Researchers built a way to tell an AI image editor exactly how much of a region to keep versus repaint, instead of just picking a shape and hoping.
The method, called SoftPaint and described in an arXiv paper, targets a real limitation in diffusion-based editing tools. Current systems use masks that are mostly binary: a pixel is either protected or it gets redrawn. SoftPaint instead reads a soft mask, one where each pixel carries its own edit strength, and uses a gradient-free sampling process to honor that strength during generation. It works across both image and video diffusion models without needing the expensive pixel-by-pixel labeled data that fine-grained control usually requires.
This matters because most practical edits are not binary either. A designer softening a shadow, smoothing a transition between a real photo and a generated element, or adjusting an edit's intensity across a gradient needs exactly this kind of dial. Zero-shot methods have struggled here, often producing visible seams or over-committing to either the original or the new content. A technique that is gradient-free and memory-efficient, and that generalizes to video, suggests this control problem is solvable without retraining every model from scratch.
The more interesting claim is the video extension. Video diffusion editing has lagged image editing specifically because fine spatial-temporal control is hard to annotate and hard to enforce frame to frame. If SoftPaint's sampler genuinely holds soft-mask strength consistently across frames, that is a bigger deal than the image-editing improvement alone.
It is one paper, not a shipped product, and arXiv preprints have a habit of looking cleaner on paper than in a user's hands.