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

A shared hex-prompt interface lets generation and editing use the same object-level color control mechanism. Prior color generation, editing, and colorization work often relies on dedicated color representations or specialized inference procedures, whereas this work uses hex values directly as prompts, drawing on language models' association between hex values and color semantics. The abstract provides a method-level description and quantitative results on FLUX.2-4B, constituting author-reported experimental evidence.

The Paint-500K data pipeline constructs training data from real images via object grounding, perceptual color labeling, and editing-pair synthesis. The pipeline scales object-level color supervision and supplies a unified training signal for both generation and editing tasks. The abstract states the pipeline comprises object grounding, perceptual color labeling, and editing-pair synthesis, without giving dataset scale details in the abstract.

Pure-color anchors complement real-image supervision and are used only at high-noise timesteps. Because shadows make real-image color labels only approximate, the work introduces pure-color anchors whose pixels exactly match their paired hex values, while leaving low-noise training to natural images. The abstract explicitly states the timing and purpose of the anchors and reports that ablations support the training recipe.

The Any Color Benchmark (ACBench), comprising ACBench-T2I and ACBench-Edit, measures object-level hex color fidelity across both tasks. It provides a comparable evaluation dimension for any-color control rather than a metric for a single task. The abstract gives the benchmark composition, the improvement margins on FLUX.2-4B, and states the highest average CompColor score among compared methods.

The work targets image generation and editing scenarios that require specifying an object's color with any hex value, suited to users with object-level color control needs such as design and content-creation workflows; its training recipe depends on the Paint-500K data pipeline together with pure-color anchors at high-noise timesteps, and evaluation centers on object-level hex color fidelity in ACBench-T2I and ACBench-Edit, with results reported on FLUX.2-4B.

The abstract does not give the specific scale of Paint-500K, the composition details and evaluation protocol of ACBench, or the proportion of pure-color anchors and how high-noise timesteps are divided; these details, along with performance on base models beyond FLUX.2-4B, remain questions for readers of the full text.