PREEN: Exemplar-Based Color Style Transfer with Smartphone Editor Presets

PREEN. Exemplar-Based Color Style Transfer with Smartphone Editor Presets. By Beomjin Ahn, from Samsung Research. Published on the twenty-seventh of October.

Color style transfer takes a content image and a style reference and gives the content the reference's color. The usual output is a new raster image. PREEN, short for Preset Regression via an Exemplar-based Editor Network, asks for something different. It casts the task as regression onto the twenty-nine parameters of a deterministic image editor. Given the two images, it predicts a preset: a vector of editor slider values. A preset is interpretable, and an editor can consume it directly, which a raster image is not.

The key obstacle is supervision. An arbitrary real reference photo comes with no ground-truth preset, so there is nothing direct to regress onto. The answer is to borrow a teacher. A style-transfer teacher produces target looks, and those looks are inverted into the editor's parameter space, giving pseudo-presets that can serve as training targets.

Training then runs in two steps. A CLIP-conditioned regressor is pretrained on those pseudo-presets. It is then refined with LoRA under an image-level loss measured in the CIELAB color space, so the final tuning is judged on the images the presets produce. Because the output lives in a low-dimensional parameter space, it also exposes a single strength knob at inference time. Turning it traces a continuous trajectory from the default look to a vivid one, in one forward pass, and every intermediate point along the way is still a valid editor preset.

On the human-curated PST50 benchmark, PREEN applies the reference's color while preserving structural fidelity, and it leads the baselines on both pixel and perceptual fidelity metrics. The preset is compact, and it is predicted with sub-second latency on the device. A human study with two axes adds a twist. Raters prefer the vivid end when the question is matching the reference, but prefer the default when judging the result as a standalone photo. That is a genuine trade-off, and the strength knob spans it without any retraining.

The move here is to change what the model outputs. Predicting slider values instead of pixels makes the result something a person can interpret and an editor can apply, keeps the structure of the photo intact, and turns the tension between matching a reference and looking good on its own into a dial rather than a choice made at training time.