The paper introduces Verifiable Visual Rewards (VVR), a framework that scores generated images with deterministic Python functions over pixels instead of learned detectors, OCR, or vision-language models. Each task is a synthetic scene of colored geometric shapes with constraints on count, color, shape, and spatial relations; both the prompt and the verifier are derived programmatically, so tasks can be generated at any scale and difficulty.
Training Stable Diffusion 3.5 Medium with reinforcement learning on VVR scores raised its accuracy on VVRBench from 2.8% to 28.3%, with gains on tasks harder than those seen in training and transfer to out-of-domain natural prompts. Mixing VVR rewards with existing post-training objectives improved general performance and human preference. The authors also release VVRBench and VVRBench-Challenge, on which the strongest evaluated model, GPT-Image-2.5, solves only 21.4% of the harder set.