PREEN Recasts Color Style Transfer as 29-Dimensional Preset Regression

Summary:PREEN (Preset Regression via an Exemplar-based Editor Network) reframes exemplar-based color style transfer as regression over the 29 parameters of a deterministic mobile image editor. Rather than synthesizing a raster output, it predicts an interpretable preset that can be applied directly to the source image. Training uses a style-transfer teacher to create pseudo-ground-truth images, inverts those images into editor parameters, pretrains a CLIP-conditioned regressor, and then applies LoRA ref

---

Core Idea: Predict an Editable Preset Instead of a New Image

PREEN (Preset Regression via an Exemplar-based Editor Network) reframes exemplar-based color style transfer as low-dimensional parameter regression. Given a content image and a style reference, the system predicts 29 parameters for a deterministic mobile image editor, covering adjustments such as exposure, contrast, saturation, and hue.

Instead of directly generating styled pixels, PREEN produces an interpretable and immediately usable vector of editor sliders. This has three practical benefits:

• Interpretability: Each dimension corresponds to a user-understandable editing operation.
• Direct execution: The output can be passed to an existing editor without another neural rendering stage.
• Continuous control: Preset parameters can be adjusted to move between natural and vivid treatments.

The method's name accurately captures its central design: the model learns a mapping from image pairs into an editor parameter space rather than a direct pixel-to-pixel transformation.

Building the Supervision Loop

The central training challenge is that an arbitrary style reference has no ground-truth preset. Directly asking a model to infer slider values would inject substantial label noise. PREEN addresses this with a three-stage procedure:

1. Generate pseudo-target images: A style-transfer teacher applies the reference style to the content image.
2. Invert the targets into presets: The generated images are passed through the target editor's parameter inversion process, producing pseudo-ground-truth vectors in the same 29-dimensional space.
3. Train the regressor: A CLIP-conditioned network is pretrained for preset prediction and then refined with LoRA using an image-level CIELAB objective.

The pipeline is:

content + style reference → teacher stylization → editor inversion → pseudo-preset → regressor

LoRA is a natural fit because full fine-tuning may overfit the style-transfer teacher and the inversion procedure. Low-rank adaptation allows the base regressor to learn style-conditioned corrections with fewer trainable parameters. The CIELAB objective directly measures the color discrepancy after applying the predicted preset, preventing the regressor from merely fitting parameter-space targets while producing weak visual results.

Why a 29-Dimensional Space Matters

A conventional image-generation model emits an RGB tensor with millions or tens of millions of values. For mobile editing, this creates three practical problems: geometric drift between the source and generated image, weak correspondence to user-visible controls, and repeated high-cost inference whenever a tonal adjustment changes.

PREEN restricts its output to a compact low-dimensional subspace. The network predicts parameter changes, while the existing editor performs rendering. The resulting system offers:

• Low storage cost: A preset contains a few dozen scalar values instead of a full-resolution image.
• Stable inference: On-device prediction requires one lightweight regression pass, without diffusion sampling or token-by-token decoding.
• Controllability: Users can inspect, override, or undo individual parameters.
• Pipeline reuse: Input resolution and rendering remain under the control of the existing mobile editor.

Low-dimensional parameterization does not eliminate all rendering cost. A sub-second on-device claim usually refers to preset regression. If the complete workflow also includes image decoding, tone mapping, and writing the result back to storage, end-to-end latency must be measured separately on the target handset.

Continuous Strength Control in One Forward Pass

Because the output occupies a low-dimensional preset subspace, PREEN can add an inference-time strength coefficient that interpolates along a default-to-vivid trajectory. Every intermediate value remains a valid editor preset, avoiding the need to train a separate strength classifier or maintain one model per intensity level.

This is more robust than interpolating hidden features because interpolation occurs in an explicit and executable parameter space. The two-axis human study identifies a clear product trade-off:

• The vivid endpoint better matches the reference image's color characteristics.
• The default endpoint is generally preferred when the result is judged as a standalone photograph.

In other words, fidelity to the reference and photographic naturalness are different objectives. The strength control lets application developers move continuously between them without retraining or changing the semantics of the preset.

Interpreting the PST50 Results

PREEN reports leading pixel- and perceptual-fidelity performance on the human-curated PST50 benchmark. Compared with direct image generation, this comparison has a structural advantage: the regressor predicts a small parameter vector, and the same editor renders the final output. That reduces opportunities for geometric drift and helps preserve the content image.

The reported results still require careful interpretation:

• PST50 is small and may not cover diverse devices, cameras, or photographic subjects.
• Rankings depend on the selected baselines, metric implementation, and editor version.
• CIELAB measures color distance but does not fully capture skin-tone realism or high-dynamic-range behavior.
• Perceptual metrics often rely on pretrained feature networks and may reflect dataset-specific bias.

Stronger validation would include cross-device testing, compatibility across editors, blind human evaluation on extreme color references and HDR images, and end-to-end energy and thermal measurements on real phones.

Deployment Recommendations

PREEN is well suited to mobile applications that need to:

• match a reference image's color while preserving composition and subject details;
• expose a continuous control from natural to vivid output;
• map AI predictions into an existing editing pipeline;
• process photos locally for privacy-sensitive use cases.

Integrators should version the 29-dimensional preset alongside the target editor and color space. Extreme references should use parameter clipping or fallback behavior, while telemetry should separately record regression latency, editor rendering time, peak memory, and power consumption.

PREEN's primary contribution is not a more spectacular filter image. It turns style transfer from opaque pixel generation into an auditable, editable, reusable control interface that fits mobile graphics pipelines.

---

Source: Hugging Face Daily Papers (2026-10-27)

— END —

Tags: #PREEN Follow #Color Style Transfer Follow #Preset Regression Follow #CLIP Follow #LoRA Follow