Document Ming Image Design-Layer checkpoints

README.md

@@ -28,4 +28,24 @@ inference variant. The original model sources and conversion notes are in
 WanGP's `models/ming_image/` directory.
 
 WanGP supports text-to-image and one-reference image editing with these files.
-The separate Design-Layer model is not included.
+
+## Ming Image 0.1 Design-Layer
+
+Design-Layer separates a flattened reference image into ordered RGBA raster layers. WanGP places every layer in the gallery, frontmost first, and saves a ZIP. Supply one reference image and a front-to-back layer plan. Start with 12 steps, guidance 2, and the 1024 working bucket; 512 is faster.
+
+Its BF16 and INT8 ConvRot transformer checkpoints are at the repository root. The Layer-specific Bailing encoder checkpoints and tokenizer are under `BailingMM2-Ming-Image-Layer/`, and runtime configs are under `ming_image_layer/`. The Design and Design-Layer checkpoints use the same VAE from `ming_image/`. Their Bailing base weights are identical, but the Layer connector, MLP, and tokenizer differ, so WanGP uses a separate complete Layer encoder.
+
+Source: [inclusionAI/Ming-Image-0.1-Design-Layer](https://huggingface.co/inclusionAI/Ming-Image-0.1-Design-Layer), revision `9fabca8b62a67f1f53a957d46389c00451c11e52` (MIT). The released Layer transformer was converted from FP32 to BF16 to match upstream BF16 inference. The merged encoder and ConvRot files were verified against their source tensors and exercised in WanGP generation.
+
+Example plan:
+
+```text
+Decompose this image into 4 layers with the following specifications:
+Number of layers: 4
+Layer 1: All clearly readable foreground text, preserving its exact wording and placement.
+Layer 2: The card or panel directly behind the text.
+Layer 3: The main foreground subject or illustration.
+Layer 4: The complete background and remaining supporting shapes.
+```
+
+The layers are raster images, including any text. The last gallery image is the background.