Text-to-image repair stops redrawing the whole frame: MLIE-CoT picks repair granularity by error size, lifting GenEval overall from 0.84 to 0.88
Synopsis
Images generated from complex prompts often keep small local errors, and existing visual chain-of-thought methods either rewrite the prompt or regenerate the whole image, discarding correct visual content or accumulating new hallucinations each round. MLIE-CoT splits correction into four actions of increasing scope, accept, local repair, whole-image editing and text-only regeneration, and a router trained on image-repair data chooses the level from the current image, target prompt and a structured diagnosis. On GenEval, MLIE-CoT raises overall score from 0.84 to 0.88 with FLUX.2-dev as backbone and from 0.80 to 0.86 with BAGEL.
Interpretation
Chain-of-thought correction for image generation can now touch only the small region that is wrong instead of redrawing the whole image each round. Earlier visual chain-of-thought methods mostly reduced reflection to a longer prompt, a layout condition or another round of whole-image generation, discarding correct content for small errors and carrying a wrong layout forward for large ones. On the 553 GenEval prompts, MLIE-CoT with FLUX.2-dev as backbone reaches an overall score of 0.88, above the 0.84 of that backbone alone; with BAGEL it reaches 0.86, above 0.80.
The system first diagnoses constraint-by-constraint differences between the current image and the target prompt, then decides how much of the existing image to keep. The router is a frozen Qwen3.5-9B multimodal encoder plus a two-layer classification head, and only the head is trained to predict L1 local repair, L2 whole-image editing or L3 text-only regeneration from the current image, target prompt and fixed-schema diagnosis. Routing training data comes from 1K prompts sampled five times per repair level, 14.8K images in total; after repair a review step keeps whichever of the original and repaired image better satisfies the prompt.
Given the same initial images, the learned router is more accurate than any single fixed repair policy. Using only L1, only L2 or only L3, or letting the model route coarsely on its own, covers only part of the error types. In a GenEval comparison on the FLUX.2-dev backbone where all four policies share the same 553 initial images, L1-only, L2-only, L3-only and self-route average 0.86, 0.85, 0.85 and 0.85, while MLIE reaches 0.88.
Perspective
The demonstrated scope is 553 GenEval prompts and 1,000 WISE prompts, with BAGEL-7B-MoT and FLUX.2-dev as backbones and the official prompts and evaluators; the authors propose future targeted optimization of repair prompts and exploring image-generation chain of thought for larger images.
The authors note a clear ceiling on the gains from more repair steps, which depend jointly on the task's reasoning demands and the correctable error space left by the base model, with one to three steps the most efficient range; extra steps do not necessarily improve quality and may only add cost.
