Diffusion Image Editing via Asynchronous Token Decoding

Text-guided image editing often suffers from global drift, where changing a prompt to modify a specific attribute (e.g., changing a dog to a cat) inadvertently alters the background or layout of the entire image. The authors investigate how to achieve targeted local edits in Diffusion Transformers (DiTs) without requiring external masks, user-provided guidance, or model fine-tuning.

ATDEdit reformulates the diffusion sampling process as an asynchronous, token-indexed operation. Instead of updating all latent tokens synchronously under the new target prompt, the framework uses a two-stage process:

1. Sensitivity Estimation: It calculates a "Token-wise Conditional Surprisal" score, which measures how much each token's prediction changes when the condition is switched. This identifies which tokens are most relevant to the requested edit.
2. Asynchronous Updating: The framework applies target-conditioned updates only to the high-surprisal (editable) tokens. For the remaining "keep" tokens, it enforces background preservation by projecting their latent states back to the source values and supplying their key/value memory from the source branch during self-attention.

This approach provides a flexible, training-free way to perform high-fidelity image editing. By moving away from static, pre-computed masks and toward dynamic, model-driven introspection, ATDEdit allows the editing process to adapt to the evolving semantics of the reverse diffusion trajectory. Experiments on the PIE-Bench dataset demonstrate that this method achieves superior preservation of the original image structure while maintaining competitive semantic alignment with the target text.