ApacheOne/HSWQ-fp8-SD15

absolutereality_v181_r32_r0.1_HSWQ_fp8e4m3fn.safetensors This is only the sd15 unet layers at HSWQ_fp8e4m3fn

Example images and workflow inside, workflow is not the best as ComfyUI and other pipelines have dropped the ball for sd15.

Positive rompt: realistic, a photo of a gothic horror woman in a black lake, beautiful woman, unsettling horror, professional, (Extremely Detailed:1.2), glow effects, godrays, intricate details, sharp focus, dramatic, photorealistic, tribal village, wooden village, burning forest background, sharp contrast, many colours, serious face, smirking, Eva Green

Negative embeddings files: BadDream, UnrealisticDream

absolutereality_v181_r32_r0.1_HSWQ_fp8e4m3fn.safetensors

absolutereality_v181.safetensors

Hybrid-Sensitivity-Weighted-Quantization (HSWQ)

High-fidelity FP8 quantization for diffusion models (SDXL). HSWQ uses sensitivity and importance analysis instead of naive uniform cast, and offers two modes: standard-compatible (V1) and high-performance scaled (V2).

Technical details: md/HSWQ_ Hybrid Sensitivity Weighted Quantization.md

How to quantize: md/HSWQ_ How to quantize SDXL.md

SDXL Benchmark Test Results: md/SDXL Benchmark Test Results.md

Credit & Special Acknowledgement

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

https://github.com/tritant/ComfyUI_Kitchen_nvfp4_Converter

https://github.com/NVIDIA/Model-Optimizer

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