FINAL-Bench/Darwin-Image-v1

1---2license: apache-2.03language:4  - en5  - ko6library_name: diffusers7pipeline_tag: text-to-image8tags:9  - darwin-image10  - aether-metacognitive11  - z-image12  - lora-merge13  - korean-text14  - vlm-judge15base_model:16  - Tongyi-MAI/Z-Image-Turbo17  - Shakker-Labs/AWPortrait-Z18  - qqnyanddld/nsfw-z-image-lora19  - renderartist/Technically-Color-Z-Image-Turbo20  - wcde/Z-Image-Turbo-DeJPEG-Lora21inference: false22---23 24# Darwin-Image-v1 — Unified DiT + VLM25 26**Darwin-Image-v1** is a **physically unified** model repository that27combines two models into a single HF repo:28 291. **Z-Image Turbo** (6B DiT) at the repo root, with 4 LoRAs fused into30   the transformer weights (no runtime adapter loading).312. **Darwin-4B-David** (Gemma4 multimodal VLM, ~16GB bf16) inside the32   `vlm_judge/` subfolder.33 34Both models coexist as real safetensors files in this single repo, so a35single `from_pretrained()` call downloads everything needed for the36AETHER metacognitive image generation pipeline.37 38## Repo Layout39 40```41FINAL-Bench/Darwin-Image-v1/42├── model_index.json            # Z-Image pipeline manifest43├── scheduler/                  # Z-Image components44├── text_encoder/               #   (Qwen3)45├── tokenizer/46├── transformer/                # ★ DiT with 4 LoRAs fused47├── vae/48├── vlm_judge/                  # ★★ Darwin-4B-David49│   ├── config.json             #    Gemma4ForConditionalGeneration50│   ├── model.safetensors       #    ~16GB bfloat1651│   ├── tokenizer.json52│   ├── chat_template.jinja53│   └── generation_config.json54├── lora_manifest.yaml55├── aether_config.json56├── fuse_report.json57└── README.md  (this file)58```59 60## LoRA Stack Fused Into DiT (v2, 2026-04-10)61 62```63Z-Image Turbo (6B DiT, bf16)64    └── + Shakker-Labs/AWPortrait-Z (scale 0.7) → portrait quality65    └── + qqnyanddld/nsfw-z-image-lora (scale 0.5) → uncensored66    └── + renderartist/Technically-Color-Z-Image-Turbo (scale 0.3) → color67```68 69All LoRAs use the ai-toolkit standard format (480 keys each,70`diffusion_model.layers.N.X.lora_A/B.weight`). Fused via direct matrix71update (`pipeline/manual_fuse.py`): `W += (B @ A) × (alpha/rank) × scale`.72 73### v2 Changelog74 75- **Removed** `wcde/Z-Image-Turbo-DeJPEG-Lora/dejpeg_v3` — caused over-smoothing76  that destroyed portrait/color detail. delta_norm was 3.67 vs 0.14~0.41 for77  other LoRAs (26× baseline). Z-Image Turbo is distilled and has minimal78  JPEG artifacts to begin with, so dejpeg was unnecessary.79- **Reduced** `Technically-Color-Z-Image-Turbo` scale 0.4 → 0.3 to prevent80  over-saturation on high-contrast/neon scenes.81 82## Usage83 84### DiT only (text-to-image)85```python86from diffusers import DiffusionPipeline87import torch88 89pipe = DiffusionPipeline.from_pretrained(90    "FINAL-Bench/Darwin-Image-v1",91    torch_dtype=torch.bfloat16,92    token="hf_...",93).to("cuda")94 95image = pipe(96    prompt="cinematic portrait of a korean woman, golden hour, 85mm f1.4",97    num_inference_steps=8,98    guidance_scale=3.5,99    height=1024,100    width=1024,101).images[0]102```103 104### Load the bundled VLM judge (from subfolder)105```python106from transformers import AutoModel, AutoProcessor107import torch108 109judge = AutoModel.from_pretrained(110    "FINAL-Bench/Darwin-Image-v1",111    subfolder="vlm_judge",112    torch_dtype=torch.bfloat16,113    device_map="cuda",114    token="hf_...",115)116processor = AutoProcessor.from_pretrained(117    "FINAL-Bench/Darwin-Image-v1",118    subfolder="vlm_judge",119    token="hf_...",120)121```122 123## AETHER Integration124 125For the full metacognitive pipeline (VLM prompt enhancement + quality126judging + Korean text inpainting), use the Darwin Image pipeline code:127 128```python129from darwin_image.pipeline import DarwinZImagePipeline, DarwinJudge, run_aether, AetherConfig130 131pipe = DarwinZImagePipeline(base_model="FINAL-Bench/Darwin-Image-v1")132judge = DarwinJudge()  # loads Darwin-4B-David133 134result = run_aether(135    user_prompt='영화 포스터: "봄의 서울" 벚꽃 남산타워',136    pipe=pipe,137    judge=judge,138    config=AetherConfig(max_iter=3, threshold=8.0),139    seed=42,140)141result.final_image.save("out.png")142```143 144## Default AETHER Config145 146See `aether_config.json` for the default metacognitive loop parameters:147- `max_iter`: 3 (up to 3 retries)148- `threshold`: 8.0 (overall score threshold to exit early)149- `enable_vlm_enhance`: true (VLM rewrites prompts)150- `enable_vlm_judge`: true (VLM scores each iteration)151- `enable_mti`: false (experimental, off by default on 8-step distilled models)152 153## License154 155Apache 2.0 — inherits from Z-Image Turbo and all LoRA base models.156 157## Citation158 159If you use Darwin-Image-v1, please cite:160 161```bibtex162@misc{darwin-image-2026,163  title={Darwin Image: VLM-Guided Metacognitive Image Generation with Korean Text Integration},164  author={VIDRAFT and 지니젠AI and FINAL-Bench},165  year={2026},166  howpublished={\url{https://huggingface.co/FINAL-Bench/Darwin-Image-v1}},167}168```169 170## Safety Notice171 172This model includes an uncensored LoRA component. Use responsibly.173The AETHER pipeline Space (`darwin-image-gen`) is deployed as **private**174(FINAL-Bench org only) for this reason.175