ColorFM: An Optimization-to-Learning Framework for Color Transfer via Flow Matching

Du Chen; Jian Yang; Kai Zhang; Xiaoming Li; Yuhang He

REVIEW 3 major objections 4 minor 47 references

ColorFM: An Optimization-to-Learning Framework for Color Transfer via Flow Matching

T0 review · 3 major / 4 minor · reviewed 2026-07-09 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read One-step color transfer distills slow optimization into real-time inference

desk verdict Solid optimization-to-learning framework for color transfer; the hierarchical color coupling strategy is the real novelty, but the evaluation is thin enough to weaken the central SOTA claim. read the letter →

arxiv 2607.07119 v1 pith:QAVDTH62 submitted 2026-07-08 cs.CV

Yuhang He , Kai Zhang , Xiaoming Li , Du Chen , Jian Yang This is my paper

classification cs.CV

keywords colorsemantictransferflowcolorfmcolorfm-loptimizationcolorfm-o

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Color transfer — recoloring a photo to match the palette of another image while keeping its structure intact — has long been split between two camps: slow per-image optimization that is precise but impractical, and fast feed-forward neural networks that are efficient but produce artifacts, color banding, or semantic mismatches. This paper proposes a two-stage framework called ColorFM that bridges the divide. The key idea is to recast color transfer as a Flow Matching problem: instead of applying a fixed color map, pixels are transported along velocity fields in RGB space, moving the content image's color distribution toward the style image's distribution through a continuous trajectory. The first stage, ColorFM-O, runs instance-specific optimization that pairs pixels across semantic regions using a hierarchical octree-based coupling strategy, producing nearly straight transport paths. This optimization generates high-quality pseudo-ground-truth pairs. The second stage, ColorFM-L, is a feed-forward network trained on those pairs. It predicts the parameters of a lightweight pixel-wise velocity field and performs color transfer in a single bidirectional step — forward to an implicit intermediate state, then backward to the target style — bypassing iterative ODE integration entirely. The critical structural enabler is that the optimization stage produces trajectories with a path-length ratio of approximately 1.009, meaning they are almost perfectly straight. This near-linearity is what licenses the learning stage to collapse the entire transport into one Euler step. The paper reports that the learned model achieves real-time inference (0.016 seconds) while outperforming prior methods on the trade-off between style fidelity and content preservation, with the lowest artifact metric among all compared approaches.

What carries the argument

The framework rests on three load-bearing components. First, explicit semantic alignment: pixels from matching semantic regions (sky-to-sky, skin-to-skin) are paired for distribution transport, but instead of processing each region independently (which causes seam artifacts), all pairs are fed into a single unified velocity field that is optimized jointly. Second, hierarchical color coupling (HCC): a recursive algorithm that partitions source and target color distributions into octants of RGB space, aligning their relative positions at each level before pairing points at leaf nodes. This preserves locality in color space, yielding nearly straight flow paths. Third, bidirectional linearized:Θ

What would settle it

Find a content-style pair where the optimization-stage trajectory has a path-length ratio significantly above 1.0 (say, 1.2 or higher), then show that the one-step bidirectional transport in ColorFM-L produces visibly incorrect or banded color output compared to the full multi-step ODE integration of ColorFM-O. If such pairs are common rather than rare, the core premise — that hierarchical coupling reliably yields straight paths that can be collapsed to a single step — fails.

Watch

Extended reading notes

Core claim

The central mechanism is the discovery that carefully constructed pixel pairings — specifically, recursive octant-based hierarchical coupling of color distributions within semantic regions — produce flow trajectories in RGB space that are quasi-linear (path-length ratio ~1.009). This geometric regularity means the entire color transport can be approximated by a single bidirectional Euler step: push the content image forward to an implicitly learned intermediate color state, then pull it backward toward the style distribution. The feed-forward model predicts the velocity-field parameters that define this two-step transport, inheriting the precision of per-instance optimization without its per

Load-bearing premise

The entire learning pipeline depends on the quasi-linearity of the optimization-stage trajectories. If this near-straightness breaks down for complex or out-of-distribution content-style pairs — which the paper itself acknowledges as a risk — the single-step Euler approximation in the feed-forward model will produce inaccurate color mappings, since there is no iterative correction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick read on ColorFM (2607.07119). The headline contribution is an optimization-to-learning pipeline: ColorFM-O does instance-specific flow matching optimization to produce pseudo-GT pairs, then ColorFM-L learns a fast feed-forward model from that data. The part that's actually new and useful is the hierarchical color coupling (HCC) strategy — recursively partitioning source and target color distributions into octants to build geometrically coherent pairings. It's a clean alternative to mini-batch OT (which loses global statistics) and Rectified Flow reflow (which is expensive), and the ablation in Table 2 shows it meaningfully reduces the Lipschitz constant. The implicit state modeling for bidirectional one-step transport is also a reasonable architectural choice, and the ablation in Table 3 supports it. Code is on GitHub, which helps. The method is well-engineered and the writing is clear. The soft spots are real but not fatal. The test set is 40 images permuted into 1,560 pairs — each image appears in ~78 pairs, so these are not independent samples. No confidence intervals or significance testing anywhere. The margin between ColorFM-L (0.320) and ColorFM-O (0.338) on Distance-to-Ideal is small enough that I wouldn't trust it without error bars. The Style Similarity metric uses a discriminator borrowed from Neural Preset, which could systematically favor outputs resembling its training distribution — and since ColorFM-L is trained on ColorFM-O outputs, there's a subtle feedback loop there that nobody acknowledges. The quasi-linearity assumption (path length ratio 1.009) is validated only on in-distribution data, and the authors admit in Section 6 that ColorFM-L may struggle with extreme styles or OOD layouts. That's honest but also means the one-step Euler approximation in Eq. 4 is untested where it matters most. The claim that ColorFM-L outperforms ColorFM-O is plausible — the inductive bias argument is reasonable — but it could also be an artifact of ColorFM-O using fixed optimization settings across all test pairs. Overall: the HCC strategy and the distillation pipeline are genuine contributions worth reading for anyone in color transfer or flow matching. The evaluation needs strengthening before the SOTA claim is airtight. I'd send it to a serious referee who can pressure-test the metrics and ask for a larger, more independent test set.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes ColorFM, a two-stage framework for color transfer that bridges optimization-based and learning-based paradigms via Flow Matching (FM). The first stage, ColorFM-O, performs instance-specific optimization of a velocity field using hierarchical color coupling (HCC) and semantic priors to generate pseudo-ground truth. The second stage, ColorFM-L, is a feed-forward model trained on this pseudo-supervision that predicts flow parameters for a bidirectional, one-step linearized transport. The core claim is that ColorFM-L achieves state-of-the-art performance in visual quality and structural fidelity while enabling real-time inference. The approach is technically coherent, and the optimization-to-learning pipeline is a reasonable way to transfer the fidelity of iterative methods to a fast inference model.

Significance. Color transfer is an established problem, and the formulation of color distribution transport via Flow Matching is a meaningful contribution to the literature. The hierarchical color coupling (HCC) strategy (Algorithm 1) is a practical heuristic for constructing geometrically coherent couplings without the cubic cost of exact Optimal Transport. The bidirectional linearized transport in ColorFM-L (Eq. 4) is an efficient approximation enabled by the quasi-linearity of the HCC-induced trajectories. The authors provide code and a large-scale dataset, which adds value. The core claim of combining optimization accuracy with feed-forward speed is well-motivated and supported by the efficiency gains reported in Table 1 (0.016s inference).

major comments (3)

§4.1 and Table 1: The evaluation is based on a test set of 40 images exhaustively permuted into 1,560 pairs. Because each image appears in approximately 78 pairs, these pairs are not statistically independent. Treating 1,560 as the sample size without accounting for this clustered structure inflates the apparent statistical power. The margin between ColorFM-L (0.320) and ColorFM-O (0.338) in 'Distance to Ideal' is small. The authors should either report standard errors or confidence intervals adjusted for the clustered design, or evaluate on a larger, independent test set to confirm that the reported differences are statistically meaningful.
§4.2 and Table 1: The Style Similarity metric relies on a discriminator pretrained by Neural Preset [14]. Since ColorFM-L is trained on pseudo-ground truth generated by ColorFM-O, and the discriminator may systematically prefer outputs resembling its own training distribution, there is a risk of evaluation bias. The authors should either validate this metric against human judgments or include an alternative, standard metric (e.g., FID or feature statistics matching using a standard backbone like VGG or CLIP) to corroborate the style transfer quality independently.
§3.4 and §4.3: The one-step bidirectional Euler approximation in ColorFM-L (Eq. 4) relies on the quasi-linearity of ColorFM-O trajectories, supported by a path length ratio of 1.009 (§4.3). However, this ratio is computed on ColorFM-O outputs, which are the training distribution for ColorFM-L. The paper acknowledges in §6 that ColorFM-L may face generalization challenges for out-of-distribution styles. The manuscript would benefit from an explicit evaluation of trajectory linearity on held-out or extreme style pairs to validate that the one-step approximation holds for the cases where ColorFM-L is applied without ColorFM-O's direct supervision.

minor comments (4)

§3.3, Algorithm 1: The notation 'O(v)' for the octant index is introduced but could be clarified for readers unfamiliar with octant partitioning. A brief inline definition would improve readability.
Table 2: The 'Time (s)' column for Mini-Batch OT (51.48s) is significantly higher than other methods. It would be useful to clarify whether this includes the coupling construction time or only the optimization, to ensure a fair comparison.
Fig. 2: The diagram is dense. The distinction between the 'Param Generator' and the 'pixel-wise MLP' in ColorFM-L could be labeled more explicitly to clarify that the generator predicts weights for the MLP rather than processing pixels directly.
§4.1: The training dataset composition (237,408 triplets) is described, but the diversity of the style images is not detailed. Given the reliance on semantic categorization, a brief note on the style distribution would help contextualize the generalization claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the derivation chain is self-contained.

full rationale

The paper's derivation chain is straightforward and non-circular. ColorFM-O reformulates color transfer as Flow Matching (citing external works [2,21,23]) and introduces a novel Hierarchical Color Coupling strategy (Algorithm 1) to construct source-target pairs, then optimizes a velocity field via the standard FM objective (Eq. 2). The resulting outputs serve as pseudo-ground truth for training ColorFM-L. This is explicitly acknowledged self-supervision — the core design of the optimization-to-learning framework — not hidden circularity. ColorFM-L is then evaluated on a separate test set (40 images, no overlap with training) using external metrics: Style Similarity (discriminator from Neural Preset [14], an external work), Content Similarity (SSIM on edge maps), and Lipschitz Constant. The quasi-linearity assumption (path length ratio 1.009) that justifies the one-step Euler approximation in Eq. 4 is empirically verified, not imported via self-citation. No self-citations are load-bearing for the central claims. No step in the derivation reduces to its inputs by construction. The concern that the Style Similarity discriminator might prefer ColorFM-O-like outputs is a metric-bias concern (correctness risk), not a circularity in the derivation chain.

Assumptions & free parameters 9 free parameters · 5 assumptions · 2 invented entities

The paper introduces 9 free parameters, most chosen by standard practice or small-scale ablation. The HCC strategy and implicit state are the main invented entities, with HCC having independent empirical support. The core axioms are domain assumptions about the applicability of flow matching to color transfer and the sufficiency of locality preservation, which are reasonable but not theoretically proven.

free parameters (9)

Dmax (max tree depth for HCC) = 3
Selected by ablation in Table 2; controls the recursion depth of hierarchical color coupling. Dmax=3 gives the best balance of style similarity and Lipschitz constant.
lambda (LPIPS loss weight) = 0.1
Stated in Eq. 5 as balancing pixel-wise accuracy with perceptual fidelity. Chosen by convention without systematic justification.
ODE integration steps (ColorFM-O inference) = 5
Selected based on saturation analysis in Fig. 6; midpoint method with 5 steps.
ColorFM-O optimization steps = 700
Fixed iteration count for all image pairs; acknowledged as a potential limitation in §5.2.
ColorFM-O learning rate = 5e-4
Adam optimizer learning rate for velocity field optimization; standard choice without systematic tuning.
ColorFM-L learning rate = 1e-5
Initial learning rate with cosine annealing; standard schedule.
Batch size (ColorFM-O) = 4096
Number of sampled pixels per optimization step.
Batch size (ColorFM-L) = 48
Training batch size for the feed-forward model.
Input resolution (ColorFM-L) = 256x256
Fixed resolution for feature extraction in ColorFM-L.

assumptions (5)

domain assumption Color transfer can be formulated as transport of pixel distributions in RGB color space via Flow Matching (§3.2).
The paper assumes that the continuous flow framework applies to discrete pixel color distributions. This is a modeling choice that treats pixels as i.i.d. samples from a distribution, ignoring spatial correlations in the transport formulation.
ad hoc to paper Preserving relative locality in color space suffices to eliminate banding and ensure flow coherence (§3.3, HCC insight).
The hierarchical color coupling strategy is based on this insight. While empirically validated in Table 2, the theoretical justification for why octant-based partitioning preserves sufficient locality is not provided.
ad hoc to paper The quasi-linearity of ColorFM-O trajectories holds sufficiently for one-step Euler approximation in ColorFM-L (§3.4).
The path length ratio of 1.009 is computed over 100 sampling steps on test data, but the generalization to unseen distributions is assumed. The paper acknowledges this may fail for extreme styles in §6.
domain assumption Semantic segmentation masks from SegFormer-B5 on ADE20K provide adequate semantic correspondence for color transfer (§3.3).
The explicit semantic alignment relies on pre-trained segmentation. The paper shows robustness to mask degradation (Fig. 7) but assumes the base segmentation is reasonable.
domain assumption The Style Similarity discriminator from Neural Preset [14] is a valid evaluator for color transfer quality (§4.2).
This learned discriminator is used as a primary metric. Its validity as an unbiased evaluator is assumed but not independently verified.

invented entities (2)

Implicit intermediate state z
purpose: Serves as a bridge between content and style distributions in ColorFM-L's bidirectional transport (Eq. 4).
The implicit state is not a fixed prior but is learned through the parameter prediction network. Fig. 9 visualizes it as an averaged intermediate style, but its existence as an 'optimal' bridge is hypothesized, not proven. The paper states: 'we hypothesize the existence of an optimal intermediate state.'
Hierarchical Color Coupling (HCC) plan independent evidence
purpose: Constructs geometrically coherent source-target pixel pairings for velocity field training in ColorFM-O (Algorithm 1).
The HCC strategy is validated through ablation (Table 2) showing improved Lipschitz constant and style similarity compared to random pairing and mini-batch OT. It produces a falsifiable improvement on standard metrics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ColorFM: An Optimization-to-Learning Framework for Color Transfer via Flow Matching." pith.science (2026). https://pith.science/paper/QAVDTH62

@misc{pith2026260707119,
  author       = {Pith},
  title        = {Pith review of: ColorFM: An Optimization-to-Learning Framework for Color Transfer via Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QAVDTH62}},
  note         = {Machine review of arXiv:2607.07119}
}

read the original abstract

Color transfer aims to align the color distribution of a source image with that of a reference image while preserving structural and semantic consistency. However, existing methods often suffer from inaccurate global mapping, semantic misalignment, and visual artifacts. To address these issues, we propose ColorFM, an optimization-to-learning framework. ColorFM connects online optimization to offline inference by reformulating color transfer as the transport of pixel distributions along velocity fields via Flow Matching. Specifically, we introduce ColorFM-O, an instance-specific optimization scheme that fits the velocity field through hierarchical color coupling guided by semantic priors. By numerically integrating the induced flow trajectories, ColorFM-O produces precise and semantically consistent color transfer results, while generating high-quality paired data as pseudo-supervision. Building upon this, we design ColorFM-L, an efficient feed-forward model trained on the generated pairs. Through implicit state modeling, ColorFM-L extracts deep semantic features to predict flow parameters for bidirectional linearized transport, ensuring accurate color transfer. Extensive experiments demonstrate that ColorFM-L outperforms state-of-the-art methods in visual quality, structural fidelity, and semantic consistency, successfully combining the accuracy of optimization with the speed of feed-forward inference.

Figures

Figures reproduced from arXiv: 2607.07119 by the authors.

**Figure 1.** Color Transfer Results of ColorFM-L. ColorFM-L enables precise color transfer across diverse styles while preserving structural fidelity and semantic consistency. methods aim to transfer its color style to the content image while maintaining photorealism. Broadly, existing methods can be categorized into two paradigms: online optimization and offline inference. Optimization-based approaches [4,17,25,28–30] are oft… view at source ↗

**Figure 2.** Overview of the ColorFM Framework. Our framework consists of two components: (a) ColorFM-O first matches pixel distributions across semantic regions using Mc and Ms, and then applies a hierarchical color coupling strategy to each pair. This process constructs the training dataset D to learn the velocity field vθ, enabling precise color transfer via ODE integration. (b) ColorFM-L extracts features from resized Ic and… view at source ↗

**Figure 3.** Qualitative Comparison. Our method exhibits distinct advantages in maintaining global color consistency, ensuring semantic-aware color alignment, and preserving image structure without artifacts. ensuring strictly no overlap with the training set. This dataset covers diverse semantic scenes, enabling a comprehensive evaluation of model performance. Network Configuration. For ColorFM-O, the velocity field is parame… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Visual Results of Distance to Ideal. ColorFM achieves the optimal trade-off between content preservation and style alignment. Content Style w/o semantic w/ semantic [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 6.** Figure 6: Impact of Inference Steps. Metrics saturate at T ≥ 5, suggesting quasi-linear trajectories. Content Style ColorFM-O [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 8.** Figure 8: Comparison Between ColorFM-O and ColorFM-L. While comparable in simple cases (top), ColorFM-L outperforms ColorFM-O in complex scenes (bottom). IA ZA IA→B IB ZB IB→A [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 10.** Figure 10: Results of Video Color Transfer. ColorFM-L ensures temporally consistent and flicker-free stylization across the sequence. decreases only slightly from 0.745 to 0.731/0.722 under 20%/40% semantic misclassification, while a more noticeable drop to 0.695 appears only …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages

[14]

In: CVPR

Ke, Z., Liu, Y., Zhu, L., Zhao, N., Lau, R.W.: Neural preset for color style transfer. In: CVPR. pp. 14173–14182 (2023) 2, 3, 7, 9, 10, 13

work page 2023
[1]

In: CVPRW

Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image super- resolution: Dataset and study. In: CVPRW. pp. 126–135 (2017) 8

work page 2017
[2]

In: ICLR (2023) 2, 4

Albergo, M.S., Vanden-Eijnden, E.: Building normalizing flows with stochastic interpolants. In: ICLR (2023) 2, 4

work page 2023
[3]

In: AAAI

An, J., Xiong, H., Huan, J., Luo, J.: Ultrafast photorealistic style transfer via neural architecture search. In: AAAI. pp. 10443–10450 (2020) 2, 3

work page 2020
[4]

arXiv (2023) 2, 3, 9, 10

Chen, Y., Yang, H., Yang, Y., Liu, Y., Wang, W., Wen, X., Xie, C.: NLUT: Neural- based 3D lookup tables for video photorealistic style transfer. arXiv (2023) 2, 3, 9, 10

work page 2023
[5]

In: WACV

Chiu, T.Y., Gurari, D.: PhotoWCT2: Compact autoencoder for photorealistic style transfer resulting from blockwise training and skip connections of high-frequency residuals. In: WACV. pp. 2868–2877 (2022) 2, 3, 6, 9, 10

work page 2022
[6]

In: CVPR

Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: StyTr2: Image style transfer with transformers. In: CVPR. pp. 11326–11336 (2022) 3

work page 2022
[7]

In: ICLR (2021) 9

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 9

work page 2021

Show all 47 references

[8]

In: AISTATS

Fatras, K., Zine, Y., Flamary, R., Gribonval, R., Courty, N.: Learning with mini- batch wasserstein : asymptotic and gradient properties. In: AISTATS. vol. 108, pp. 2131–2141 (2020) 4 16 Y. He et al

2020
[9]

In: CVPR

Gatys, L.A., Ecker, A.S., Bethge, M.: Image style transfer using convolutional neural networks. In: CVPR. pp. 2414–2423 (2016) 3

2016
[10]

In: ICCV

Gong, Z., Wu, Z., Tao, Q., Li, Q., Loy, C.C.: SA-LUT: Spatial adaptive 4d look-up table for photorealistic style transfer. In: ICCV. pp. 18294–18303 (2025) 2, 3, 4, 9, 10

2025
[11]

In: WACV

Ho, M.M., Zhou, J.: Deep preset: Blending and retouching photos with color style transfer. In: WACV. pp. 2113–2121 (2021) 2, 3, 9, 10

2021
[12]

In: ICCV

Hong, K., Jeon, S., Yang, H., Fu, J., Byun, H.: Domain-aware universal style trans- fer. In: ICCV. pp. 14609–14617 (2021) 3

2021
[13]

In: ICCV

Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: ICCV. pp. 1501–1510 (2017) 3

2017
[15]

In: ICLR (2015) 9

Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) 9

2015
[16]

In: AAAI

Larchenko, M., Lobashev, A., Guskov, D., Palyulin, V.V.: Color transfer with mod- ulated flows. In: AAAI. pp. 4464–4472 (2025) 2, 3, 4, 7, 9, 10

2025
[17]

In: WACV

Li, M., Wang, G., Zhang, X., Liao, Q., Xiao, C.: D-LUT: Photorealistic style trans- fer via diffusion process. In: WACV. pp. 9188–9196 (2025) 2, 3, 9, 10

2025
[18]

NeurIPS30(2017) 3

Li, Y., Fang, C., Yang, J., Wang, Z., Lu, X., Yang, M.H.: Universal style transfer via feature transforms. NeurIPS30(2017) 3

2017
[19]

In: ECCV

Li, Y., Liu, M.Y., Li, X., Yang, M.H., Kautz, J.: A closed-form solution to photo- realistic image stylization. In: ECCV. pp. 453–468 (2018) 2, 3

2018
[20]

In: AAAI

Lin, T., Lin, H., Li, F., He, D., Wu, W., Wang, M., Li, X., Liu, Y.: AdaCM: adaptive colormlp for real-time universal photo-realistic style transfer. In: AAAI. pp. 1613–1621 (2023) 2, 3

2023
[21]

In: ICLR (2023) 2, 4, 6

Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023) 2, 4, 6

2023
[22]

IEEE TIP32, 4742–4756 (2023) 2

Liu, C., Yang, H., Fu, J., Qian, X.: 4D LUT: learnable context-aware 4d lookup table for image enhancement. IEEE TIP32, 4742–4756 (2023) 2

2023
[23]

In: ICLR (2023) 2, 4, 6

Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: ICLR (2023) 2, 4, 6

2023
[24]

In: ICLR (2024) 4

Liu, X., Zhang, X., Ma, J., Peng, J., et al.: Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In: ICLR (2024) 4

2024
[25]

In: CVPR

Luan, F., Paris, S., Shechtman, E., Bala, K.: Deep photo style transfer. In: CVPR. pp. 4990–4998 (2017) 2, 3, 6

2017
[26]

TMLR (2024) 9

Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jégou, H., Mairal, J., ...

2024
[27]

In: CVPR

Park, D.Y., Lee, K.H.: Arbitrary style transfer with style-attentional networks. In: CVPR. pp. 5880–5888 (2019)

2019
[28]

In: CVMP

Pitié, F., Kokaram, A.: The linear monge-kantorovitch linear colour mapping for example-based colour transfer. In: CVMP. pp. 1–9 (2007) 2, 3

2007
[29]

In: ICCV

Pitie, F., Kokaram, A.C., Dahyot, R.: N-dimensional probability density function transfer and its application to color transfer. In: ICCV. pp. 1434–1439 (2005) 2, 3

2005
[30]

CVIU107(1-2), 123–137 (2007) 2, 3 ColorFM for Color Transfer 17

Pitié, F., Kokaram, A.C., Dahyot, R.: Automated colour grading using colour dis- tribution transfer. CVIU107(1-2), 123–137 (2007) 2, 3 ColorFM for Color Transfer 17

2007
[31]

In: ICML

Pooladian,A.A.,Ben-Hamu,H.,Domingo-Enrich,C.,Amos,B.,Lipman,Y.,Chen, R.T.: Multisample flow matching: straightening flows with minibatch couplings. In: ICML. pp. 28100–28127 (2023) 4, 6

2023
[32]

In: ICML

Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Bengio, Y., Courville, A.: On the spectral bias of neural networks. In: ICML. pp. 5301–5310 (2019) 11

2019
[33]

In: ICLR Workshop (2018) 9

Ramachandran, P., Zoph, B., Le, Q.V.: Searching for activation functions. In: ICLR Workshop (2018) 9

2018
[34]

NeurIPS32(2019) 3

Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. NeurIPS32(2019) 3

2019
[35]

IEEE Access10, 68281–68290 (2022) 10

Soria, X., Pomboza-Junez, G., Sappa, A.D.: LDC: Lightweight dense cnn for edge detection. IEEE Access10, 68281–68290 (2022) 10

2022
[36]

TMLR (2024) 4, 6

Tong, A., Fatras, K., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Wolf, G.,Bengio,Y.:Improvingandgeneralizingflow-basedgenerativemodelswithmini- batch optimal transport. TMLR (2024) 4, 6

2024
[37]

In: CVPR

Ulyanov, D., Vedaldi, A., Lempitsky, V.: Deep image prior. In: CVPR. pp. 9446– 9454 (2018) 11

2018
[38]

arXiv preprint arXiv:2404.02733 (2024) 3

Wang, H., Spinelli, M., Wang, Q., Bai, X., Qin, Z., Chen, A.: InstantStyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733 (2024) 3

2024 arXiv
[39]

IEEE TIP13(4), 600–612 (2004) 10

Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP13(4), 600–612 (2004) 10

2004
[40]

In: CVPR

Wen, L., Gao, C., Zou, C.: CAP-VSTNet: Content affinity preserved versatile style transfer. In: CVPR. pp. 18300–18309 (2023) 2, 3, 6, 9, 10

2023
[41]

In: ECCV

Xia, X., Zhang, M., Xue, T., Sun, Z., Fang, H., Kulis, B., Chen, J.: Joint bilateral learning for real-time universal photorealistic style transfer. In: ECCV. pp. 327–342 (2020) 2, 3

2020
[42]

NeurIPS 34(2021) 9

Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. NeurIPS 34(2021) 9

2021
[43]

In: CVPR

Ye, Z., Huang, H., Wang, X., Wan, P., Zhang, D., Luo, W.: Stylemaster: Styl- ize your video with artistic generation and translation. In: CVPR. pp. 2630–2640 (2025) 3

2025
[44]

In: ICCV

Yoo, J., Uh, Y., Chun, S., Kang, B., Ha, J.W.: Photorealistic style transfer via wavelet transforms. In: ICCV. pp. 9036–9045 (2019) 2, 3, 6, 9, 10

2019
[45]

IEEE TPAMI44(4), 2058–2073 (2022) 2, 3

Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time. IEEE TPAMI44(4), 2058–2073 (2022) 2, 3
[46]

In: CVPR

Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018) 8

2018
[47]

In: CVPR

Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ade20k dataset. In: CVPR. pp. 633–641 (2017) 9

2017

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.

[1] [14]

In: CVPR

Ke, Z., Liu, Y., Zhu, L., Zhao, N., Lau, R.W.: Neural preset for color style transfer. In: CVPR. pp. 14173–14182 (2023) 2, 3, 7, 9, 10, 13

work page 2023

[2] [1]

In: CVPRW

Agustsson, E., Timofte, R.: NTIRE 2017 challenge on single image super- resolution: Dataset and study. In: CVPRW. pp. 126–135 (2017) 8

work page 2017

[3] [2]

In: ICLR (2023) 2, 4

Albergo, M.S., Vanden-Eijnden, E.: Building normalizing flows with stochastic interpolants. In: ICLR (2023) 2, 4

work page 2023

[4] [3]

In: AAAI

An, J., Xiong, H., Huan, J., Luo, J.: Ultrafast photorealistic style transfer via neural architecture search. In: AAAI. pp. 10443–10450 (2020) 2, 3

work page 2020

[5] [4]

arXiv (2023) 2, 3, 9, 10

Chen, Y., Yang, H., Yang, Y., Liu, Y., Wang, W., Wen, X., Xie, C.: NLUT: Neural- based 3D lookup tables for video photorealistic style transfer. arXiv (2023) 2, 3, 9, 10

work page 2023

[6] [5]

In: WACV

Chiu, T.Y., Gurari, D.: PhotoWCT2: Compact autoencoder for photorealistic style transfer resulting from blockwise training and skip connections of high-frequency residuals. In: WACV. pp. 2868–2877 (2022) 2, 3, 6, 9, 10

work page 2022

[7] [6]

In: CVPR

Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: StyTr2: Image style transfer with transformers. In: CVPR. pp. 11326–11336 (2022) 3

work page 2022

[8] [7]

In: ICLR (2021) 9

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) 9

work page 2021