Inline Critic Steers Image Editing

2026-05-12

AI Summary

Existing image editing methods rely on post-hoc corrections and struggle to rectify errors in challenging regions during the generation process. This work proposes a novel inline critic mechanism embedded within the forward pass: by introducing learnable Inline Critic tokens, it enables real-time assessment and guidance of the generative process at intermediate layers of a frozen diffusion model, facilitating dynamic and fine-grained editing. For the first time, critique signals are integrated directly into the forward propagation, combined with a three-stage training strategy, intermediate-layer error analysis, and attention modulation to substantially enhance editing fidelity. The method achieves state-of-the-art performance among open-source approaches across multiple benchmarks—scoring 7.89 on GEdit-Bench, gaining +9.4 points on RISEBench, and attaining 81.92 on KRIS-Bench—surpassing even GPT-4o.

Abstract

Instruction-based image editing exhibits heterogeneous difficulty not only across cases but also across regions of an image, motivating refinement approaches that allocate correction to where the model struggles. Existing refinement signals arrive late, after a fully generated image or a completed denoising step. We ask whether such a signal can act within an ongoing forward pass. To investigate this, we probe a frozen image-editing model and find that although generation capability emerges only in the last few layers, the error pattern is already set in early layers (rank correlation \r{ho} = 0.83 with the final-layer error map). Based on this, we introduce Inline Critic, a learnable token that critiques a frozen model's predictions at its intermediate layers and steers its hidden states to refine generation during the forward pass. A three-stage recipe is proposed to stabilize the training from learning how to critique to steering generation. As a result, we achieve state of the art on GEdit-Bench (7.89), a +9.4 gain on RISEBench over the same backbone, and the strongest open-source result on KRIS-Bench (81.92, surpassing GPT-4o). We further provide analyses showing that the critic genuinely shapes the model's attention and prediction updates at subsequent layers.

Problem

• image editing
• instruction-based editing
• refinement signal
• forward pass
• error pattern

Innovation

• Inline Critic
• instruction-based image editing
• intermediate-layer refinement
• frozen model steering
• real-time error correction