Hybrid-Sensitivity-Weighted-Quantization (HSWQ)
High-fidelity ConvRot NVFP4 quantization for diffusion models (Z-Image / SDXL). HSWQ uses sensitivity and importance analysis instead of naive uniform cast. This is highly useful for users who need to strictly manage their VRAM resources while maintaining maximum image quality.
ComfyUI Load Diffusion Model nvfp4 pack with FULL ConvRot (Linear→NVFP4, Conv2d→INT8 int8_tensorwise) after DualMonitor + V4 pack-MSE FP16 protection under a fixed budget. calib writes NVFP4 .input_scale. Z-Image pack scripts: hswq_convert_nvfp4_1.0.py (HSWQ) and native_convert_nvfp4.py (native).
Technical details: https://github.com/ussoewwin/Hybrid-Sensitivity-Weighted-Quantization
How to quantize (SDXL/Z-Image ConvRot NVFP4): md/How to quantize SDXL NVFP4.md
ComfyUI Loader for ConvRot NVFP4: To use these models in ComfyUI, please use this custom node: ComfyUI-HSWQ-Loader-and-Tools
Benchmark (Reference)
| Model | SSIM (Avg) | File size | Compatibility |
|---|---|---|---|
| Original FP16 | 1.0000 | 100% | High |
| HSWQ Z-Image ConvRot NVFP4 | 0.97 | 60% (FP16 mixed) | High (ComfyUI NVFP4) |
📦 Available Models
| Filename | Base Model | Version | License |
|---|---|---|---|
moodyProMix_zitV13_hswq_int8protect60_convrot_nvfp4.safetensors |
Moody Pro Mix | v1.3 | CreativeML Open RAIL++-M |
moodyRealMix_zitV7_hswq_int8protect60_convrot_nvfp4.safetensors |
Moody Real Mix | v7.0 | CreativeML Open RAIL++-M |
📜 Credits & License
🏆 Special Acknowledgement
We extend our deepest respect and gratitude to the Nunchaku Team for their groundbreaking work on SVDQ quantization and for sharing their models with the community. This collection relies heavily on their research and original implementation.
- Original Repository: nunchaku-tech/nunchaku-sdxl
Base Models
These models are derivatives of their respective creators. All credit for aesthetic tuning and model training belongs to the original creators.
- Moody Pro Mix / Moody Real Mix: Created by catlover1937.
Disclaimer: These models are provided for optimization and research purposes. Please adhere to the original licenses of the base models.