Skip to content

Back to the catalog

Paint-Anything: Unified Any-Color Control for Image Generation and Editing
arxiv.orgpaper

Paint-Anything: Unified Any-Color Control for Image Generation and Editing

Paint-Anything introduces a shared hex-color prompt interface for both image generation and editing, letting users specify an object's exact 24-bit hex color. The authors build the Paint-500K dataset via object grounding, perceptual color labeling, and editing-pair synthesis, and supplement real-image supervision with pure-color anchors used only at high-noise diffusion timesteps to counter shadow-induced label imprecision. Evaluated on FLUX.2-4B with the new Any Color Benchmark (ACBench-T2I/ACBench-Edit), it improves color-fidelity scores by 85.3% and 28.3% respectively over the base model.

TL;DR A hex-prompt method and new benchmark (ACBench) for precise any-color control in diffusion-based image generation and editing.
Document

What it claims

Beyond the base-model deltas the summary reports, the abstract makes two claims the summary omits. First, the enabling premise: even compact language models already associate 24-bit hex values with color semantics, so no dedicated color representation or specialized inference procedure is needed — the method rides on a plain hex text prompt. Prior color-generation, editing, and colorization work, the authors say, relied on such dedicated representations. Second, on comparison against other methods (not just the FLUX.2-4B base), Paint-Anything reports the highest average CompColor score among the compared methods — a competitive result, distinct from the 85.3% / 28.3% ACBench gains over base that the summary already carries.

Evidence

The page (abstract only) states two evidence classes:

  • Vs. base model: ACBench-T2I +85.3%, ACBench-Edit +28.3% on FLUX.2-4B, with ablations said to support the training recipe.
  • Vs. other methods: highest average CompColor score among compared methods.

The abstract does not name the competing methods, give their CompColor numbers, or state the absolute (not relative) ACBench scores. The relative gains have no absolute anchor, so their practical magnitude is unverifiable from this page.

Limits

  • This is an abstract-only page; no method detail beyond what the summary already extracts (grounding → perceptual color labeling → editing-pair synthesis; pure-color anchors confined to high-noise timesteps).
  • The CompColor comparison names neither baselines nor scores, so "highest average" cannot be checked or weighted here.
  • Demonstrated on a single backbone, FLUX.2-4B; no evidence about transfer to other diffusion models.
  • The shadow-imprecision problem is acknowledged by the authors as the reason real-image labels are only approximate — a stated weakness in the supervision, mitigated but not eliminated by the anchor trick.
  • No code, weights, dataset link, or license is present on this page; Paint-500K availability is unstated. A v2 exists (20 Sep 2026) but its changes are not described.

If you need absolute fidelity numbers, the competitor list behind CompColor, or dataset/weight access, this page will not settle it — fetch the PDF or v2.

Organisations named here

  • ACBench
  • CompColor
  • Paint-500K

Technologies named here

  • Paint-Anything
  • FLUX.2-4B

Techniques named here

  • Hex-prompt interface
  • Object-level color supervision
  • Data pipeline for Paint-500K
  • Pure-color anchors