UniSlider: Perceptually Uniform Sliders for Continuous Image Editing

Forum topic · 小凯 · 2026-10-07

Summary

UniSlider (arXiv: 2610.06831) addresses a key usability problem in AI image editing: sliders in current generative editing methods simply rescale internal strength parameters (adapter coefficients, prompt weights, or interpolation factors), which correlate poorly with perceived visual change. Images can partially revert as the slider moves, show no visible difference over long ranges, then change abruptly. The authors decouple the slider from strength, requiring that perceptual distance from the input image grows linearly with slider value. UniSlider is a lightweight LoRA trained on a few-step editing backbone so that strength approaches an ideal slider. Few-step sampling enables enforcing this objective directly in pixel space without intermediate ground truth, while the backbone output remains untouched at full strength. Since a low-rank adapter alone cannot make strength fully uniform, UniSlider applies an inference-time adaptive remapping of strength; because training makes trajectories monotonic, this remapping closes the remaining gap without extra training or parameters. Evaluated on a new 300-edit benchmark measuring uniformity, monotonicity, edit fidelity, and identity preservation, UniSlider outperforms all prior methods and is preferred in user studies.

Overview

Research area: Image Editing
Authors: David Serrano-Lozano, Duygu Ceylan, Yannick Hold-Geoffroy, Iliyan Georgiev, Javier Vazquez-Corral, Anna Frühstück
Published: 2026-10-05
arXiv: 2610.06831

Key Points

Sliders provide an intuitive interface for continuous image editing, but in current generative methods they are merely a rescaling of a method's strength parameter — adapter coefficients, prompt weights, or interpolation factors.
This strength does not relate well to perceived change: images may partially revert as the slider moves, show no visible difference over long ranges, and change abruptly over short ones.
Remapping strength can fix non-uniform pacing, but only if trajectories are monotonic — something current methods do not guarantee.
The authors therefore decouple the slider from strength, requiring perceptual distance from the input to grow linearly with the slider value.

Method

UniSlider is a lightweight LoRA trained on a few-step editing backbone to make strength approximate an ideal slider.
Few-step sampling allows enforcing this objective directly in pixel space without intermediate ground truth, while the backbone output remains unchanged at full strength.
A low-rank adapter cannot fully uniformize strength, so UniSlider applies an inference-time adaptive remapping of the strength. Because training optimizes for monotonic trajectories, this remapping bridges the remaining gap with no additional training or parameters.

Results

Evaluated on a new benchmark of 300 continuous edits measuring uniformity, monotonicity, edit fidelity, and identity preservation.
UniSlider outperforms all prior methods and is preferred in a user study.