Impression Sliders: On-Demand Sliders from Impression Descriptions for Parametric Design

Abstract

In parametric visual design, users often explore a parameter space, guided by a desired impression to convey to viewers—for example, seeking to make a photo look more “cinematic.” However, manipulating the parameters in a direction that strengthens the intended impression is particularly difficult when the user does not understand the parameter space well. We present Impression Sliders, a system that turns a natural-language description of a desired impression into a slider for strengthening that impression. With a vision-language model (VLM) as the evaluator, the system uses preferential Bayesian optimization to estimate, in roughly 45 seconds, a design that strongly conveys the described impression. The resulting slider then lets users move the current design toward this estimate. We implemented the system in three domains, and a study with 19 raters indicated that the estimated designs conveyed the described impressions to viewers. An exploratory user study suggested that Impression Sliders help users adjust designs toward target impressions, particularly when the parameter space is difficult to interpret.