SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models
Source: https://arxiv.org/abs/2411.05007
Structured fields
{
"checked_at": "2026-10-17T00:00:00.000Z",
"usefulness": "Primary paper for 4-bit weight-and-activation quantization of diffusion transformers; explains the low-rank outlier-absorption branch and the Nunchaku fused kernels. Basis for judging 4-bit FLUX claims.",
"source_type": "paper",
"availability": "available",
"source_owner": "MIT HAN Lab and collaborators (Li et al.)",
"next_check_at": "2026-10-24T00:00:00.000Z",
"known_limitations": [
"Speed/memory numbers are the authors' reports on specific GPUs; image-quality loss relative to BF16 not independently assessed here."
]
}Evidence claims
- [sourced_fact] arXiv:2411.05007 by Li et al., first submitted 7 November 2024, latest revision v4 on 8 November 2025; ICLR 2025 Spotlight per the page. source
- [upstream_claim] Abstract claims 3.5x memory reduction for 12B FLUX.1 models, 3.0x speedup over W4A16 on a 16GB laptop RTX 4090 (INT4) and 3.1x on RTX 5090 (NVFP4); experiments cover SDXL, PixArt-Sigma and FLUX.1. source
- [upstream_claim] Research review verdict at export: reviewed (as of 2026-10-17). Recorded by the research pipeline, not a hub review outcome; evidence.review stays unreviewed until an independent reviewer records one.
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | arXiv-nonexclusive-distribution-1.0 | unknown | unknown | unknown | unknown | Page links arXiv non-exclusive distribution license v1.0; no other terms stated. |