PREEN: Smartphone Color Preset Regression with Just 29 Parameters

From "Generating Images" to "Predicting Presets": PREEN Redefines Color Style Transfer

The common approach to color style transfer is to directly generate a new image, but such methods are often difficult to reproduce in mobile editors. PREEN (Preset Regression via an Exemplar-based Editor Network) proposes a different path: given a content image and a style reference image, it predicts a set of smartphone editor preset parameters rather than outputting pixels. This set of presets is a 29-dimensional vector of slider values that can be consumed directly by the editor and is interpretable.

The core insight of this work lies in transforming style transfer into a regression problem over the editor's parameter space. Compared with other methods, PREEN's output is not a black-box image that "looks like" the target but a set of parameters that users can manually adjust and re-edit. This is significant in real-world workflows—photographers and ordinary users can fine-tune on top of the preset rather than facing a generated image that cannot be reversed.

Where Does the Supervision Signal Come From: Teacher Model and Pseudo-Presets

The core obstacle to training such a model is the lack of supervision signals: any real reference image has no corresponding "ground-truth" editor preset. PREEN's solution is to introduce a style transfer teacher model and invert its target appearance into the editor parameter space to generate pseudo-presets.

The specific pipeline is divided into two steps: first pretrain a CLIP-conditioned regressor, then fine-tune it with LoRA under an image-level CIELAB loss. This design allows the model to still learn a reasonable parameter mapping in the absence of real preset annotations. The CLIP condition ensures semantic understanding of the style reference, while the CIELAB loss directly constrains the perceptual quality of the color output.

CLIP (Contrastive Language–Image Pre-training) is a vision-language pretrained model proposed by OpenAI that aligns the semantic space of images and text through contrastive learning on 400 million image-text pairs. In PREEN, CLIP is used as the conditional encoder for the style reference: the reference image is processed by the CLIP image encoder to extract semantic features, which serve as the conditional input to the regressor, enabling the model to understand "what style the reference image wants" rather than merely memorizing pixel statistics. Compared with directly using the RGB pixels of the reference image as the condition, CLIP features are more robust to semantic content and can alleviate the interference of specific objects in the reference image on style judgment.

LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique that decomposes weight updates into the product of two low-rank matrices, freezing the original model weights and training only a small number of newly added parameters. PREEN uses LoRA for fine-tuning under an image-level CIELAB loss after pretraining the regressor, with the benefits that fine-tuning is low-cost, the general mapping learned during pretraining is not easily damaged, and the model can adapt to the output distribution of a specific teacher model. For scenarios requiring on-device deployment, LoRA's small parameter count also facilitates model version iteration and multi-style adaptation.

Benchmark Performance: Superior in Both Fidelity and Structure

On the manually curated PST50 benchmark, PREEN applies the color of the reference image while maintaining structural fidelity, leading baseline methods on both pixel-level and perceptual fidelity metrics. Because the output is a compact preset vector, inference achieves sub-second latency on-device, making it suitable for real-time smartphone application scenarios.

Compared with methods that directly generate images, PREEN's advantage lies in the low dimensionality of its output space. The subspace formed by 29 parameters is not only compact but also brings an additional benefit—controllability.

CIELAB is a device-independent color space defined by the International Commission on Illumination (CIE), where L* represents lightness and a* and b* represent the green-red and blue-yellow chromaticity axes. Compared with the RGB space, CIELAB is designed so that Euclidean distance is closer to human perceptual difference (ΔE), and is therefore often used as the perceptual loss space for color transfer. In the fine-tuning stage, PREEN adopts an image-level CIELAB loss to directly constrain the perceptual consistency between the color rendered by the editor from the output preset and the target appearance of the teacher model, rather than per-pixel RGB differences. This helps avoid the problem in RGB space of "numerically close but perceptually off" and makes the loss function better aligned with human judgment of color style.

PST50 is the manually curated benchmark constructed in this paper, containing 50 pairs of content images and style reference images for evaluating the fidelity of color style transfer. The "pixel-level" metrics mentioned in the paper usually refer to measures of reconstruction similarity such as PSNR and SSIM, while the "perceptual fidelity" metrics include perceptual distances based on deep features such as LPIPS. PREEN leads the baselines on both types of metrics, indicating that its preset parameters can not only reproduce the reference colors but also preserve the structural details of the content image.

One Knob for Continuously Adjusting Style Strength

The low-dimensional parameter subspace exposes a practical inference-time strength knob. In a single forward pass, users can adjust along a continuous trajectory from "default" to "vivid," and every intermediate state is a valid editor preset. This means users can find a balance between style matching and natural appearance without retraining the model.

A two-axis human study shows that raters prefer the vivid end when matching the reference image, but prefer the default end when appreciating the result as a standalone photo. This seemingly contradictory preference is precisely covered by the strength knob—users can slide to choose according to the usage scenario (matching style or pursuing good looks) without re-running inference.

The "two-axis human study" refers to a study design containing two independent evaluation dimensions: first, the "match the reference image" axis, which asks raters to judge the similarity of the result to the style reference; second, the "standalone photo appreciation" axis, which asks raters to evaluate the naturalness and aesthetics of the result as a photo without seeing the reference. The results of the two axes diverge—the matching task prefers the vivid end, while the appreciation task prefers the default end—indicating that the quality of style transfer cannot be measured by a single metric, and user intent varies with the usage scenario. This finding supports the necessity of the strength knob design: rather than fixing a compromise point during training, it is better to leave the choice to the user at inference time.

Implications for Mobile Imaging Workflows

PREEN's approach suggests a more pragmatic paradigm for style transfer: rather than pursuing end-to-end image generation, constrain the output to a parameter space that the editor can consume. This not only lowers the deployment barrier (sub-second, on-device) but also preserves the user's control over the result. For smartphone manufacturers and imaging application developers, this "predict the preset" approach is easier to integrate into existing editors and easier for professional users to accept.

Of course, the limitation of this method is that it depends on a specific 29-parameter editor—if the editor's parameter system changes, the model needs to be re-adapted. In addition, the quality of pseudo-presets is affected by the teacher model, and the teacher model's bias may be passed on to the final presets. Nevertheless, PREEN's exploration of interpretability and interactivity offers a direction worth attention for color style transfer.

One-line Summary

PREEN turns color style transfer into predicting 29-dimensional smartphone editor presets, enabling interpretable, adjustable, real-time on-device color grading.

Paragraph Summary

PREEN proposes a color style transfer path different from image generation: given a content image and a style reference image, the model directly regresses the 29-dimensional preset parameters of a smartphone editor, outputting interpretable, re-editable slider values rather than pixels. To address the lack of preset annotations for real reference images, PREEN uses a style transfer teacher model to invert the target appearance into pseudo-presets, first pretraining a CLIP-conditioned regressor, then fine-tuning with LoRA under a CIELAB perceptual loss. On the PST50 benchmark, PREEN leads baselines on pixel-level and perceptual fidelity metrics while maintaining structural fidelity; the compact preset vector enables sub-second on-device inference latency. The low-dimensional parameter space also exposes an inference-time strength knob, allowing continuous adjustment between default and vivid in a single forward pass. Human studies show a preference for the vivid end when matching the reference and for the default end when appreciating independently, and the knob precisely covers this trade-off.