Hugging Face's logo

ussoewwin
/
Hybrid-Sensitivity-Weighted-Quantization-SDXL-ConvRot-INT8

Text-to-Image
nunchaku
sdxl
svdq
quantized
int8
illustrious
realvisxl
photorealistic
comfyui
controlnet
anime
faipl-1.0-sd
creativeml-openrail-m
Model card Files
xet
 Community
Hybrid-Sensitivity-Weighted-Quantization-SDXL-ConvRot-INT8 / README.md
ussoewwin's picture
ussoewwin
Upload README.md with huggingface_hub
1d58b7e verified
preview code
|
Raw
History Blame Contribute Delete
6.8 kB
metadata 
license: other
tags:
  - text-to-image
  - sdxl
  - nunchaku
  - svdq
  - quantized
  - int8
  - illustrious
  - realvisxl
  - photorealistic
  - comfyui
  - controlnet
  - anime
  - faipl-1.0-sd
  - creativeml-openrail-m
library_name: nunchaku

Hybrid-Sensitivity-Weighted-Quantization (HSWQ)

High-fidelity ConvRot INT8 quantization for diffusion models (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-compatible int8_tensorwise pack with FULL ConvRot on remaining Linear/Conv2d after DualMonitor + V4 weighted-histogram FP16 protection under a fixed 300 MiB budget. Keep ratio is 0 (r0); critical layers stay FP16 via automatic analysis, not a keep-ratio percentage. SDXL pack scripts: quantize_sdxl_hswq_v3.1.py (HSWQ) and native_convert_int8_sdxl.py (native).

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

How to quantize (SDXL ConvRot INT8): md/How to quantize SDXL.md

ComfyUI Loader for ConvRot INT8 / INT8: To load these INT8 models in ComfyUI, please use the unofficial loader node: ComfyUI-nunchaku-unofficial-loader

Post-quantize fidelity bench (integrated, default ON): After save, quantize_sdxl_hswq_v3.1.py and native_convert_int8_sdxl.py clear parent VRAM, then automatically run benchmark/int8bench_sdxl.py with --fp16 = the FP16 input, --int8 = the saved pack, and a fixed --prompt / --seed (not inventable parent CLI overrides). Pass --no-bench to skip. Standalone re-runs use the same int8bench_sdxl.py command shape as in the How-to.

SDXL ConvRot INT8 Benchmark Test Results (published tables): test/benchmark_sdxl_int8.md

Benchmark (Reference)

Model SSIM (Avg) File size Compatibility
Original FP16 1.0000 100% High
Naive INT8 0.95-0.97 50% High
HSWQ ConvRot INT8 0.94-0.98 68% (FP16 mixed) High (ComfyUI INT8)

📦 Available Models

Filename Base Model Version License
JANKUTrainedChenkinNoobai_v777_hswq_r32_1off_convrot_int8.safetensors JANKU Trained Chenkin & Noobai-Rouwei (Illustrious-XL) v777 Fair AI Public License 1.0-SD
bluePencilXL_v031_hswq_r32_1off_convrot_int8.safetensors blue_pencil-XL v0.3.1 CreativeML Open RAIL++-M
epicrealismXL_pureFix_hswq_r32_1off_convrot_int8.safetensors epiCRealism XL pureFix CreativeML Open RAIL++-M
koronemixIllustrious_v70_sci_1on_covrot_int8.safetensors koronemixIllustrious v70 Fair AI Public License 1.0-SD
koronemixVpred_v20_sci_1off_convrot_int8.safetensors koronemixVpred v2.0 CreativeML Open RAIL++-M
novaAnimeXL_ilV190_hswq_r32_1on_convrot_int8.safetensors Nova Anime XL ilV190 Fair AI Public License 1.0-SD
novaAsianXL_illustriousV70_hswq_r32_1off_convrot_int8.safetensors Nova Asian XL v7.0 Fair AI Public License 1.0-SD
oneObsession_v23_hswq_r32_1off_convrot_int8.safetensors OneObsession v23 CreativeML Open RAIL++-M
prefectIllustriousXL_v8_hswq_r32_1on_convrot_int8.safetensors Prefect Illustrious XL v8 Fair AI Public License 1.0-SD
realvisxlV30_v30TurboBakedvae_hswqr32_r32_1on_convrot_int8_.safetensors RealVisXL V3.0 (Turbo) v3.0 Turbo CreativeML Open RAIL++-M
realvisxlV50_v40Bakedvae_hswq_r32_ioff_convrot_int8.safetensors RealVisXL V5.0 (Lightning) v4.0 BakedVAE CreativeML Open RAIL++-M
realvisxlV50_v50Bakedvae_hswq_r32_1on_covrot_int8.safetensors RealVisXL V5.0 (Lightning) v5.0 BakedVAE CreativeML Open RAIL++-M
uwazumimixILL_v50_hswq_r32_1on_convrot_int8.safetensors UwazumiMix v5.0 Fair AI Public License 1.0-SD
waiIllustriousSDXL_v170hswq_r32_1off_convrot_int8.safetensors Illustrious-XL v1.7 (WAI-illustrious-SDXL) v17.0 (HF weight) Fair AI Public License 1.0-SD
waiREALCN_v150_hswq_r32_1on_convrot_int8.safetensors WAI-REAL_CN v15.0 Fair AI Public License 1.0-SD
waiREALISM_v10_hswq_r32_1on_convrot_int8.safetensors WAI-REALISM v1.0 Fair AI Public License 1.0-SD

📜 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.

  • JANKU Trained Chenkin & Noobai-Rouwei (Illustrious-XL): Created by janxd.
  • blue_pencil-XL: Created by Euge_us.
  • epiCRealism XL: Created by epinikion.
  • WAI-illustrious-SDXL / WAI-REAL_CN / WAI-REALISM: Created by WAI0731.
  • koronemixIllustrious / koronemixVpred: Created by koronen.
  • Nova Anime XL / Nova Asian XL: Original creator on Civitai.
  • Prefect Illustrious XL: Created by Goofy_Ai.
  • OneObsession: Created by Polyhedron.
  • RealVisXL: Created by SG_161222.
  • UwazumiMix: Created by UWAZUMI.

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