0.205124). Like MXFP8, its native fast multiplication requires SM 10.0; the recorded 24.925 s sampling time uses the fallback path.
+8. **INT4 ConvRot W4A4** is the smallest checkpoint and fastest of the three additions on this RTX 4060 Ti, but has the largest measured fidelity loss.
 
 No weighted composite score is used. See [the full technical report](TECHNICAL_REPORT.md) and [`tables/decision_table.csv`](tables/decision_table.csv).