CoolFace
Modelpublic

ussoewwin/Hybrid-Sensitivity-Weighted-Quantization-Z-Image-Hybrid-ConvRot-NVFP4

sourceHugging Faceotherupdated 25d agoView on Hugging Face
0likes
Model Card

Hybrid-Sensitivity-Weighted-Quantization (HSWQ)

<p align="center"> <img src="https://raw.githubusercontent.com/ussoewwin/Hybrid-Sensitivity-Weighted-Quantization/main/icon.png" width="128"> </p>

High-fidelity Hybrid ConvRot NVFP4 quantization for Z-Image Turbo diffusion models. Built from a complete native ConvRot INT8 UNet via the reverse method (converting lowest-impact layers to NVFP4 in ascending order of trajectory impact). This is highly useful for users who need to strictly manage their VRAM resources (~53–58% savings) while maintaining high image fidelity and completely eliminating trajectory bifurcations.

Technical details: https://github.com/ussoewwin/Hybrid-Sensitivity-Weighted-Quantization

How to quantize (Z-Image Hybrid NVFP4): md/How to quantize Z Image - Hybrid NVFP4.md

ComfyUI Loader for ConvRot NVFP4: To use these models in ComfyUI, please use this custom node: ComfyUI-HSWQ-Loader-and-Tools

Z Image ConvRot NVFP4 Benchmark Test Results (published tables): benchmark result/benchmark_zi_nvfp4.md

Benchmark (Deterministic Latent Trajectory Divergence)

Multi-seed trajectory evaluation (20 seeds per model, 200 total evaluations across 10 Z-Image models):

MetricHSWQ ConvRot Hybrid NVFP4Native NVFP4 (Full Model)Advantage
Final Cosine (Avg)0.960 – 0.9740.900 – 0.947+0.020 to +0.060 higher fidelity
Bifurcation Rate0/200 (0.0%)3/200 (1.5%)Zero bifurcations across all seeds
Latent MSE Reduction0.74 – 1.251.45 – 2.5436% – 61% lower quantization error
VRAM Footprint~53% – 58% Savings~50% SavingsSignificantly reduced VRAM usage

📦 Available Models

FilenameBase ModelVersionLicense
moodyProMix_zitV13_hswq_hybrid_nv80_convrot_nvfp4.safetensorsMoody Pro Mixzit v1.3 (nv80)CreativeML Open RAIL++-M
moodyProMix_collectorsEdition_hswq_hybrid_nv90_convrot_nvfp4.safetensorsMoody Pro MixCollector's Edition (nv90/nv100)CreativeML Open RAIL++-M
moodyRealMix_zitV7_hswq_hybrid_nv100_convrot_nvfp4.safetensorsMoody Real Mixzit v7.0 (nv100)CreativeML Open RAIL++-M
moodyRealMix_xhsEdition_hswq_hybrid_nv110_convrot_nvfp4.safetensorsMoody Real MixXHS Edition (nv100/nv110)CreativeML Open RAIL++-M
darkBeast30BF16INT8_dbzit9DIMRclaw_hswq_hybrid_nv100_convrot_nvfp4.safetensorsDark Beastdbzit9 DIMRclaw (nv100)CreativeML Open RAIL++-M
unstableRevolution_V3Fp16_hswq_hybrid_nv90_convrot_nvfp4.safetensorsUnstable Revolutionv3.0 FP16 (nv90/nv100)CreativeML Open RAIL++-M
gonzalomoZpop_insta2_hswq_hybrid_nv80_convrot_nvfp4.safetensorsgonzalomo Z-Popinsta2 (nv80/nv100)CreativeML Open RAIL++-M
divingZImageTurbo_v70Fp16_hswq_hybrid_nv90_convrot_nvfp4.safetensorsDiving Z-Image Turbov7.0 FP16 (nv90/nv99)CreativeML Open RAIL++-M
beyondREALITY_V30_hswq_hybrid_nv100_convrot_nvfp4.safetensorsBeyond REALITYv3.0 (nv100)CreativeML Open RAIL++-M
2127ZImageAsianUtopian_v40Turbo_hswq_hybrid_nv100_convrot_nvfp4.safetensors2127 Z-Image Asian Utopianv4.0 Turbo (nv100)CreativeML Open RAIL++-M
copaxTimeless_xplusZ13_hswq_hybrid_nv100_convrot_nvfp4.safetensorsCopax Timelessxplus Z13 (nv100)CreativeML Open RAIL++-M
unstablebastard_v14_hswq_hybrid_nv100_convrot_nvfp4.safetensorsUnstable Bastardv1.4 (nv100)CreativeML Open RAIL++-M
zimageTurboByStable_2602BF16_hswq_hybrid_nv100_convrot_nvfp4.safetensorsZ-Image Turbo2602 BF16 (nv99/nv100)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.

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 / Dark Beast: Created by catlover1937.
  • —Unstable Revolution / Unstable Bastard: Created by Yamer.
  • —gonzalomo Z-Pop: Created by gonzalomo.
  • —Diving Z-Image Turbo: Created by Diving.
  • —Beyond REALITY: Created by Lykon.
  • —2127 Z-Image Asian Utopian: Created by 2127.
  • —Copax Timeless: Created by Copax.
  • —Z-Image Turbo: Created by Tongyi-MAAS / Civitai Community.

Disclaimer: These models are provided for optimization and research purposes. Please adhere to the original licenses of the base models.