PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Beomjin Ahn

arXiv:2610.36896v1cs.CVcs.LG

TL;DR

Reference-based color transfer lacks preset ground truth for arbitrary real examples, while raster outputs do not expose editing parameters. PREEN learns to predict a compact editor preset from a content image and style reference using teacher-derived pseudo-presets. On PST50, it leads baselines on PSNR, SSIM, and LPIPS, while its chroma-strength control traces a continuous preset trajectory.

  • Problem

    Real content–reference pairs lack expert-graded target edits and matching presets, limiting supervision for reference-based preset regression.

    Source: §1 · pp. 1–2

  • Method

    PREEN inverts style-transfer teacher outputs into pseudo-presets, then trains a CLIP-conditioned regressor to predict editor controls from content–reference pairs.

    Source: §1 · p. 2, §3.3 · p. 3

  • Results

    On PST50, PREEN leads every baseline on PSNR, SSIM, and LPIPS, with +1.34 dB PSNR over Deep Preset’s rendered output.

    Source: §4.3 · p. 5

  • Takeaways & Limitations

    A scalar α control rescales preset chroma along a continuous default-to-vivid trajectory of valid, re-editable presets.

    Source: §5 · p. 8

    PST50’s restrained expert targets mean higher fidelity scores do not establish greater expressivity than the spatial LUT teacher.

    Source: §5 · p. 8

Abstract

from arXiv · show

We present PREEN (Preset Regression via an Exemplar-based Editor Network), which casts color style transfer as regression onto the 29 parameters of a deterministic image editor. Given a content image and a style reference, PREEN predicts a preset, a vector of editor slider values that is interpretable and editor-consumable, not a raster image. The key obstacle is supervision, since an arbitrary real reference lacks a preset ground truth. We therefore supervise with a style-transfer teacher, whose target looks we invert into the editor's parameter space as pseudo-presets, then pretrain a CLIP-conditioned regressor and refine it with LoRA under an image-level CIELAB loss. On the human-curated PST50 benchmark, PREEN applies the reference's color while preserving structural fidelity, leading the baselines on pixel- and perceptual-fidelity metrics. The preset is compact and predicted at sub-second on-device latency. The low-dimensional subspace also exposes a single inference-time strength knob that traces a continuous default-to-vivid trajectory in one forward pass, with every intermediate state a valid editor preset. A two-axis human study shows that raters prefer the vivid end for matching the reference but the default as a standalone photo, a trade-off the knob spans without retraining.

1. Introduction

PREEN addresses the lack of reference-conditioned preset supervision by using a style-transfer teacher to train a regressor that outputs editor-consumable parameters. It positions this approach between sampled-preset regression, dense LUTs, and pixel-generative stylization.

1. Introduction

  • Real reference pairs lack expert-graded target edits, so PREEN uses teacher supervision to predict 29-parameter presets from content–reference pairs.Its contributions also include an inference-time strength control and a two-axis human study of the resulting trajectory.Source: §1 · pp. 1–2

2.1. Photo Preset Transfer

  • Deep Preset learns from references re-rendered with sampled presets, confining its learned styles to that sampled set.PREEN instead uses a style-transfer teacher on real reference pairs to supply targets inverted into editor parameters as pseudo-presets.Source: §2.1 · p. 2

2.2. Look-Up-Table Stylization

  • LUT methods map colors through learned grids, but their dense outputs are harder for users to inspect or refine than a few named sliders.Spatially adaptive LUTs include SA-LUT, which also supplies the PST50 evaluation benchmark.Source: §2.2 · p. 2

2.3. Pixel-Generative Stylization

  • Pixel-generative stylization methods emit raster images without exposing user-editable parameters, unlike preset-based outputs.The paper compares experimentally against pixel-generative and LUT methods, as well as Deep Preset [10].Source: §2.3 · p. 2

3. Method

PREEN maps a content image and style reference to a preset for a deterministic 29-control editor. It trains from teacher-generated targets by inverting edits into presets, then regresses those pseudo-labels with a CLIP-conditioned network.

3.1. Problem Formulation

  • The editor maps a content image and 29 bounded controls to an edited image, splitting controls into tone and chromaticity groups.The same differentiable editor supports both pseudo-label inversion and the image-level loss.Source: §3.1 · p. 3

3.2. Teacher Supervision

  • SA-LUT supplies structure-preserving teacher targets dominated by color and tone transforms near the editor’s subspace.The published teacher is used unmodified in both training phases.Source: §3.2 · p. 3

3.3. Phase 1: Pseudo-Label Recovery and Regression

  • Gradient-based optimization through the differentiable editor projects each teacher edit into a pseudo-preset, creating triples for regressor training.The regressor combines frozen CLIP embeddings and color-statistic features, then uses separate tone and chromaticity trunks to predict the controls.Source: §3.3 · p. 3

X Y SA-LUT Deep Preset PREEN Target

Phase 2 adapts the regressor with an image-level loss on editor outputs, and inference-time chroma scaling provides a continuous edit-strength control while leaving tone unchanged.

3.4. Phase 2: Adaptation under Image-Level Loss

  • Image-level refinement addresses perceptual nonuniformity in parameter-space loss, where equal slider shifts can have very different visual effects.The CIELAB terms route luminance and chromaticity gradients to their respective trunks, while an RGB anchor reaches both.Source: §3.4 · p. 4
  • Phase 2 trains LoRA adapters against a CIELAB reconstruction loss between the editor output and the full teacher’s edit of an augmented reference.The objective also includes an ℓ2 preset regularizer; identity pairs use the content image as target.Source: §3.4 · p. 4

3.5. Strength Interpolation via Iab α-Scaling

  • At editing time, α rescales the chromaticity parameters while leaving tone parameters unchanged, providing an inference-time intensity control.α=1 recovers the predicted preset; α>1 intensifies saturation and hue shifts, emphasizing color shift over fidelity to the teacher distribution.Source: §3.5 · p. 4
  • Restricting scaling to chroma preserves the tonal anchor; scaling luminance too can co-saturate tone, clip highlights, and crush shadows.The same regressor serves both default and intensified modes.Source: §3.5 · p. 4

4. Experiments

Experiments train and evaluate PREEN on PST50 using teacher-derived pseudo-presets and standard image-fidelity metrics. PREEN leads the reported baselines on several metrics, while ablations identify gains from the two-phase pipeline.

4.1. Experimental Setup

  • Training pairs combine OpenImages content with semantically compatible AVA references and SA-LUT teacher supervision, yielding a few hundred thousand pairs.Phase 1 recovers pseudo-presets from teacher outputs; the references are sampled from AVA’s top 20% by mean aesthetic score.Source: §4.1 · p. 5

4.2. Evaluation

  • PST50 pairs content, reference, and expert-edited target images; evaluation averages five metrics, including PSNR, SSIM, HCorr, LPIPS, and ∆E∗ ab.Outputs and targets are resized to 256×256 for evaluation.Source: §4.2 · p. 5

4.3. Main Results on PST50

  • PREEN leads every baseline on PST50 PSNR, SSIM, and LPIPS, with +1.34 dB PSNR over Deep Preset’s rendered output and +1.20 dB over next-best NLUT.HCorr rises from 0.348 for no-edit to 0.473; ∆E∗ ab is 9.07, tied with LUT methods within per-image noise.Source: §4.3 · p. 5
  • PREEN’s global preset lacks spatial adaptivity, but its PST50 fidelity is within range of LUT grids over 1000× larger in dimension.The authors attribute the benchmark scores to restrained expert targets rather than greater expressivity.Source: §4.3 · p. 5
  • Qualitative DPST examples show restyling on par with Deep Preset while retaining α control, though paired ground truth and metrics are unavailable.Source: §4.3 · p. 5

4.4.1. Component Ablation

  • The component ablation reports that Phase 1 reaches 23.60 dB PSNR, while Phase 2 adds the largest gains, especially raising HCorr from 0.383 to 0.473.The supplied ablation caption identifies the top block as single-phase training and remaining rows as one-factor changes to the full pipeline.Source: §4.4.1 · p. 5, Table 3 · p. 5

X Y SA-LUT Deep Preset PREEN PREEN Vivid

PREEN’s experiments examine its training choices, preset controls, benchmark comparisons, human preferences, and on-device deployment. The results expose trade-offs between fidelity, reference matching, and standalone aesthetic appeal.

X Y SA-LUT Deep Preset PREEN PREEN Vivid

  • Single-stage training does not surpass the full pipeline, while removing CIELAB terms gains pixel fidelity but cedes every color-facing metric.The single-stage variant reaches comparable ∆E∗ ab (9.12 vs. 9.07) but trails on HCorr (0.453 vs. 0.473) and LPIPS; MLP-only LoRA recovers under half the chromaticity gain (HCorr 0.423).Source: X Y SA-LUT Deep Preset PREEN PREEN Vivid · p. 6

4.4.2. Teacher Choice

  • Collapsing SA-LUT’s context map brings no consistent improvement; among tested backbones, CLIP gives the best ∆E∗ ab.The global teacher variant edges PSNR by ∼0.1 dB and SSIM by 0.003 but lowers HCorr and worsens ∆E∗ ab; removing the backbone degrades every metric, while CLIP and handcrafted features contribute complementarily.Source: §4.4.2 · p. 6, §4.4.3 · p. 6

4.5. Strength Interpolation (α Sweep)

  • Figure 5 sweeps α from 1 to 10 in one forward pass; PSNR peaks at default and degrades monotonically as perceptual saturation increases.Vivid is fixed at α=5 by visual inspection, not benchmark tuning.Source: Figure 5 · p. 7, §4.5 · p. 7

4.4.4. Preset-Group Contribution

  • Tone parameters add the largest single preset-group jump (+2.2 dB over white balance), while per-hue HSL channels close the remaining HCorr gap (0.443 → 0.473).Global saturation improves fidelity but lowers HCorr (0.451 → 0.443).Source: §4.4.4 · p. 7

4.6.1. Direct Parameter Prediction

  • Guided VLM preset predictions reach 21.4–22.5 dB PSNR, 2.1–3.2 dB below PREEN; pixel synthesis matches HCorr but trails in fidelity.The pixel-synthesis method trails by ∼5 dB PSNR, 0.17 SSIM, and over 2× LPIPS, and produces a raster without editable parameters.Source: §4.6.1 · p. 7, §4.6.2 · p. 7

4.7. User Study

  • On style match, Vivid wins preset opponents (pooled 66.9%), but on aesthetic appeal it falls to 43.3% pooled as preference shifts to default and Claude.Against the teacher, Vivid roughly ties on style match (44.6%); both PREEN modes are preferred to the teacher’s full-strength edit on aesthetic appeal.Source: §4.7 · pp. 7–8

4.8. On-Device Deployment

  • The exported INT8 model predicts a preset in 283 ms per pair on a Snapdragon 8 Elite Gen 5, CPU-only, with accuracy preserved.The reported PST50 scores are 24.34/0.915/0.479/0.093/9.02, within 0.3 dB PSNR of the main-paper result and matching its other four metrics.Source: §4.8 · p. 8

5. Discussion and Limitations

The discussion frames PREEN’s compact preset space as both a practical control and a constraint. Its fidelity results reflect calibration to PST50’s restrained edits, while its applicability is bounded by the editor and tested reference setting.

5. Discussion and Limitations

  • The chroma-only α control spans valid, re-editable presets, but the human study makes the preferred operating point a user choice across different goals.Figure 6 encodes pairwise win rates by labeled method, with the dashed line marking 50%.Source: §5 · p. 8, Figure 6 · p. 8
  • PREEN’s higher PST50 fidelity does not imply greater expressivity than the spatial LUT teacher; restrained targets and identity sampling calibrate its edits.The ℓ2 preset regularizer also keeps edits calibrated, and damping the teacher toward the input improves four of five PST50 metrics.Source: §5 · p. 8
  • The 29-parameter editor cannot represent palette-remapping looks or local effects such as grain, vignettes, and local burns.For palette-remapping looks, even the per-image oracle fails.Source: §5 · p. 8
  • Transfer to disjoint references is untested; shrinking the CLIP-dominated model footprint with a lighter encoder is left to future work.Source: §5 · p. 8

6. Conclusion

PREEN recasts exemplar-based color transfer as regression onto a deterministic editor’s 29 parameters. Its PST50 results combine color transfer with structural fidelity and compact on-device prediction, while its strength control spans different user goals.

6. Conclusion

  • PREEN predicts a deterministic editor’s 29 parameters, using an inverted style-transfer teacher to create pseudo-presets for CLIP-conditioned regression and LoRA adaptation.Source: §6 · p. 8
  • On PST50, PREEN transfers the reference’s color shift while preserving structure, with a preset ∼1000× more compact than the teacher’s LUT and sub-second on-device runtime.Source: §6 · p. 8
  • A chromaticity scalar spans default and vivid re-editable presets, whose ends serve different goals: matching the reference and standing alone as a photo.Source: §6 · p. 8

1. Editor Pipeline Details

The supplementary material details the editor, the full α sweep, metric protocol, and per-image statistical tests. It reports where α departs from default performance and qualifies which benchmark advantages are statistically supported.

1. Editor Pipeline Details

  • The 29 editor parameters run in a fixed sequence across RGB, luma, and HSL stages, with differentiable operations allowing Phase 2 loss gradients to reach the predicted preset.The editor parameters act in [−1, 1] and map to specific operators and stages.Source: §1 · p. 11, Table S1 · p. 12

2. Full α Sweep Analysis

  • On PST50, performance stays close to α=1 for α∈[0.5, 2] but falls more clearly by α=3, where PSNR drops 24.63→23.53 and ∆E∗ ab rises 9.07→10.29.At α=5 and 10, oversaturation degrades every metric; at α=5, only 0.3% of per-hue channels rotate by more than 30°.Source: §2 · p. 11, Figure S1 · p. 12

3. Metric Protocol

  • PST50 outputs and targets are resized to 256×256; the protocol specifies PSNR, SSIM, HCorr, LPIPS, and ∆E∗ ab computations and uses a held-out validation split for checkpoint selection.PST50 is evaluation-only, with no training-image duplicates found by the reported perceptual-hash inspection.Source: §3 · p. 11

4. Per-Image Statistical Tests

  • At 0.05, PSNR and SSIM advantages are significant against every baseline; LPIPS is significant against all but Deep Preset (Ŷ) (p=.077).PREEN’s ∆E∗ ab lead over the two LUT methods (9.07 vs. 9.63/9.81) is within per-image noise (p≥.42), while its HCorr gain over no-edit is significant (p<.001).Source: §4 · p. 13

5. Training-Free Probes and Reference Points

Training-free probes show that matching a reference’s color statistics can sacrifice fidelity, while blending the teacher output and PREEN’s restrained presets better track PST50 targets. The oracle fit exposes substantial prediction headroom, but trades perceptual texture for color/tone fidelity.

5. Training-Free Probes and Reference Points

  • Per-pair probes compare 29-parameter color-statistic fitting, damped teacher blends, test-time inversion, and an oracle fit to the expert target.Test-time inversion requires the teacher and about one minute per 256×256 image, versus PREEN’s 0.36 s without the teacher.Source: §5 · p. 13, Table S5 · p. 14
  • Reference-fitting reaches HCorr 0.455/0.471 yet collapses every fidelity metric below the no-edit baseline because the reference is a different scene.The authors characterize this as a distribution-only proxy failure and caution that HCorr must be read alongside fidelity.Source: §5 · p. 13
  • At γ=0.5, the damped teacher blend surpasses most main-paper methods, while no other learned baseline’s γ sweep exceeds PREEN’s 24.63 dB PSNR.Best blends reach NLUT 24.05, ModFlows 24.16, Deep Preset ( Ŷ ) 23.52, and AdaIN 23.51 dB.Source: §5 · p. 13
  • Teacher projection changes PSNR only slightly, while PREEN improves from 22.67 to 24.63 dB and ∆E∗ ab from 9.70 to 9.07 against its undamped supervision projection.PREEN’s presets have median norm ratio 0.24 and median cosine alignment 0.28 with the teacher projection; its restraint objective includes an ℓ2 preset regularizer and identity sampling.Source: §5 · pp. 13–14
  • The oracle reaches 32.56 dB PSNR and HCorr 0.672, but its LPIPS 0.137 is worse than the no-edit baseline’s 0.127.The oracle’s CIELAB-weighted reconstruction objective trades perceptual texture for color/tone fidelity, so it upper-bounds the editor’s color/tone subspace rather than perceptual quality.Source: §5 · p. 14

6. Predicted Preset Breakdown

PREEN’s predicted preset exposes real editor-slider values and a vivid mode that scales chromaticity while preserving tone settings. Qualitative examples also show that the global editor cannot reproduce some target looks, and that the backbone checkpoints were compared under matched settings.

6. Predicted Preset Breakdown

  • The Vivid preset rescales 17 chromaticity parameters while leaving 12 tone parameters unchanged across the illustrated PST50 pairs.The 29 slider values comprise eight global controls and 21 per-hue HSL channels; the displayed bars use real values in [−1, 1].Source: §6 · p. 14

7. Qualitative Gallery

  • Four of PREEN’s five largest PST50 ∆E∗ ab errors involve targets beyond the editor’s render space, while another case is prediction error despite a close oracle fit.In the editor-limited cases, oracle ∆E∗ ab is 8.7–11.0 versus an across-benchmark median of 4.1; the last case has oracle ∆E∗ ab 4.6 and PREEN 22.2. The gallery includes the three lowest-PSNR fit-to-reference examples alongside PREEN and expert targets.Source: §7 · p. 14, Figure S2 · p. 15

8. Backbone Ablation Checkpoints

  • The backbone checkpoints use native input resolution while holding the MLP head, training schedule, and loss identical, so rows vary by frozen encoder and resolution.The listed encoders are CLIP ViT-L/14, DINOv3 ViT-L, SigLIP 2 large, and InternViT-300M.Source: §8 · p. 15

9. Additional Implementation Details

The supplementary implementation details specify the training phases, prompting setup, and user-study protocol. Guided VLM prompts recover the reported levels, whereas neutral prompts produce larger edits and degraded metrics.

9. Additional Implementation Details

  • The implementation uses 123 handcrafted color-statistic features alongside two 768-dimensional CLIP image embeddings, with checkpoints at 1.1k Phase 1 and 4.6k Phase 2 steps.Phase 1 inversion uses AdamW for 1,000 steps per pair, and both explicit and decoupled weight decay nudge recovered presets toward neutral.Source: §9 · p. 15, Table S6 · p. 15
  • Random-base LoRA reaches ∆E∗ ab 9.57 on PST50, versus 9.07 for the full pipeline and 9.12 for Phase 2-only full fine-tuning.The authors describe the low-rank update as underexpressive without the Phase 1 pretrained base.Source: §9 · p. 15

10. VLM Prompt Template

  • Under neutral prompting, all three VLMs make 1.7–3.7× larger mean per-pixel edits than under guided prompting, and every metric degrades.The guided prompt uses generic slider semantics, a magnitude prior, and generic examples rather than PST50- or target-specific values; VLM outputs are non-deterministic and were queried once per image.Source: §10 · pp. 15, 17, Table S7 · p. 17

11. User-Study Protocol

  • The preference studies recruited 50 raters each, targeted five independent ratings per image pair, and showed the reference beside two anonymized candidate edits.The two questions asked which edit better matches the style and which looks more beautiful as a standalone photo; participants were restricted to desktop users.Source: §11 · p. 17

X Y Deep Preset ( ˆP) PREEN PREEN Vivid

The supplementary materials document qualitative transfer examples and the complete PST50 galleries, alongside the pairwise rating protocol and win-rate reporting. The ratings distinguish style match from standalone beauty, with reported majority agreement of 74–81%.

X Y PREEN PREEN Vivid X Y PREEN PREEN Vivid

  • Five DPST pairs and extended in-the-wild galleries show content, reference, PREEN default, and vivid outputs, with no paired ground truth for the DPST examples.The in-the-wild examples span landscape, indoor, and low-light scenes; the authors describe transfer of global color and tone character across diverse scenes.Source: Figure S5 · p. 18, Figure S6 · p. 19, Figure S7 · p. 20

X Y Deep Preset ( ˆP) PREEN PREEN Vivid

  • Across the pairwise studies, majority agreement averages 74–81% across both questions, with Fleiss’ κ of 0.15–0.34.The questions separate style matching from standalone-photo beauty; the study comparisons include PREEN default and vivid against Deep Preset, Claude, and SA-LUT.Source: X Y Deep Preset ( ˆP) PREEN PREEN Vivid · p. 18, Table S8 · p. 18

12. Full PST50 Results

  • The complete PST50 galleries show paired cases with expert targets and unpaired cases without them, each comparing PREEN default and vivid outputs.The paired examples are the 50 triples underlying reported numbers; the 50 unpaired cases provide qualitative coverage only.Source: §12 · p. 18, Figure S9 · p. 22, Figure S10 · p. 23

X Y Deep Preset ( ˆP) PREEN PREEN Vivid

  • The VLM prompt figure shows both images attached and a schema-enforced 29-field JSON response; reported results use the guided prompt variant.The worked example abbreviates its 21 per-hue fields.Source: Figure S8 · p. 21

Related papers

Loading 2610.36896v1…