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

Beomjin Ahn

cs.CVOct 27, 2026

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

Authors: Beomjin Ahn

Organizations: Samsung Research, Seoul, Korea

Abstract

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.

Figures & tables

Panel of Figure 1 — Figure 1: PREEN restyles an input toward a reference via a predicted $29$ -parameter editor preset (shown at the Vivid strength setting).

Figure 2: Overview of PREEN. Given a content image $X$ and a style reference $Y$ , PREEN predicts a $29$ -parameter editor preset $P_{\mathrm{pred}}$ in a single forward pass. Applying $P_{\mathrm{pred}}$ with our differentiable editor $E$ yields $X^{\prime}_{\mathrm{pred}}$ ; the strength knob $\alpha$ scales the preset’s chroma parameters before editing. PREEN’s output is a compact, human-editable preset rather than a rendered image. Flame icons mark the modules updated during training (CLIP only through low-rank adapters), and the snowflake marks the fixed editor.

Method	Output	Dim	PSNR $\uparrow$	SSIM $\uparrow$	HCorr $\uparrow$	LPIPS $\downarrow$	$\mathbf{\Delta E^{*}_{ab}\downarrow}$	Time (s)
No edit (identity)	–	–	22.67	0.906	0.348	0.127	12.32	–
Reinhard [ 11 ]	Raster	–	20.13	0.835	0.447	0.122	13.40	0.004
AdaIN [ 13 ]	Raster	–	19.06	0.704	0.504	0.314	14.38	0.12
Deep Preset ( $\hat{Y}$ ) [ 6 ]	Raster	–	23.29	0.893	0.411	0.107	10.53	0.19
ModFlows [ 18 ]	Raster	–	19.98	0.853	0.510	0.118	12.77	1.57
NLUT [ 3 ]	LUT	$\mathbb{R}^{3{\times}33^{3}}$	23.43	0.885	0.545	0.129	9.81	N/A †

Table 1: PST50 (paired) metrics. The Output column gives each method’s edit representation, and the Dim column gives its dimensionality. Raster outputs have no fixed dimension because their size scales with the image resolution. Time is the per-image CPU latency (supplementary); for LUT and preset methods it covers generating the LUT or preset, not applying it. Bold marks the best value per metric column. Per-image Wilcoxon significance is reported in the supplementary.

Panel of Figure 3 — Figure 3: Head-to-head on PST50 (first five samples). Columns: content $X$ , reference $Y$ , SA-LUT [ 4 ] , Deep Preset ( $\hat{P}$ ) [ 6 ] , PREEN ( $\alpha\!=\!1$ ), and the expert target.

Setting	PSNR	SSIM	HCorr	LPIPS	$\mathbf{\Delta E^{*}_{ab}}$
Phase 1 only	23.60	0.907	0.383	0.123	10.60
Phase 2 only	24.43	0.918	0.453	0.105	9.12
MLP-only LoRA	24.39	0.915	0.423	0.115	9.89
w/o gradient routing	24.25	0.898	0.473	0.097	9.44
w/o identity samples	22.36	0.833	0.391	0.118	10.87
RGB loss only	24.89	0.922	0.462	0.103	9.59

Table 2: Pipeline ablation of PREEN on PST50. The top block trains a single phase; each remaining row changes one factor of the full pipeline (bottom row).

Teacher	PSNR	SSIM	HCorr	LPIPS	$\mathbf{\Delta E^{*}_{ab}}$
SA-LUT Global	24.74	0.917	0.465	0.094	9.23
SA-LUT (Our setting)	24.63	0.914	0.473	0.094	9.07

Table 3: Teacher ablation: PREEN trained under the published SA-LUT teacher vs. a global variant (per-pixel context map collapsed to a per-image mean). Metrics are PREEN’s.

Backbone	Params	PSNR	SSIM	HCorr	LPIPS	$\mathbf{\Delta E^{*}_{ab}}$
No backbone	6M	23.03	0.883	0.359	0.141	11.36
DINOv3 [ 12 ]	317M	23.04	0.884	0.370	0.141	11.39
SigLIP 2 [ 13 ]	330M	23.25	0.889	0.388	0.128	11.58
InternViT [ 2 ]	318M	23.40	0.906	0.371	0.127	11.22
CLIP [ 10 ]	315M	23.60	0.907	0.383	0.123	10.60

Table 4: Phase 1 backbone ablation. Params = backbone + MLP head. Each backbone uses its native input resolution.

Panel of Figure 4 — Figure 4: Qualitative comparison on three DPST [ 8 ] pairs with distinct content (no paired ground truth). Columns: $X$ , $Y$ , SA-LUT [ 4 ] , Deep Preset ( $\hat{P}$ ) [ 6 ] , PREEN ( $\alpha\!=\!1$ ), and PREEN Vivid ( $\alpha\!=\!5$ ).

Active groups	#	PSNR	SSIM	HCorr	LPIPS	$\mathbf{\Delta E^{*}_{ab}}$
No edit	0	22.67	0.906	0.348	0.127	12.32
+ G1 (WB)	2	21.99	0.898	0.403	0.102	11.71
+ G2 (global tone)	7	24.21	0.913	0.451	0.099	10.08
+ G3 (saturation)	8	24.40	0.919	0.443	0.096	9.60
+ G4 (per-hue HSL)	29	24.63	0.914	0.473	0.094	9.07

Table 5: Incremental preset-group contribution on PST50. The predicted preset is masked at inference to retain only the listed parameter groups, with the rest set to their neutral value.

Panel of Figure 5 — Figure 5: Visual $\alpha$ sweep on the last three PST50 samples. One forward pass, with only the inference-time chroma scalar changing between operating points. Columns: $X$ , $Y$ , $\alpha\!=\!1$ , $\alpha\!=\!3$ , $\alpha\!=\!5$ , $\alpha\!=\!10$ .

Parameter prediction (Output: Preset)
Method	PSNR	SSIM	HCorr	LPIPS	$\mathbf{\Delta E^{*}_{ab}}$
Gemini 3.1 Pro	21.40	0.788	0.343	0.157	12.57
Claude Opus 4.8	22.49	0.857	0.364	0.124	11.41
Qwen3-VL 235B	21.87	0.886	0.352	0.124	12.02
Image-to-image (Output: Raster)
Nano Banana Pro	19.66	0.745	0.461	0.198	15.24

Table 6: Off-the-shelf foundation models on PST50: frontier VLMs predicting the preset and a generative image model, with PREEN for reference.

Panel of Figure 6 — Figure 6: Human pairwise preference on PST50. Each bar splits the judgments between the two labeled methods. The printed percentage is the win rate of the segment it sits in, and the dashed line marks $50\%$ .

Stage	Parameter	Idx	Update rule
1	Temperature	0	$X_{c}\leftarrow m_{c}\,X_{c}$ ( $m_{c}$ : white-balance LUT, Table S2 )
2	Contrast	1	$X\leftarrow\tfrac{1}{2}+\gamma\,(X-\tfrac{1}{2}),\hskip 8.50012pt\gamma=\big((s+400)/400\big)^{2}$
3	Tint	2	$X_{r}\mathrel{+}=\rho\,\phi\cdot(0.0032\,s)\,X_{r},\ \ X_{g}\mathrel{+}=(1{-}\rho)\,\phi\cdot(0.0034\,(-s))\,X_{g},\ \ X_{b}\mathrel{+}=\rho\,\phi\cdot(0.0034\,s)\,X_{b}$
			$\ell=0.2126X_{r}{+}0.7152X_{g}{+}0.0722X_{b},\ \ \phi=(\ell/\sigma)^{2}$ if $\ell{<}\tau$ else $((1{-}\ell)/\sigma)^{2}$
			$\rho=\mathrm{sig}(0.5s),\ \tau{=}\tfrac{128}{255},\ \sigma{=}\tfrac{170}{255}$
4	Saturation	3	$X_{c}\leftarrow\big(WX_{c}+\mathrm{Gain}\,(g_{r}X_{r}{+}g_{g}X_{g}{+}g_{b}X_{b})\big)/1024$

Table S1 : The

29

-parameter editor in fixed application order, with the per-stage update rule (

s\equiv 100\,p

for the stage’s parameter; stage-1 multipliers

m_{c}

in Table S2 ;

h{=}128/255

; constants rounded and hard branches shown for readability). Stages 1–4 act in RGB, 5–8 in YCbCr luma

Y

, 9 in HSL.

knot	$-100$	$-75$	$-50$	$-25$	$0$	$25$	$50$	$75$	$100$
$M_{R}$	1.00	1.00	1.00	1.00	1.00	1.06	1.13	1.20	1.25
$M_{G}$	1.67	1.34	1.17	1.08	1.00	1.06	1.10	1.13	1.16
$M_{B}$	4.71	2.02	1.40	1.16	1.00	1.00	1.00	1.00	1.00

Table S2 : Stage-1 white-balance calibration LUT: per-channel multipliers

M_{c,j}

at the 9 temperature knots (UI units, here mapped to

p_{0}\in[-1,1]

$\alpha$	PSNR $\uparrow$	SSIM $\uparrow$	HCorr $\uparrow$	LPIPS $\downarrow$	$\mathbf{\Delta E^{*}_{ab}}\downarrow$
No edit	22.67	0.906	0.348	0.127	12.32
$0.25$	24.11	0.909	0.412	0.112	10.45
$0.5$	24.40	0.912	0.437	0.102	9.71
$1.0$ (default)	24.63	0.914	0.473	0.094	9.07
$2.0$	24.26	0.913	0.481	0.095	9.36
$3.0$	23.53	0.908	0.445	0.105	10.29

Table S3 : Full

\alpha

sweep on PST50. The first row is the identity (all

29

parameters neutral, the “No edit (identity)” row of Table 1 of the main paper), shown for reference; it is not the

\alpha{=}0

point of the main paper’s Eq. (3), which would zero only the chroma slice while retaining the predicted tone parameters.

Panel of Figure 16 — Figure S1 : Chroma-only versus full-parameter $\alpha$ -scaling on one PST50 sample with a pronounced warm edit. Top: scaling only the chromaticity slice (ours) intensifies color while preserving tone. Bottom: scaling all parameters co-saturates the tonal channels, degrading the photograph as $\alpha$ grows. Columns: content $X$ , reference $Y$ , and $\alpha\!=\!1$ , $3$ , $5$ , $10$ .

vs.	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ HCorr	$\Delta$ LPIPS	$\Delta\Delta E^{*}_{ab}$
No edit	$+1.96$ ( $<$ .001)	$+.009$ (.041)	$+.125$ ( $<$ .001)	$-.034$ ( $<$ .001)	$-3.25$ ( $<$ .001)
Reinhard	$+4.50$ ( $<$ .001)	$+.079$ ( $<$ .001)	$+.026$ (.58)	$-.028$ (.002)	$-4.34$ (.002)
AdaIN	$+5.57$ ( $<$ .001)	$+.210$ ( $<$ .001)	$-.031$ (.19)	$-.221$ ( $<$ .001)	$-5.32$ ( $<$ .001)
Deep Preset ( $\hat{Y}$ )	$+1.34$ (.011)	$+.022$ (.007)	$+.062$ (.12)	$-.014$ (.077)	$-1.46$ (.014)
ModFlows	$+4.64$ ( $<$ .001)	$+.062$ ( $<$ .001)	$-.037$ (.22)	$-.025$ (.041)	$-3.70$ (.021)
NLUT	$+1.20$ (.015)	$+.030$ (.002)	$-.072$ (.028)	$-.035$ ( $<$ .001)	$-0.74$ (.42)

Table S4 : Per-image paired comparison, PREEN (

\alpha\!=\!1

) minus baseline: mean difference with two-sided Wilcoxon

p

in parentheses. Positive

\Delta

PSNR/

\Delta

SSIM/

\Delta

HCorr and negative

\Delta

LPIPS/

\Delta\Delta E^{*}_{ab}

favor PREEN. The last row instead compares the two teacher variants (PREEN trained under each); it is global minus published, so a positive

\Delta E^{*}_{ab}

favors the published teacher.

Method	PSNR	SSIM	HCorr	LPIPS	$\mathbf{\Delta E^{*}_{ab}}$
No edit	22.67	0.906	0.348	0.127	12.32
Fit-to-ref (mean+std) ‡	19.84	0.812	0.455	0.191	13.86
Fit-to-ref (marginals) ‡	19.59	0.793	0.471	0.214	14.13
Damped teacher output ( $\gamma{=}0.5$ , raster) †	25.09	0.922	0.501	0.088	8.71
TTO inversion of teacher †‡	22.67	0.892	0.505	0.120	9.70
TTO inversion of damped teacher output †‡	25.05	0.922	0.471	0.091	8.93

Table S5 : Training-free probes and reference points on PST50, evaluated against the expert target.

\dagger

= requires the SA-LUT teacher at inference;

\ddagger

= per-image test-time optimization. The bottom row is the oracle fit to the target (upper bound on the editor’s subspace).

Panel of Figure 19 — Figure S2 : Training-free fit-to-reference (marginals) over-applies the reference look and distorts the photograph. The three lowest-PSNR fit-to-reference results on PST50. Columns: input $X$ , reference $Y$ , fit-to-reference, PREEN ( $\alpha\!=\!1$ ), and the expert target.

Figure S3 : Predicted presets on PST50 (first five pairs). Per row: inputs ( $X$ , $Y$ ), rendered outputs, and the predicted $29$ -parameter preset for PREEN ( $\alpha\!=\!1$ ) and PREEN Vivid ( $\alpha\!=\!5$ ) as real slider values (top eight: global controls; shaded block: the $21$ per-hue HSL channels). Bar shade tracks magnitude, and the three largest-magnitude parameters (by PREEN $\alpha\!=\!1$ magnitude) are bold; Vivid scales chromaticity only, leaving tone bars unchanged.

Panel of Figure 21 — Figure S4 : PREEN’s five worst per-image $\Delta E^{*}_{ab}$ cases on PST50: target looks outside the editor’s render space (rows 1–4) and a reachable target the regressor misses (last row). Columns: input $X$ , reference $Y$ , oracle fit to the target, PREEN, and the expert target.

Panel of Figure 22 — Figure S5 : Qualitative results on five DPST [ 8 ] pairs (distinct content; no paired ground truth). Columns: $X$ , $Y$ , Deep Preset ( $\hat{P}$ ) [ 6 ] , PREEN ( $\alpha\!=\!1$ ), and PREEN Vivid ( $\alpha\!=\!5$ ).

Panel of Figure 23 — Figure S6 : Extended in-the-wild gallery (1/2): PREEN applied to photographs with selected references. Each group: content $X$ , reference $Y$ , PREEN ( $\alpha\!=\!1$ ), and PREEN Vivid ( $\alpha\!=\!5$ ).

Panel of Figure 24 — Figure S7 : Extended in-the-wild gallery (2/2): $X$ , $Y$ , PREEN ( $\alpha\!=\!1$ ), and PREEN Vivid ( $\alpha\!=\!5$ ).

Phase	Settings
$1$ (inversion)	AdamW, lr $0.2$ , $1{,}000$ steps per pair; CIELAB weight $\lambda_{\mathrm{lab}}{=}20$ ; pre-activation decay $10^{-3}$ ; AdamW weight decay $0.01$ .
$1$ (regression)	AdamW, lr $3{\times}10^{-5}$ ; loss $=$ Huber( $\beta{=}0.1$ ) on $(P_{\mathrm{pred}},\tilde{P})$ $+$ $\ell_{2}$ preset regularizer ( $\lambda{=}10^{-2}$ ); checkpoint at $1.1$ k steps.
$2$ (LoRA)	AdamW, lr $10^{-4}$ , effective batch $64$ ( $16{\times}4$ gradient accumulation), bf16 ; LoRA rank $16$ , alpha $32$ ; image-level loss $w_{\mathrm{pix}}{=}0.2$ , $w_{\mathrm{ab}}{=}5.0$ , $w_{\mathrm{L}}{=}1.0$ ; $\ell_{2}$ preset regularizer $\lambda{=}10^{-2}$ ; identity ratio $0.1$ ; $Y_{\mathrm{aug}}{=}$ random $80$ – $100\%$ crop $+$ horizontal flip; checkpoint at $4.6$ k steps, i.e., ${\approx}294$ k $(X,Y)$ pairs seen ( $4.6\text{k}{\times}64$ ).

Table S6 : Training hyperparameters by phase. Throughout, the backbone is CLIP ViT-L/14 with frozen base weights and the regressor is two parallel residual MLP trunks

[1024,1024,1024,256]

(tone and chroma); training uses the few hundred thousand

(X,Y)

pairs described in the main paper.

Model	Prompt	PSNR $\uparrow$	SSIM $\uparrow$	HCorr $\uparrow$	LPIPS $\downarrow$
Gemini 3.1 Pro	neutral	18.27	0.675	0.211	0.234
	guided	21.40	0.788	0.343	0.157
Claude Opus 4.8	neutral	19.57	0.725	0.261	0.199
	guided	22.49	0.857	0.364	0.124
Qwen3-VL 235B	neutral	15.91	0.713	0.163	0.293
	guided	21.87	0.886	0.352	0.124

Table S7 : Prompt variants on PST50: neutral prompt (a) versus guided (b, reported).

Study	Pairing	Style match (%)	Aesthetic (%)	$n$
1	Default vs. Deep Preset ( $\hat{P}$ )	$61.8\,[53.0,70.6]$	$67.5\,[60.6,74.0]$	$246$
	Default vs. Claude Opus 4.8	$52.0\,[43.2,60.8]$	$52.5\,[44.7,60.6]$	$244$
2	Vivid vs. Deep Preset ( $\hat{P}$ )	$61.6\,[53.3,69.7]$	$47.2\,[38.4,56.0]$	$250$
	Vivid vs. Claude Opus 4.8	$65.7\,[56.3,74.4]$	$40.3\,[33.2,47.6]$	$248$
	Vivid vs. Default	$73.4\,[66.4,79.9]$	$42.5\,[35.6,49.4]$	$252$
3	Default vs. SA-LUT	$28.3\,[22.9,34.0]$	$68.8\,[60.8,76.3]$	$240$

Table S8 : Every pairing of Figure 6 of the main paper: win rate of the first-named PREEN mode, its

95\%

pair-clustered bootstrap interval (

10

k pair resamples), and

n

valid judgments per question.

Panel of Figure 28 — Figure S9 : Full PST50 paired test set (1/2). Each group: content $X$ , reference $Y$ , PREEN ( $\alpha\!=\!1$ ), PREEN Vivid ( $\alpha\!=\!5$ ), and the expert target.

Panel of Figure 29 — Figure S10 : Full PST50 paired test set (2/2). Each group: $X$ , $Y$ , PREEN ( $\alpha\!=\!1$ ), PREEN Vivid ( $\alpha\!=\!5$ ), and the expert target.

Panel of Figure 30 — Figure S11 : Full PST50 unpaired test set (1/2). Each group: $X$ , $Y$ , PREEN ( $\alpha\!=\!1$ ), and PREEN Vivid ( $\alpha\!=\!5$ ).

Panel of Figure 31 — Figure S12 : Full PST50 unpaired test set (2/2). Each group: $X$ , $Y$ , PREEN ( $\alpha\!=\!1$ ), and PREEN Vivid ( $\alpha\!=\!5$ ).

Image Style Transfer

Image style transfer changes an image’s visual appearance to match an artistic, photographic, or other reference while preserving its content.

See all 61 papers

May 22, 2025

CDST: Color Disentangled Style Transfer for Universal Style Reference Customization

We introduce Color Disentangled Style Transfer (CDST), a novel and efficient two-stream style transfer training paradigm which completely isolates color from style and forces the style stream to be color-blinded. With one same model, CDST unlocks universal style transfer capabilities in a tuning-free manner during inference. Especially, the characteristics-preserved style transfer with style and content references is solved in the tuning-free way for the first time. CDST significantly improves the style similarity by multi-feature image embeddings compression and preserves strong editing capability via our new CDST style definition inspired by Diffusion UNet disentanglement law. By conducting thorough qualitative and quantitative experiments and human evaluations, we demonstrate that CDST achieves state-of-the-art results on various style transfer tasks.

Shiwen Zhang, Zhuowei Chen, Lang Chen +1

ByteDance

Disentangled Representation Learning Image Style Transfer

Ask PDF

Jun 1, 2026

Hist2Style: Histogram-Guided Stylization with Bilateral Grids

Photorealistic style transfer aims to match the color and tone of an input image to that of a style target while preserving the content and details of the original scene. Although existing large image models can facilitate these kinds of appearance edits, their high computational demands, potential for hallucinations, and limited user control make them unsuitable for high-resolution, real-time workflows. We introduce Hist2Style, a bilateral-grid formulation for fast, edge-aware stylization that preserves visual fidelity by constraining operations to locally affine transforms in bilateral space. Our model distills a large image editing model into a lightweight network by training on a large supervised corpus generated with language and vision-language models, targeting spatially varying color edits. The network conditions on a histogram-based embedding of the style target to provide an interpretable interface for adjusting the output style by modifying the target color distribution. Overall, Hist2Style maintains content structure by construction, avoids hallucinations, and supports real-time, high-resolution photorealistic stylization with interactive user-controllable color and tone adjustments.

Dekel Galor, Adam Pikielny, Zhoutong Zhang +4

Adobe Nextcam · University of California, Berkeley

Image Style Transfer

Ask PDF

Aug 5, 2026

StyleComposer: Training-Free Multi-Reference Style Composition

The style of a painting is not monolithic: color, texture, and structure may come from different sources. Existing reference-guided methods transfer them as one style signal, leaving each attribute's source and strength outside the user's control. We ask where in a diffusion model one attribute can change while the others hold, and find that no single representation isolates all three. The proposed StyleComposer therefore routes each style attribute through the representation where it separates best and coordinates the routes over denoising time. Without training or inversion, it satisfies three references and the prompt jointly more closely than prior methods, and exposes one strength slider per attribute. Project page: https://lexxsh.github.io/StyleComposer

Sanghyeok Lee, Jihye Kang, Namhyuk Ahn

Inha University

Image Style Transfer Compositional T2I Generation T2I Diffusion Models Diffusion-Based Denoising

Ask PDF

Abstract

Image Style Transfer

Explore similar work

CDST: Color Disentangled Style Transfer for Universal Style Reference Customization

Hist2Style: Histogram-Guided Stylization with Bilateral Grids

StyleComposer: Training-Free Multi-Reference Style Composition