Darwin-Image-v1 — Unified DiT + VLM

Darwin-Image-v1 is a physically unified model repository that combines two models into a single HF repo:

1. Z-Image Turbo (6B DiT) at the repo root, with 4 LoRAs fused into the transformer weights (no runtime adapter loading).
2. Darwin-4B-David (Gemma4 multimodal VLM, ~16GB bf16) inside the vlm_judge/ subfolder.

Both models coexist as real safetensors files in this single repo, so a single from_pretrained() call downloads everything needed for the AETHER metacognitive image generation pipeline.

Repo Layout

FINAL-Bench/Darwin-Image-v1/
├── model_index.json            # Z-Image pipeline manifest
├── scheduler/                  # Z-Image components
├── text_encoder/               #   (Qwen3)
├── tokenizer/
├── transformer/                # ★ DiT with 4 LoRAs fused
├── vae/
├── vlm_judge/                  # ★★ Darwin-4B-David
│   ├── config.json             #    Gemma4ForConditionalGeneration
│   ├── model.safetensors       #    ~16GB bfloat16
│   ├── tokenizer.json
│   ├── chat_template.jinja
│   └── generation_config.json
├── lora_manifest.yaml
├── aether_config.json
├── fuse_report.json
└── README.md  (this file)

LoRA Stack Fused Into DiT (v2, 2026-04-10)

Z-Image Turbo (6B DiT, bf16)
    └── + Shakker-Labs/AWPortrait-Z (scale 0.7) → portrait quality
    └── + qqnyanddld/nsfw-z-image-lora (scale 0.5) → uncensored
    └── + renderartist/Technically-Color-Z-Image-Turbo (scale 0.3) → color

All LoRAs use the ai-toolkit standard format (480 keys each, diffusion_model.layers.N.X.lora_A/B.weight). Fused via direct matrix update (pipeline/manual_fuse.py): W += (B @ A) × (alpha/rank) × scale.

v2 Changelog

• Removed wcde/Z-Image-Turbo-DeJPEG-Lora/dejpeg_v3 — caused over-smoothing that destroyed portrait/color detail. delta_norm was 3.67 vs 0.14~0.41 for other LoRAs (26× baseline). Z-Image Turbo is distilled and has minimal JPEG artifacts to begin with, so dejpeg was unnecessary.
• Reduced Technically-Color-Z-Image-Turbo scale 0.4 → 0.3 to prevent over-saturation on high-contrast/neon scenes.

Usage

DiT only (text-to-image)

from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained(
    "FINAL-Bench/Darwin-Image-v1",
    torch_dtype=torch.bfloat16,
    token="hf_...",
).to("cuda")

image = pipe(
    prompt="cinematic portrait of a korean woman, golden hour, 85mm f1.4",
    num_inference_steps=8,
    guidance_scale=3.5,
    height=1024,
    width=1024,
).images[0]

Load the bundled VLM judge (from subfolder)

from transformers import AutoModel, AutoProcessor
import torch

judge = AutoModel.from_pretrained(
    "FINAL-Bench/Darwin-Image-v1",
    subfolder="vlm_judge",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    token="hf_...",
)
processor = AutoProcessor.from_pretrained(
    "FINAL-Bench/Darwin-Image-v1",
    subfolder="vlm_judge",
    token="hf_...",
)

AETHER Integration

For the full metacognitive pipeline (VLM prompt enhancement + quality judging + Korean text inpainting), use the Darwin Image pipeline code:

from darwin_image.pipeline import DarwinZImagePipeline, DarwinJudge, run_aether, AetherConfig

pipe = DarwinZImagePipeline(base_model="FINAL-Bench/Darwin-Image-v1")
judge = DarwinJudge()  # loads Darwin-4B-David

result = run_aether(
    user_prompt='영화 포스터: "봄의 서울" 벚꽃 남산타워',
    pipe=pipe,
    judge=judge,
    config=AetherConfig(max_iter=3, threshold=8.0),
    seed=42,
)
result.final_image.save("out.png")

Default AETHER Config

See aether_config.json for the default metacognitive loop parameters:

• max_iter: 3 (up to 3 retries)
• threshold: 8.0 (overall score threshold to exit early)
• enable_vlm_enhance: true (VLM rewrites prompts)
• enable_vlm_judge: true (VLM scores each iteration)
• enable_mti: false (experimental, off by default on 8-step distilled models)

License

Apache 2.0 — inherits from Z-Image Turbo and all LoRA base models.

Citation

If you use Darwin-Image-v1, please cite:

@misc{darwin-image-2026,
  title={Darwin Image: VLM-Guided Metacognitive Image Generation with Korean Text Integration},
  author={VIDRAFT and 지니젠AI and FINAL-Bench},
  year={2026},
  howpublished={\url{https://huggingface.co/FINAL-Bench/Darwin-Image-v1}},
}

Safety Notice

This model includes an uncensored LoRA component. Use responsibly. The AETHER pipeline Space (darwin-image-gen) is deployed as private (FINAL-Bench org only) for this reason.