TL;DR
PREEN (Preset Regression via an Exemplar-based Editor Network) performs color style transfer by predicting 29 interpretable parameters for a standard smartphone image editor. By "inverting" a high-quality teacher model into this parameter space, it achieves state-of-the-art results that are fast, editable, and controllable via a real-time "strength knob."
What Problem Does This Paper Address?
Most research in photo-realistic style transfer (PST) focuses on generating new pixels. While visually impressive, these "pixel-generative" methods (like diffusion models) are often slow, stochastic, and produce "black-box" outputs that users can't easily tweak.
On the other hand, "parametric" methods aim to predict slider values for an editor. This is much more practical for mobile apps, but it's hard to train because we don't have large datasets of (Content, Reference, Ground-Truth Preset) triples. Previous attempts relied on synthetic data, which limited the variety of styles the models could learn.
Methods
The authors introduce PREEN, which bridges the gap between high-quality pixel-based style transfer and practical parametric editing.
Key technical moves include:
- Teacher Inversion: Instead of needing ground-truth presets, they use a state-of-the-art LUT-based model (SA-LUT) as a "teacher." They take its output and use gradient-based optimization through a differentiable editor to find the 29 parameters that best recreate that look. These become "pseudo-presets" for training.
- Two-Phase Training:
- Phase 1: Pre-train a regressor (CLIP + MLP) to predict these pseudo-presets.
- Phase 2: Refine the model using an image-level CIELAB loss. By passing the predicted preset through the differentiable editor and comparing the resulting image to the teacher's output, the model learns which parameters are perceptually most important.
- Strength Interpolation: They introduce a simple scalar $\alpha$ that rescales the chromaticity parameters at inference time. This allows users to sweep from a "default" look to a "vivid" one in a single forward pass.
Datasets
- Training: A custom dataset of several hundred thousand pairs. Content images come from OpenImages, and style references are high-aesthetic photos from the AVA database. Pairs are matched semantically using CLIP embeddings.
- Evaluation: The primary benchmark is PST50, a human-curated set of 50 triples (content, reference, and expert-edited target).
Key Insights and Ideas
- The Power of Constraints: A global 29-parameter preset is ~1000x more compact than a typical 3D LUT, yet it outperforms dense LUT methods on fidelity metrics. The constraint acts as a regularizer, preventing the "over-editing" common in free-form methods.
- The "Three-Way Split": A fascinating finding from their human study is that "matching the reference" and "looking good as a photo" are often conflicting goals.
- Vivid settings ($\alpha=5$) win on style matching.
- Default settings ($\alpha=1$) win on aesthetic appeal.
- The $\alpha$ knob allows the user to navigate this trade-off without retraining.
- Hybrid Features: While CLIP provides great semantic context, adding simple handcrafted color statistics (like RGB histograms and log-ratios) significantly boosts performance.
Key Results
- State-of-the-Art Fidelity: PREEN leads baselines on the PST50 benchmark, achieving 24.63 dB PSNR and 0.094 LPIPS, outperforming both previous parametric methods (Deep Preset) and LUT-based ones.
- Mobile-Ready Speed: On a Snapdragon 8 Elite Gen 5, PREEN runs in just 283 ms (CPU-only), making it well-suited for real-time interactive use in mobile photo editors.
- User Preference: In human studies, PREEN's outputs were consistently preferred over the teacher model's full-strength edits, likely because the parametric bottleneck keeps the results more "natural."
Limitations and Future Work
- Global Only: Because it uses a global preset, PREEN can't handle spatially varying edits like vignettes, local "burns," or complex color remappings.
- Semantic Dependency: The model assumes the reference image is semantically compatible with the content. Transferring style between wildly different scenes wasn't the focus.
- Backbone Efficiency: Most of the latency comes from the CLIP ViT-L/14 backbone. The authors suggest that moving to lighter encoders like MobileCLIP could further improve speed.
Summary based on "PREEN: Exemplar-Based Color Style Transfer with Smartphone Editor Presets" by Beomjin Ahn (Samsung Research).
Summary updated