Training-free object removal typically works by redirecting self-attention away from the target region. But on few-step distilled diffusion models, this approach often leaves behind object-shaped residuals, limiting its practical use.
The authors of ReState trace this failure to a temporal mismatch: the correction is applied at the wrong point in the denoising trajectory. Their method adjusts the latent state according to the current noise level, aligning the removal operation with the model's step schedule.
Because ReState requires no training or fine-tuning, it can be dropped into existing few-step pipelines. The paper reports that this noise-level-aware correction suppresses residual artifacts while preserving the surrounding image content.