Every generative image tool has a strength slider and every one of them is subtly maddening. You drag it a quarter of the way and nothing happens. You drag it to half and the image transforms completely. Somewhere in the middle it briefly goes back toward the original. And no two edits behave the same way, so you cannot build any intuition.
UniSlider: Perceptually Uniform Sliders for Continuous Image Editing, posted 5 October 2026, diagnoses why and fixes it. The authors are David Serrano-Lozano, Duygu Ceylan, Yannick Hold-Geoffroy, Iliyan Georgiev, Javier Vazquez-Corral and Anna Frühstück — an Adobe-weighted lineup, which is unsurprising given the subject.
The diagnosis
A strength slider rescales a method’s internal parameter — an adapter coefficient, a prompt weight, a guidance scale. That number is a knob in the model’s machinery, and there is no reason for it to be linearly related to how different the result looks.
The paper names three concrete symptoms:
Uneven progression. Some ranges produce no visible change, others produce abrupt transformation. The mapping from parameter to perception is wildly non-uniform.
Non-monotonic behaviour. Nothing enforces that increasing the slider increases the change. The image can partially revert as you move it — which is the one that makes the control feel actively broken, because it violates the single assumption every user has about what a slider is.
Poor strength-to-perception relationship. Small parameter adjustments cause large visual changes, or vice versa, and which is which varies by image and by edit.
The fix
A lightweight LoRA trained on a few-step editing backbone, such that perceptual distance from the input grows linearly with the slider value — with adaptive sampling at inference time.
The reframing is the contribution: treat the slider’s range as a perceptual coordinate rather than as a parameter.
Instead of set the adapter coefficient to 0.5, the slider means produce an image that is halfway, perceptually, between the input and the maximum edit. Which is what a user has always assumed it meant.
“Perceptual distance” here will be something in the LPIPS or DreamSim family — a learned metric that correlates with human judgements of similarity far better than pixel difference does. So the training objective is: at slider value t, the output’s perceptual distance from the input should be t × (distance at maximum).
And monotonicity falls out of it. If distance is forced to grow linearly, it cannot decrease, so the reversion artefact is structurally impossible rather than merely discouraged.
Evaluated on a new benchmark measuring uniformity, monotonicity, edit fidelity and identity preservation, with a user study.
Why this is a more important paper than its subject suggests
It would be easy to file this as a quality-of-life improvement to an Adobe feature. It is better understood as a worked example of a problem that is everywhere in generative tooling.
Almost every control in these tools exposes a model-internal quantity and labels it with a human word. Guidance scale. Denoise strength. LoRA weight. CFG. Temperature. Each is a number the implementation needed, surfaced because it was available, with a name chosen afterwards.
The consequence is that users cannot form intuitions, because the same control behaves differently on different inputs. Which is why so much practical knowledge in this field takes the form of folk rules — denoise around 0.35 for this kind of image — rather than understanding. People are memorising the quirks of an un-calibrated instrument.
We ran into exactly this in our AI upscaling guide three days ago: the single most useful piece of advice we could give was if your generative upscale looks wrong, denoise is too high, with a table of empirical ranges. That advice exists because the control is not perceptually calibrated. A UniSlider-style treatment of denoise strength would make the table unnecessary.
The general principle: a control should be calibrated in the units the user perceives, not the units the implementation computes. That is an old idea in interface design — it is why audio faders are logarithmic and why colour pickers use perceptual spaces like OkLab rather than raw RGB — and generative tools have mostly not done it yet.
What to watch for
Does it generalise across edit types? A slider calibrated for style transfer may not be calibrated for object removal. The paper trains on an editing backbone; whether one LoRA covers the space or you need one per edit class is the practical question.
What does it cost? “Lightweight LoRA” and “adaptive sampling at inference” suggest a modest overhead, but adaptive sampling means the number of steps varies with the slider position, which complicates latency budgeting.
And is identity preservation really preserved? It is one of their four benchmark axes, which is a good sign — forcing perceptual distance to grow linearly could in principle be satisfied by changing the wrong things.