CacheEdit: Efficient Multi-round Image Editing via Adaptive Token-wise Reuse.

Instruction-based image editing (IIE) is a vital tool for iterative content creation, enabling multi-round interactions that refine visual details while preserving cross-round consistency. However, this workflow is constrained by the compute-bound nature of Diffusion Transformers (DiTs): because DiTs process tokens uniformly, they waste substantial computation on regions untouched by the instruction. We investigate the Round--Step--Layer hierarchy of DiT-based editing and identify a phenomenon we term Delayed Latent Emergence (DLE). Although pronounced latent changes emerge only in the late denoising stages, deep-layer activations within transformer blocks at the very first sampling step already diverge markedly in edited regions. Building on this insight, we propose CacheEdit, a training-free framework centered on an Adaptive Activation Cache (Acache) that exploits early-step sensitivity to detect invariant tokens and reuse their cached activations across subsequent sampling steps, thereby bypassing redundant computation. Experiments on FLUX.1 Kontext and Qwen-Image-Edit show that CacheEdit achieves up to $2.5\times$ end-to-end acceleration. Moreover, by isolating and reusing static features, CacheEdit mitigates stochastic drift and improves instruction-following and structural consistency over full-recomputation baselines.

Instruction-based image editing lets users refine an image through multiple rounds of natural-language instructions, such as changing one object while keeping the rest of the scene unchanged. However, modern editing models based on Diffusion Transformers are computationally expensive because they process every image region equally, even when only a small part needs to be changed. In this work, we study how these models behave across editing rounds, denoising steps, and network layers. We find that although visible image changes appear only near the end of the generation process, the model’s internal features already reveal which regions will be edited at the very beginning. Based on this observation, we introduce CacheEdit, a training-free method that identifies image tokens likely to remain unchanged and reuses their stored activations in later denoising steps. This avoids redundant computation while preserving important visual details. Experiments on FLUX.1 Kontext and Qwen-Image-Edit show that CacheEdit speeds up editing and can also improve consistency by reducing unintended changes in regions that should stay fixed.