Semantic Granularity Navigation in Image Editing

Despite the generative capabilities of diffusion and flow models, real-image editing remains constrained by a persistent trade-off between semantic editability and structural fidelity. We trace a primary cause of this limitation to the implicit coupling of edit progress with model scale in existing paradigms. Under this coupling, stronger edits typically require visiting noisier states, which spends computation on destabilizing layout before the semantic change is well localized. We introduce NaviEdit, a training-free inference-time controller that decouples edit progress from model scale traversal through a strict self-consistency contract. NaviEdit operates at the rollout level and leaves the underlying pretrained model unchanged. It treats scale as a control input and reallocates a fixed step budget toward semantically responsive intermediate scales instead of destructive high-noise regimes. Experiments show positive average gains across compatible editors and flow backbones, supporting decoupling as a portable inference-time control principle.

Modern AI tools can edit real photos from text instructions, such as changing an object’s color or replacing an item. However, making a stronger edit often causes unwanted changes elsewhere: the background may shift, the layout may warp, or the original object may lose its identity. This paper investigates this problem and identifies a common cause in how many image editors organize the editing process. To make a large change, they often move through highly uncertain intermediate states, where the system has more freedom but less grip on the original image. We propose NaviEdit, a method that guides existing image editors without requiring them to be retrained. NaviEdit directs the editor to spend more steps in the part of the process where meaningful changes are possible while the original image structure is still preserved. This leads to more reliable edits that better match the user’s request while leaving unrelated parts unchanged.