Treating Refinement Itself as the Editing Interface: How RefineEdit Edits Images Without Training in a Generative Refinement Network

It proposes what the authors describe as the first training-free prompt-to-prompt editing framework for GRN-generated images, formulating editing as a coupled refinement process in which localization and content generation evolve together. Prior training-free editors mostly add spatial control to diffusion models (masks, attention maps, injected features) or select between source and editing predictions under a fixed causal decoding order in autoregressive models; this work moves the editing interface onto GRN's globally revisable binary codes, which the authors say extends GRN beyond image generation and prior training-based video editing. The paper provides method derivations, an algorithm, and quantitative comparison across nine PIE-Bench editing categories, plus ablations, hyperparameter slices, and a user study; this is a methods paper with benchmark evaluation.

It uses the signed probability drop between the two branches for the same source-sampled bits to define both a spatial mask and a bitwise mask, separating which positions may change from which bits within them may change. Diffusion editors mediate spatial control through masks or attention whose accuracy directly shapes editing quality; here the editing evidence is attached to individual binary coordinates and is reassessed as the image evolves. The paper gives the definition of the probability drop, the thresholding rules, and the source-anchored routing equation, and ablates spatial freezing and bit locking separately to observe each mechanism's role.

Adaptive spatial freezing and finite bit locking stabilize routing decisions across refinement steps: the former fixes the spatial mask when the initial editing response is strong enough to limit mask expansion, and the latter keeps bits editable if they satisfied both spatial and bitwise criteria at least once within the latest steps. Instantaneous selection can switch bits on and off as scores fluctuate; these mechanisms decouple editing permission from bit value, since locking preserves permission rather than the value. The paper defines the response measure and the locking window, and reports qualitative differences when either mechanism is removed (without freezing the mask spreads beyond the robot into the surrounding roof and background near its hands; without bit locking the fingers retain more source appearance), while the quantitative ablation table is marked in the text as pending replacement with measured results.

Across nine editing categories of PIE-Bench, RefineEdit achieves the best background-preservation scores in PSNR, LPIPS, MSE, and SSIM among the compared methods, together with the highest whole-image and edited-region CLIP scores. Relative to the flow-based FlowEdit, the paper reports PSNR rising from 25.03 to 30.50 and LPIPS falling from 0.062 to 0.037; the authors note the edited-region advantage over LEDits++ is small and interpret it as comparable semantic alignment alongside improved preservation rather than a substantial semantic gain. Evaluation uses 560 source/editing prompt pairs across nine categories, all baselines edit the same GRN source image, and evaluation masks come from Grounded-SAM and are not given to the method; each edit takes 26.77 seconds on average (single A100, ten repetitions, excluding I/O), comparable to FlowEdit and lower than RF-Inversion, with roughly a speedup over PnP and PnP-DirectInv.

The result targets settings where a source image is generated by GRN and the user wants prompt-to-prompt editing without retraining, external masks, or attention control; it lets edit localization and content generation evolve within the same refinement process, and is especially relevant to localized edits that demand strong background preservation (replacement, addition and removal, pose, color, material, background, and style). The paper names extending to arbitrary image inputs and reducing category-specific tuning as future directions.

Open questions the paper itself raises include: the method depends on GRN source trajectories and does not support arbitrary real images; performance depends on the pretrained generator and the accuracy of probability-based localization; freezing an incomplete mask can exclude parts of the intended editing region while continued mask expansion can alter unrelated content; finite bit locking keeps bits editable but does not guarantee the intended semantic change; and the switch step and thresholds still require tuning, with category-specific settings selected on the evaluation benchmark and generalization to unseen prompts and editing distributions left to be tested. In addition, the quantitative ablation table is marked in the text as pending replacement with measured results, so the quantitative contribution of each stabilization mechanism still awaits final data; this reading covered the full text, but how figures and appendix details are presented affects how individual conclusions can be checked.