Researchers have unveiled Paint-Anything, a method that aims to give designers precise control over the color of individual objects in generated and edited images. The work targets what the authors call "any-color control": the ability to specify an object's target color using any 24-bit hex value.
Rather than building dedicated color representations or specialized inference pipelines, the approach learns a shared hex-prompt interface for both generation and editing, relying on object-level color supervision. The authors note that even compact large language models can already associate hex values with color semantics, offering a simple starting point.
To train the system, the team developed a data pipeline that constructs Paint-500K from real images via object grounding, perceptual color labeling, and editing-pair synthesis. Because shadows make real-image color labels only approximate, the pipeline adds pure-color anchors whose pixels exactly match their paired hex values. These anchors are applied only at high-noise timesteps, while low-noise training uses natural images.
The researchers also introduce the Any Color Benchmark (ACBench), split into ACBench-T2I and ACBench-Edit, to measure object-level hex color fidelity across both tasks. On FLUX.2-4B, Paint-Anything reportedly improves ACBench-T2I and ACBench-Edit scores by 85.3% and 28.3% respectively over the base model, with ablations supporting the training recipe. It also achieves the highest average CompColor score among compared methods.
The paper page also invites readers to try any-color control in Seedream 5.0 Pro, with more in the Seedream series promised.