Chain-of-thought (CoT) reasoning has proven effective in improving the reasoning abilities of large language models by breaking complex tasks into step-by-step solutions. In recent years, researchers have attempted to extend CoT reasoning to visual generative models to improve their ability to handle complex tasks. However, existing CoT methods for text-to-image generation still rely mainly on textual reasoning and therefore cannot fully use and preserve the correct semantic information already present in an image, as even the most detailed language cannot capture all visual information. With advances in image generation, models such as FLUX and BAGEL can generate images conditioned on visual inputs. Recent work has attempted to carry images forward as context in the chain of thought, but this introduces another limitation: unsuitable earlier images can strongly interfere with the model, while repeated whole-image regeneration introduces new hallucinations at each round and causes them to accumulate. To address these issues, we propose MLIE-CoT, a visual chain-of-thought framework based on multi-level image editing. MLIE-CoT learns to select the repair granularity from three levels, preserving correct semantics whenever possible during reasoning while reducing the effect of incorrect content. To improve the ability of multimodal models to route errors of different severity to appropriate repair levels, we construct an image-repair dataset covering different error cases and use it to train a repair-level router. Extensive experiments demonstrate that MLIE-CoT significantly outperforms both its base models and methods using fixed repair strategies, producing clear improvements across multiple benchmarks.
Figures & tables
Figure 1: Failure cases caused by negative visual anchors and collateral changes during image-conditioned repair.
Figure 2: Quality degradation caused by repeated whole-image regeneration. Across both sequences, successive corrections introduce artifacts and overwrite content that was initially correct.
Figure 3: Representative results for the three routed repair levels.
Figure 4: Overview of the MLIE-CoT framework. The analyzer identifies mismatches between the prompt and current image. A frozen multimodal reasoning backbone feeds its representation to a trainable three-way prediction head, which selects L1 local editing with cropped regions, L2 whole-image editing conditioned on the current image, or L3 regeneration by the prompt refiner and generator. The critic examines the new image: a pass returns the final image and terminates, whereas a failure sends the result back for another diagnosis and repair round.
Level
Action
Typical scenario
L0
Accept current image
The current image satisfies the constraints in the target prompt.
L1
Local repair
Errors in color, attributes, counts, or a few objects that affect only a small number of regions.
L2
Whole-image editing
Spatial relations, background, or layout require adjustment, but the current image remains worth preserving.
L3
Text-conditioned regeneration
The main concept or overall composition is incorrect, or multiple connected errors are difficult to edit one by one.
Table 4: Fixed-level and router ablations on the same set of initial images.
Figure 5: Effect of the number of MLIE-CoT steps on generation quality and cost for BAGEL and FLUX.2 on GenEval and WISE.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Comparison
Base [95% CI]
MLIE-CoT [95% CI]
Δ [95% CI]
Bagel on GenEval
0.8082 [0.7791, 0.8359]
0.8566 [0.8279, 0.8840]
+0.0484 [+0.0430, +0.0555]
Bagel on WISE
0.7040 [0.6780, 0.7300]
0.7600 [0.7330, 0.7860]
+0.0560 [+0.0400, +0.0720]
Flux.2 on GenEval
.848 [.819, .876]
.884 [.857, .909]
+.036 [+.018, +.054]
Flux.2 on WISE
.734 [.708, .760]
.778 [.752, .803]
+.044 [+.028, +.060]
Appendix
Table S1: Paired uncertainty and significance results for the two specified same-prompt comparisons. Raw/Holm-adjusted randomization p -values are .1100/.1100 for WISE and .0003/.0006 for GenEval. The WISE interval for the improvement includes zero and is not significant after correction; the GenEval improvement remains significant.
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.