← News

ResearchSources: arXiv · 20 Nov 2026

Chameleon picks quantization formats per layer for diffusion models

An arXiv paper proposes post-training quantization that treats the number format itself as a tunable choice, reporting better image quality and a measured speedup on diffusion models.

In plain words

Diffusion models that generate images are usually compressed so they fit in limited memory, but existing compression methods fix the number format in advance. Chameleon instead chooses the best format for each weight channel and each layer, using simple statistics. The authors report better image quality than fixed-format methods and a modest speedup on one tested model.

What was announced

The paper argues that at a fixed bit-width, the best number format depends on the distribution being encoded, and those distributions differ across weight channels, layers, and diffusion timesteps, where activations shift from heavy-tailed and noise-dominated to tightly clustered. Chameleon therefore holds the bit-width fixed and treats the format as a discrete variable, selected per channel and per layer, with optional refinement per timestep bucket. Activation formats are drawn from a set including INT8, FP8 variants, MXFP8, and MXINT8, chosen from two cheap statistics and stored in a lookup table; weight formats come from 8-bit and 4-bit sets selected offline by reconstruction error. The authors say an architectural fork adapts the selection layer to multi-step UNets, single-step distilled models, and Diffusion Transformers.

On SDXL, SDXL-Turbo, and PixArt-alpha over COCO-2014, the authors report the best FID in all six backbone and bit-width settings, with CLIP scores within 0.24 of the FP16 reference and the best result among quantized methods at W4A8. On PixArt-alpha, a real mixed-format kernel is reported at 1.09x faster than FP16 end-to-end with 3x smaller weights. The authors attribute most of the gain to the per-layer format choice, noting the timestep refinement helps mainly at 6-bit activations. These are the paper's own measurements on its abstract, not independent confirmations.

What it does not show

  • The evidence is the paper's abstract only; the full article, implementation details, and code availability are not covered.
  • All quality and speed figures are the authors' own reported results and have not been independently verified.
  • The abstract does not state whether the code or kernels are publicly available.

For security and IT teams

Scope
The reported 1.09x speedup and 3x smaller weights apply to PixArt-alpha specifically, not to all tested backbones.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.