UniSlider: Perceptually Uniform Sliders for Continuous Image Editing
UniSlider remaps generative edit strength so perceptual change grows linearly, beating prior sliders on a 300-edit benchmark.
UniSlider separates the user-facing slider from a generative model's strength parameter so perceptual distance from the input grows linearly. A lightweight LoRA trained on a few-step editing backbone makes the edit trajectory more monotone while preserving the backbone output at full strength. Because a low-rank adapter cannot fully uniformize strength, an inference-time adaptive remapping closes the remaining gap without extra parameters. On a new 300-edit benchmark of uniformity, monotonicity, fidelity, and identity preservation, UniSlider beats prior methods and is preferred in a user study.
- Slider position is decoupled from raw generative edit strength
- A LoRA on a few-step backbone makes the trajectory more monotone
- Inference-time adaptive remapping finishes perceptual uniformity
- Outperforms prior methods on a 300-edit benchmark and a user study
From the original article
1 min read at arxiv.orgSliders provide an intuitive interface for continuous image editing. In current generative approaches, however, the slider is simply a rescaling of the method's strength parameter, such as an adapter coefficient, a prompt weight, or an interpolation factor. This strength relates poorly to perceptual change.
By David Serrano-Lozano · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.