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

Paint-Anything. Unified Any-Color Control for Image Generation and Editing. By Ji Xie and four others, from ByteDance Seed, Zhejiang University, and one other institution. Published on the seventeenth of September. No code repository is linked to it.

Professional design needs any-color control: the ability to specify an object's target color as any twenty-four-bit hex value, when generating an image and when editing one. Prior work on color generation, editing, and colorization often relies on dedicated color representations or specialized inference procedures. Advances in large language models suggest a simpler starting point, since even compact models can associate hex values with the meaning of colors. If a language model already knows roughly what a hex code looks like, the code itself can go straight into the prompt.

Paint-Anything learns a single shared hex-prompt interface for both generation and editing, supervised by color at the level of objects. The same kind of instruction, an object and a hex value, drives both tasks, so there is no separate color representation to maintain and no special procedure at inference. What makes it work is the data and the way it is used in training.

The supervision comes from Paint-500K, which a data pipeline builds from real images through object grounding, perceptual color labeling, and synthesis of editing pairs. Real images have shadows, so their labels only approximate a color. To complement them, the authors add pure-color anchors, whose pixels exactly match the hex value they are paired with. The anchors are used only at high-noise timesteps. Low-noise training is left to natural images.

To measure the result, the authors also introduce a benchmark of their own. Measurement comes from the Any Color Benchmark, ACBench, in two parts, ACBench-T2I and ACBench-Edit, which test hex color fidelity at the object level in both tasks. The question each one asks is simple: did the named object come out in the named color.

On FLUX point two, in its four-billion-parameter version, Paint-Anything improves ACBench-T2I scores by eighty-five point three percent and ACBench- Edit scores by twenty-eight point three percent, relative to the base model. Ablations support the training recipe. On a separate measure, it also has the highest average CompColor score among the methods compared.

The recipe pairs two kinds of supervision: exact pure-color anchors while the image is still mostly noise, and approximate labels from real images everywhere else. Through it, one interface, a hex code in the prompt, serves generation and editing alike.