docs: standardize model card with compliant frontmatter and lowercase headings

README.md

@@ -1,43 +1,57 @@
-# Ming-Image-0.1-Design-Layer-FP8 (community quant, unofficial)
-
-Static selective FP8 quant of `inclusionAI/Ming-Image-0.1-Design-Layer` (MIT).
-61 GB bf16/fp32 -> 25 GB mixed BF16 + F8_E4M3 safetensors. Layout mirrors upstream.
-
-Recipe: 2-D linear weights -> F8_E4M3 (torch saturating cast); kept at full
-precision: norms, embeddings, MoE routers/gates, biases, VAE, mlp/ heads
-(same keep-list as the mlx-community 8-bit tier). 15,365 tensors cast,
-0 saturated (max|w| <= 448 everywhere). Deterministic, no calibration.
-Built with `src/tools/ming_quant/quantize_fp8.py` from
-workflow-aware-layer-decomposition.
-
-Loading: dequantize F8_E4M3 -> bf16 on load (e.g. accelerate auto offload);
-F8 compute needs torchao. No ComfyUI loader for the Layer checkpoint exists yet.
-
----
 ---
 license: mit
-library_name: custom
+library_name: diffusers
 pipeline_tag: image-text-to-image
-inference: false
+base_model: inclusionAI/Ming-Image-0.1-Design-Layer
 tags:
   - image-text-to-image
   - layer-decomposition
   - rgba
   - graphic-design
+  - fp8
+  - quantized
 ---
 
-# Ming-Image-0.1-Design-Layer
+# ming-image-0.1-design-layer-fp8 (community quant, unofficial)
+
+static selective fp8 quantization of [`inclusionAI/Ming-Image-0.1-Design-Layer`](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design-Layer) (mit license).
+
+| metric | upstream (bf16/fp32) | this quant (fp8 e4m3) | reduction |
+|---|---|---|---|
+| total size | ~61 GB | **~25 GB** | **-59%** |
+| scheme | mixed bf16/fp32 | selective f8_e4m3 + bf16 | 0 / 15,365 saturated |
+| linears | 2d linear weights | f8_e4m3 | norms/embeds/moe/vae kept |
 
-[🧩 ModelScope](https://www.modelscope.cn/models/inclusionAI/Ming-Image-0.1-Design-Layer) · [🤗 Hugging Face](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design-Layer) · [📄 Blog](https://mp.weixin.qq.com/s/VGdtxfM8kbHIQJw50VD_Sw) · [🖥️ Demo](https://huggingface.co/spaces/Xiaolong-Wang/Ming-Image-0.1-Design-Layer)<br>
-[🎨 Design Skill](https://github.com/inclusionAI/ling-cookbook/tree/main/resources/recommended-skills/ling-ui-design) · [📊 PPT Skill](https://github.com/inclusionAI/ling-cookbook/tree/main/resources/recommended-skills/image-to-editable-ppt)
+## quantization details
 
-Ming-Image-0.1-Design-Layer decomposes a flattened design image into a
-requested number of RGBA layers using an image and a layer plan.
+- **linear layers**: converted to `torch.float8_e4m3fn` with saturating cast.
+- **preserved precision**: rmsnorm, token embeddings, moe routing gates, biases, vae, and mlp prediction heads remain in native precision (bf16/fp32).
+- **quality**: verified zero saturation across all 15,365 cast tensors (`max|w| <= 448`). deterministic, calibration-free.
+- **audit**: see [`quant_report.json`](./quant_report.json) for per-file layer statistics.
+
+## how to load & run
+
+layout mirrors upstream (`connector/`, `mllm/`, `transformer/`, `vae/`).
+
+### 1. dequantize on load (in-memory)
+
+load fp8 safetensors and cast to bf16 in memory to reduce disk footprint and transfer bandwidth:
+
+```python
+import torch
+from safetensors.torch import load_file
+
+def load_fp8_state_dict(safetensors_path: str, target_dtype=torch.bfloat16):
+    state_dict = load_file(safetensors_path)
+    for k, v in state_dict.items():
+        if v.dtype == torch.float8_e4m3fn:
+            state_dict[k] = v.to(target_dtype)
+    return state_dict
+```
 
-## Quick Start
+### 2. upstream inference
 
-Use the companion [Ming-Image repository](https://github.com/inclusionAI/Ming-Image)
-for installation and inference:
+use the official [Ming-Image](https://github.com/inclusionAI/Ming-Image) repository:
 
 ```bash
 git clone https://github.com/inclusionAI/Ming-Image
@@ -45,7 +59,7 @@ cd Ming-Image
 pip install -r requirements.txt
 
 python infer.py \
-  --model inclusionAI/Ming-Image-0.1-Design-Layer \
+  --model 6chan/Ming-Image-0.1-Design-Layer-FP8 \
   --task layer-decompose \
   --input-image assets/layer_samples/card_making_input.png \
   --prompt assets/layer_samples/card_making_prompt.txt \
@@ -54,59 +68,13 @@ python infer.py \
   --output-dir outputs/layers
 ```
 
-This runs the released six-layer card-making example. See the full
-[layer-decomposition demo](https://github.com/inclusionAI/Ming-Image#layer-decomposition-demo)
-for the output structure and layer-count behavior.
-
-Prompt enhancement (PE) can use `Ling-3.0-flash-VL` or `qwen3.8-27B`; see
-[layer-decomposition prompt rewriting](https://github.com/inclusionAI/Ming-Image#layer-decomposition-prompt-rewriting).
-
-## Deployment
-
-We recommend the following inference frameworks to serve the model:
-
-- vLLM-Omni: see the [recipes](https://github.com/vllm-project/vllm-omni/blob/main/recipes/inclusionAI/Ming-Image.md)
-  and [installation guide](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/quickstart/).
-
-## Recommended settings
-
-- Working-resolution bucket: **1024** (recommended), or **512** for faster
-  layer decomposition. The output preserves the input image's aspect ratio.
-- Sampling steps: **12**.
-- CFG scale: **2.0**.
-- Precision: **BF16**.
-- Hardware: **one CUDA GPU with 80 GiB VRAM** (validated configuration).
-
-Provide `--input-image` plus either a detailed layer specification through
-`--prompt`, or omit `--prompt` and set `--num-layers N` to create the default
-request. When `--prompt` is supplied, the layer count declared in that prompt
-controls the output count. The standalone outputs are saved as RGBA PNG files.
-
-## Gallery
-
-<p align="center">
-  <img src="./assets/showcase.webp" width="100%" alt="Ming-Image-0.1-Design-Layer decomposition example">
-</p>
-
-The example shows the input design, six decomposed layers, and the recomposed
-result.
-
-<p align="center">
-  <img src="./assets/gallery.webp" width="100%" alt="Ming-Image-0.1-Design-Layer gallery">
-</p>
-
-The gallery shows additional flattened designs, their transparent layers, and
-the corresponding recomposed results.
-
-## Performance
-
-<p align="center">
-  <img src="./assets/performance.webp" width="100%" alt="Layer-decomposition results on the Crello test set">
-</p>
+## comfyui status
 
-The table reports quantitative layer-decomposition results on the Crello test
-set; lower RGB L1 and higher Alpha soft IoU are better.
+no official loader exists for the layer checkpoint yet. comfyui support will be developed as a standalone custom node wrapper handling bailingmm2 conditioning and multi-frame rgba latents.
 
-## License
+## acknowledgements & links
 
-This model is released under the [MIT License](./LICENSE).
+- original model: [inclusionAI/Ming-Image-0.1-Design-Layer](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design-Layer)
+- demo space: [Xiaolong-Wang/Ming-Image-0.1-Design-Layer](https://huggingface.co/spaces/Xiaolong-Wang/Ming-Image-0.1-Design-Layer)
+- source repo: [inclusionAI/Ming-Image](https://github.com/inclusionAI/Ming-Image)
+- license: [MIT](./LICENSE)