| | --- |
| | license: apache-2.0 |
| | license_name: wan-ai-license |
| | license_link: https://github.com/Wan-Video/Wan2.2/blob/main/LICENSE.txt |
| | base_model: Video-Reason/VBVR-Wan2.2 |
| | library_name: diffusers |
| | tags: |
| | - wan2.2 |
| | - i2v |
| | - fp8 |
| | - comfyui |
| | - video-generation |
| | - surgical-quant |
| | --- |
| | # Wan2.2-I2V-14B: HiFi-Surgical-FP8 & BF16 (ComfyUI Optimized) |
| |
|
| | This model follows the Wan-AI Software License Agreement. Please refer to the original repository for usage restrictions. |
| |
|
| | This repository provides two high-performance versions of **Wan2.2-I2V-14B**, meticulously optimized for the **ComfyUI** ecosystem. We offer a standard **BF16** version and a specialized **HiFi-Surgical-FP8** mixed-precision version. |
| |
|
| | * **Original Project**: [Video-Reason Wan2.2](https://video-reason.com/) |
| | * **Original Weights**: [HuggingFace - VBVR-Wan2.2](https://huggingface.co/Video-Reason/VBVR-Wan2.2) |
| |
|
| | --- |
| |
|
| | ## 💎 The HiFi-Surgical Optimization Strategy |
| |
|
| | Unlike generic "one-click" quantization scripts that often cause visual degradation in Wan2.2, our **HiFi-Surgical-FP8** version uses a data-driven, diagnostic-led approach to preserve cinematic quality. |
| |
|
| | ### 1. Layer-Wise SNR Calibration |
| | We performed a deep medical-grade scan on all 406 linear weight tensors of the FP32 Master. Only layers maintaining an **SNR (Signal-to-Noise Ratio) > 31.5dB** were converted to FP8. This ensures that the mathematical "soul" of the model remains intact. |
| |
|
| | ### 2. High-Outlier Protection |
| | Wan2.2 weights are notoriously "fragile" with sharp numerical peaks. Our strategy identifies layers with a high **Outlier Index** (Max/Std deviation > 12) and locks them in **BF16**. This specifically targets and eliminates the "sparkle" noise and flickering artifacts common in standard FP8 conversions. |
| |
|
| | ### 3. Structural Integrity (Blocks 30-39) |
| | We have physically isolated the **Cross-Attention** layers in the final blocks of the DiT architecture. By keeping these critical layers in BF16, we ensure that prompt adherence and temporal consistency are not compromised. |
| |
|
| | --- |
| |
|
| | ## 📊 Comparison & Specs |
| |
|
| | | Feature | Standard BF16 | **HiFi-Surgical-FP8 (Recommended)** | |
| | | :--- | :--- | :--- | |
| | | **File Size** | ~27.2 GB | **~22.4 GB** | |
| | | **Precision** | Pure Bfloat16 | **Hybrid FP8-E4M3 / BF16** | |
| | | **VRAM Requirement** | 24GB+ | **16GB - 24GB** | |
| | | **Visual Fidelity** | Reference Grade | **99% Reference Match** | |
| | | **Inference Speed** | Base Speed | **Accelerated on Blackwell/Hopper** | |
| |
|
| | --- |
| |
|
| | ## 🛠️ ComfyUI Integration & Usage |
| |
|
| | These models are specifically converted and tested for **ComfyUI**. |
| |
|
| | 1. **Native Scaling Support**: We have included the `scale_weight` metadata for every quantized tensor. This allows ComfyUI loaders to utilize hardware-level scaling on **NVIDIA Blackwell (RTX 50-series)** and **Hopper** architectures for maximum speed. |
| | 2. **How to Use**: |
| | * Place the `.safetensors` file in your `ComfyUI/models/diffusion_models/. |
| | * Use the **CheckpointLoaderSimple** or the specialized **UNETLoader**. |
| | * Ensure your ComfyUI is up-to-date to support the `float8_e4m3fn` type. |
| |
|
| | --- |
| |
|
| | ## 📝 Diagnostic Methodology |
| |
|
| | Each weight in the HiFi version was selected based on the following diagnostic results: |
| | * **Total Layers Scanned**: 406 |
| | * **FP8 Layers**: 184 (Non-sensitive FFN & Attention layers) |
| | * **BF16 Layers**: 222 (Sensitive Cross-Attention & Outlier-heavy layers) |
| | * **Target Hardware**: Optimized for RTX 4090, 5090, and H100/H200. |
