docs: update ControlNet ConvRot INT8 model card

README.md

@@ -1,3 +1,137 @@
 ---
 license: apache-2.0
+tags:
+- controlnet
+- text-to-image
+- image-to-image
+- inpainting
+- qwen-image
+- qwen-image-2512
+- videox-fun
+- quantized
+- int8
+- convrot
+- comfyui
+pipeline_tag: image-to-image
+library_name: videox_fun
 ---
+
+# ControlNet Models (ConvRot INT8)
+
+<p align="center">
+  <img src="https://raw.githubusercontent.com/ussoewwin/Hybrid-Sensitivity-Weighted-Quantization/main/icon.png" width="128">
+</p>
+
+High-fidelity **ConvRot INT8** quantized weights for diffusion ControlNet models. This repository provides memory-efficient, production-grade INT8 quantized checkpoints maintaining high structural fidelity and multi-condition guidance.
+
+---
+
+## 🌟 Model Overview
+
+This repository hosts the **ConvRot INT8** quantized edition of **`Qwen-Image-2512-Fun-Controlnet-Union-2602`**, originally developed and trained by **Alibaba PAI / VideoX-Fun**.
+
+The **ConvRot INT8** format applies rotational transformation matrix optimizations to eliminate outlier activation skew and ensure minimal reconstruction error, reducing the model footprint to **~1.64 GB** while preserving multi-condition structural precision.
+
+- **Base Model:** [alibaba-pai/Qwen-Image-2512-Fun-Controlnet-Union](https://huggingface.co/alibaba-pai/Qwen-Image-2512-Fun-Controlnet-Union)
+- **Architecture:** Multi-condition Union ControlNet added across 5 layer blocks of Qwen-Image-2512.
+- **Quantization:** Full ConvRot INT8 (weights stored as `int8_tensorwise` with rotation matrices).
+- **License:** Apache-2.0
+
+---
+
+## 📦 Available Models
+
+| Filename | Base Architecture | Supported Conditions | Quantization | Size | License |
+| :--- | :--- | :--- | :--- | :--- | :--- |
+| `Qwen-Image-2512-Fun-Controlnet-Union-2602_convrot_int8.safetensors` | Qwen-Image-2512 (5 Layer Blocks) | Canny, HED, Depth, Pose, MLSD, Scribble, Gray, Inpaint | ConvRot INT8 | ~1.64 GB | Apache-2.0 |
+
+---
+
+## 🛠️ Supported Control Modalities & Features
+
+1. **Union Condition Architecture**:
+   Supports unified multi-condition control within a single checkpoint:
+   - **Pose** (OpenPose / DWPose keypoints)
+   - **Canny** (Edge detection)
+   - **HED** (Soft edge detection)
+   - **Depth** (Z-depth estimation)
+   - **MLSD** (Straight-line wireframe extraction)
+   - **Scribble** (Interactive doodle / sketch guidance)
+   - **Gray** (Grayscale / luminance guidance)
+2. **Inpainting Mode Support**:
+   Fully compatible with masked inpainting pipelines (`Pose + Inpaint`, `Depth + Inpaint`, etc.).
+3. **Multi-Resolution Conditioning**:
+   Extracting control maps in multi-resolution formats improves generalization across arbitrary target aspect ratios.
+
+---
+
+## ⚙️ Recommended Inference Settings
+
+- **`control_context_scale`**: Optimal range is **`0.70` – `0.95`**. Higher values yield stronger adherence to the control input, while lower values offer increased prompt stylization freedom.
+- **Prompt Formulation**: Detailed natural language descriptions significantly enhance semantic stability and detail preservation.
+- **Resolution**: Multi-scale aspect ratios matching the base Qwen-Image model resolution specifications.
+
+---
+
+## 🚀 How to Use
+
+### 1. VideoX-Fun Framework
+
+Clone the upstream [VideoX-Fun](https://github.com/aigc-apps/VideoX-Fun) repository:
+
+```bash
+git clone https://github.com/aigc-apps/VideoX-Fun.git
+cd VideoX-Fun
+
+mkdir -p models/Diffusion_Transformer
+mkdir -p models/Personalized_Model
+```
+
+Place the quantized model under `models/Personalized_Model/`:
+
+```
+models/
+├── Diffusion_Transformer/
+│   └── Qwen-Image-2512/
+└── Personalized_Model/
+    └── Qwen-Image-2512-Fun-Controlnet-Union-2602_convrot_int8.safetensors
+```
+
+Execute standard prediction pipelines:
+```bash
+# Text-to-Image with ControlNet
+python examples/qwenimage_fun/predict_t2i_control.py
+
+# Image-to-Image with Inpainting + ControlNet
+python examples/qwenimage_fun/predict_i2i_inpaint.py
+```
+
+### 2. ComfyUI
+
+For ComfyUI workflows, use the dedicated loader node pack:
+- **[ComfyUI-HSWQ-Loader-and-Tools](https://github.com/ussoewwin/ComfyUI-HSWQ-Loader-and-Tools)**
+- **[ComfyUI-QwenImageLoraLoader](https://github.com/ussoewwin/ComfyUI-QwenImageLoraLoader)**
+
+---
+
+## 📜 Credits & Citation
+
+### Upstream Creators
+Deep gratitude and acknowledgement to the **Alibaba PAI VideoX-Fun Team** for developing, training, and open-sourcing the Qwen-Image ControlNet series.
+- **Original Model:** [alibaba-pai/Qwen-Image-2512-Fun-Controlnet-Union](https://huggingface.co/alibaba-pai/Qwen-Image-2512-Fun-Controlnet-Union)
+- **Official Repository:** [aigc-apps/VideoX-Fun](https://github.com/aigc-apps/VideoX-Fun)
+
+```bibtex
+@misc{videox_fun,
+  author = {VideoX-Fun Team},
+  title = {VideoX-Fun: A Flexible Video and Image Generation Toolset},
+  year = {2025},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/aigc-apps/VideoX-Fun}}
+}
+```
+
+---
+
+**Disclaimer:** These model weights are released strictly for research, non-commercial, and optimization workflows under the Apache-2.0 License.