HSVDQuant preprint reports faster four-bit model conversion with smaller correction components

Researchers report that HSVDQuant improves four-bit model quantization while using smaller high-precision correction components than SVDQuant. Quantization reduces the precision of model weights and intermediate values; the new method jointly optimizes the low-bit representation and its correction component while accounting for errors in intermediate values. Across five diffusion backbones, rank-4 HSVDQuant outperformed rank-4 SVDQuant on PSNR and LPIPS measures. It also surpassed rank-32 SVDQuant on SANA-1.6B, FLUX.1-schnell and FLUX.1-dev with an 8× smaller rank and up to 6.25× faster quantization. On Qwen3-8B, it raised MMLU accuracy from 61.50% to 68.17% against rank-32 SVDQuant. These are results reported in a preprint.