A research team has trained a single image model to inspect and revise its own generated pictures. On GenEval, a 553-prompt benchmark that checks details such as object counts, colors and positions, the method repaired 64.9% of images that failed the first check, compared with 20.6% for a supervised fine-tuned version of the same model.
Called UMM-Reflection, the system builds on BAGEL, a model that can both understand and generate images. It creates an initial picture, checks that picture against the prompt, writes a specific edit instruction and renders a revision. The process can repeat for up to three rounds.
The researchers changed how the model learns to make those edits. Reinforcement learning compares 16 complete correction paths that start from the same image. A frozen verifier scores the paths during training, and the resulting full-path score updates both the model’s written diagnosis and its image generator. The verifier is not used when the model runs.
The gain came from editing rather than a better first render: the paper reports no statistically significant difference in initial GenEval scores, and the RL stage does not update initial image generation. The authors’ analysis suggests the training mainly helps the model choose effective corrections it could already produce, rather than giving it a new way to perceive images. Counting accuracy remained 67.5% from the RL model’s first image to its final image.
The authors released the training and evaluation code, model checkpoints and supervised-training data on GitHub and Hugging Face. The result is a benchmark study of one BAGEL-based research system, not evidence that image generators generally can repair arbitrary mistakes.
Sources
Image credit: Yijia Fan et al., UMM-Reflection paper, CC BY 4.0.