Paint-Anything reports colour-fidelity gains on a benchmark it wrote

single source· 1 articles · confidence: medium · first seen 2026-09-16 20:00 UTC

What this means for you

Nothing to act on yet. No weights, code or API are released, and the headline improvement is measured on a benchmark the same authors wrote, so there is no independent number to check. If exact brand colour matters to you, this is a paper to track, not a tool to adopt.

A preprint posted to arXiv on 16 September describes Paint-Anything, a training method that lets an image model take a 24-bit hex value — the code for one exact colour — and apply it to a named object, in both generation and editing. It is trained on Paint-500K, built from real photographs, plus synthetic colour anchors used only in high-noise training steps. The paper reports gains on FLUX.2-4B of 85.3% and 28.3% on ACBench-T2I and ACBench-Edit — benchmarks it introduces itself, with no independent evaluation and no release of weights or code.

Key facts

  • ·Paint-Anything learns a shared hex-prompt interface for image generation and editing, evaluated on the FLUX.2-4B base model. source
  • ·It is trained on Paint-500K, a dataset built from real images through object grounding, perceptual colour labelling and editing-pair synthesis. source
  • ·The paper reports ACBench-T2I and ACBench-Edit improvements of 85.3% and 28.3% respectively relative to the base model. source
  • ·ACBench is introduced in the same paper and measures object-level hex colour fidelity across generation and editing. source
  • ·Pure-colour anchors whose pixels exactly match their paired hex values are used only at high-noise timesteps; low-noise training uses natural images. source
  • ·The paper was posted to arXiv on 16 September 2026. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire