SliderEdit++: Improved Slider Trajectories in Instruction-Based Image Editing

SliderEdit++. Improved Slider Trajectories in Instruction-Based Image Editing. By Arman Zarei and seven others, from Adobe Research and the University of Maryland. Published on the fourteenth of October. No code repository is linked to it.

In instruction-based image editing, a slider is meant to give the user continuous, fine-grained control over how strongly an individual edit is applied. Slide a little, get a little more of the edit. Existing approaches do not reliably deliver that. Moving along a slider can produce abrupt visual jumps, and the change is not always monotonic, so pushing further can partly undo what came before. When several sliders are controlled together, they can interfere with one another, so turning up one edit drags another along with it.

To study those failures systematically, the authors first built SlideBench, a benchmark for continuous image editing that spans diverse editing categories. It comes with a unified suite of metrics for three properties: monotonic progression, perceptual continuity, and disentanglement between sliders. With that in hand they propose SliderEdit++. Its central move is to regularize the whole intermediate editing trajectory explicitly, instead of supervising only where the trajectory starts and ends.

The trajectory is shaped in two spaces. Regularization is applied in velocity space and in semantic space, and it is paired with spectral control, all aimed at making slider traversal smooth and predictable. Alongside that, a separate disentanglement objective reduces interference between edits that are meant to be controlled independently. The method also runs on a few-step editing backbone.

On SlideBench, SliderEdit++ reaches state-of-the-art performance, improving the overall score by eighteen point five percent relative to the strongest baseline. The few-step backbone is what brings the speed: the method is roughly three to twenty times faster than existing slider-based approaches.

Speed changes what a slider is for. A control that lags is something you set and wait on, while one that answers quickly is something you can sweep back and forth while watching the image. That is the use the authors point to, substantially more responsive interactive editing, and it only pays off because the path between the endpoints is now regularized too, so what you see mid-sweep is meant to be as predictable as where you stop.