This paper introduces Paint-Anything, a method that lets users specify an object’s exact color using any 24‑bit hex value and have a model generate or edit images to match. A 24‑bit hex value is the familiar six‑digit color code (for example #RRGGBB) that designers use to pick precise colors. The goal is a simple, unified “hex‑prompt” interface that works both when making new images and when changing colors in existing ones.

The team trained models to understand and follow hex color prompts at the object level. They built a dataset called Paint‑500K from real images by finding objects in photos, estimating their perceptual colors, and making pairs of images for editing examples. The authors also relied on large language models’ surprising ability to link hex codes to color descriptions as a starting point for the interface.

Real photos are imperfect because lighting and shadows change how a color appears. To handle that, the researchers added pure‑color anchors: small images whose pixels exactly match a target hex value. These anchors provide exact color supervision, but they used them only during high‑noise parts of training and left the low‑noise training steps to real images. That balance lets the model learn exact color values while still matching natural image appearance.

To measure how well the system follows hex color instructions, they created the Any Color Benchmark (ACBench). ACBench has two parts: ACBench‑T2I for text‑to‑image color fidelity and ACBench‑Edit for editing fidelity. On the FLUX.2‑4B base model, Paint‑Anything improved ACBench‑T2I scores by 85.3% and ACBench‑Edit scores by 28.3% compared with the same base model without the Paint‑Anything training. The paper also reports the highest average CompColor score among compared methods and presents ablation studies that support their training choices.