Paint-Anything: Unified Any-Color Control for Image Generation and Editing
The paper’s claim is precise hex-level color control for both generation and editing, using one shared prompt interface.
Paint-Anything trains on object-level color supervision from a 500K-image pipeline, then adds pure-color anchors to handle cases where shadows make real-image color labels imprecise. The authors test it with ACBench, a benchmark split across text-to-image and editing tasks. On FLUX.2-4B, they report gains of 85.3% on ACBench-T2I and 28.3% on ACBench-Edit over the base model. ArXiv · AI/CL/LG's note
Paint-Anything trains on object-level color supervision from a 500K-image pipeline, then adds pure-color anchors to handle cases where shadows make real-image color labels imprecise. The authors test it with ACBench, a benchmark split across text-to-image and editing tasks. On FLUX.2-4B, they report gains of 85.3% on ACBench-T2I and 28.3% on ACBench-Edit over the base model. ArXiv · AI/CL/LG's note
score 5