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

arXiv cs.GR details SliderEdit++, a new approach to instruction-based image editing aimed at fixing a familiar UX problem: sliders that feel jumpy, non-monotonic, or overly coupled. The paper frames this as a controllability issue, where moving a slider should produce a stable progression instead of sudden visual shifts.

To measure the problem, the authors introduce SlideBench, a benchmark for continuous image editing with metrics for monotonic progression, perceptual continuity, and inter-slider disentanglement. That matters for anyone building creative tools, because the difference between a useful control and a frustrating one often comes down to whether edits behave consistently across the full range.

SliderEdit++ changes the training objective so it regularizes the intermediate trajectory, not just the endpoints. It combines velocity- and semantic-space trajectory regularization with spectral control, plus a disentanglement objective to reduce interference between independently controlled edits.

The claimed payoff is both quality and responsiveness: on SlideBench, the method reaches state-of-the-art results and improves the overall score by 18.5% versus the strongest baseline. Using a few-step editing backbone also makes it roughly 3x to 20x faster than existing slider-based methods, which could make interactive image-editing tools feel much more immediate in practice.