HSWQ NVFP4 Quantization for SDXL Significantly Outperforms Native Approach

Zestyclose_Bake3680 · reddit · 2026-08-05

The author benchmarked HSWQ ConvRot NVFP4 quantization on SDXL models. Data shows that while NVFP4 degrades image quality more than INT8, applying HSWQ significantly outperforms native NVFP4.

Across multiple fine-tuned models, HSWQ won in both MSE and SSIM metrics. The author notes this experience will help apply the quantization method to upcoming models like Z Image and Krea2 to minimize degradation and save storage.

Original post →

More from Multimodal

Multimodal channel →