Inline Critic: Real-Time Image Editing Correction

Researchers propose Inline Critic, a method that corrects image editing errors during a model's forward pass. By probing a frozen image-editing model, they found that error patterns are established in early layers (rank correlation 0.83 with final-layer error map), while generation capability emerges only in later layers. Inline Critic uses a learnable token to critique intermediate predictions and steer hidden states, enabling refinement without waiting for full generation. A three-stage training recipe stabilizes learning from critique to steering. The approach addresses heterogeneous difficulty across image regions, offering a more efficient alternative to post-hoc refinement. The paper is published on arXiv (2605.12724).