Paint-Anything fine-tunes generative model colors with hexadecimal precision: an 85 percent jump in generation tests

Color accuracy remains one of the silent obstacles that hinder the integration of image generation models in advertising production pipelines and professional design, where typical textual prompts fail to precisely match brand visual identities. To address this shortcoming, researchers disclosed in a detailed technical report published on the “Archive” platform a new system called “Paint-Anything”, which provides a unified interface for full control of colors via 24-bit hexadecimal codes, combining text-to-image generation with editing according to strict color specifications at the level of independent elements.

Previous attempts to tune colors in generative models relied on custom complex representations or hybrid inference mechanisms that increase computational cost and limit editing flexibility.The new system starts from a core observation that integrated language models have become capable of automatically linking hexadecimal codes to color meanings.This allowed researchers to build a unified interface that accepts digital codes directly within the guiding text, imposing precise supervision at the level of each element in the visual scene without the need for separate inference processors.

To train the model on this capability, the team developed a data processing pipeline that produced the “Paint-500K” dataset of realistic images, relying on spatial identification of elements, perceptual color classification, and synthesis of image pairs for editing. Training encountered a physical obstacle in that shadows and lighting in real photos make color tags approximate rather than matching the abstract numeric value. The researchers overcame this by inventing “pure color anchors”, pixels whose values mathematically correspond to the hexadecimal code, restricting their use to high-noise timesteps in diffusion models and leaving low-noise steps for natural images to preserve scene realism and fine details.

To measure performance accurately, independent of general visual impressions, the paper introduced the ACBench metric, with two parts dedicated to generation and editing, to assess adherence to the target color code for each element. When the technique was evaluated on the FLUX.2-4B model, the system achieved an 85.3 % jump in the generation test compared with the baseline model, and a 28.3 % improvement in the editing test, also recording the highest average score on the “CompColor” benchmark among competing solutions.

This shift has a directly practical impact on design studios, advertising agencies, and e-commerce platforms in the Gulf and Egypt, where teams waste long hours on manual post-editing to adjust corporate identity or product color shades after AI-generated output. Enabling models to read precise digital codes reduces visual production costs and eliminates the need for additional color-correction tools, moving generative models from early-draft utilities to production-ready engines ready for immediate integration into local digital marketing campaigns and e-commerce.