Towards High-Fidelity, Identity-Preserving Real-Time Makeup Transfer: Decoupling Style Generation

Lydia Kin Ching Chau; Ruowei Jiang; Zhi Yu

REVIEW 4 major objections 5 minor 41 references

Towards High-Fidelity, Identity-Preserving Real-Time Makeup Transfer: Decoupling Style Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that splitting makeup transfer into transparent RGBA mask extraction and graphics-based application yields real-time, identity-preserving virtual try-on that outperforms six existing methods.

desk verdict A solid engineering story for real-time makeup transfer, but the SOTA claim rests on a same-renderer train/test loop that does not measure real-world transfer. read the letter →

arxiv 2509.02445 v2 pith:AED4ZSKU submitted 2025-09-02 cs.CV

Lydia Kin Ching Chau , Zhi Yu , Ruowei Jiang This is my paper

classification cs.CV

keywords makeuptransfervirtualtry-onRGBAmaskssemi-transparentpseudo-ground-truthtrainingidentitypreservationtemporalconsistencyreal-timevideoinference

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

This paper tries to make virtual makeup try-on practical for live video by changing what the neural network outputs. Instead of generating a full makeup-applied face, the proposed framework trains a model to estimate a transparent RGBA makeup layer—a mask with per-pixel color and opacity—that is free of the wearer's identity and skin tone. The mask is extracted once from a reference makeup image and then applied to each video frame with a lightweight warp-and-blend step, which is what makes real-time, temporally consistent transfer possible. To learn such masks, the authors build pseudo-ground-truth data from a graphics rendering engine with controllable makeup styles and from a k-means skin-tone analysis of real eye makeup images, and they add alpha-weighted reconstruction and lip-color losses. They report that the approach outperforms six prior methods on fidelity, temporal stability, and identity preservation, including for faces with diverse skin tones.

What carries the argument

The load-bearing object is the transparent RGBA makeup mask: a four-channel image in which RGB stores the makeup color and A stores per-pixel opacity. The method's key move is to make this mask the only thing a generator must predict, so makeup extraction and makeup application become two independent stages. The alpha channel is learned under the assumption that cosmetics are semi-transparent and mix with skin tone, formalized for eye makeup by Eq. (1) (alpha = max(0, 1 − CosSim(pixel, skinTone)) in LAB). At inference, the extracted mask is warped with a lightweight affine alignment and alpha-blended onto each target frame; because the same mask is reused across frames, temporal consistency

What would settle it

A controlled lab test: photograph the same face with no makeup, then with a known lipstick or eyeshadow layer whose physical coverage is measured (for example, by weighing applied product or using a calibrated swatch), and compare the model's predicted alpha mask against the known coverage. If predicted alpha does not track coverage across skin tones, the transparency model at the core of the method breaks.

Watch

Extended reading notes

Core claim

The paper's central claim is that a semi-transparent makeup style can be captured as an identity-free RGBA mask and that this representation is enough to perform makeup transfer by compositing. An RGBA mask stores, for every pixel, the makeup pigment color and an alpha value saying how opaque the makeup is; because the mask is separated from the face underneath, the same mask can be warped onto a different face without dragging the reference identity along. The authors argue this is achievable by training a UNet-based generator on pseudo ground-truth masks produced two ways: a graphics-based renderer that composites procedurally generated makeup styles onto natural faces, and a k-means clust

Load-bearing premise

The whole pipeline stands or falls on the assumption that the synthetic rendering pipeline—especially the cosine-similarity alpha rule for eye makeup—faithfully represents how real semi-transparent makeup combines with skin, so training and evaluation on this synthetic data predicts real-world performance.

Editorial extensions

If this is right

The same extracted RGBA mask can be applied to any aligned face, so the network does not need to re-run per frame; the remaining work is affine warping and alpha blending, which is the basis of the real-time claim.
Because the alpha channel encodes transparency separately from color, the method claims to avoid the shadow-like artifacts and skin-tone shifts that occur when makeup color is regressed directly on the face.
Applying one mask across a video makes temporal consistency follow from the fixed mask, not from post-hoc smoothing between frames.
The synthetic paired-data generation can be scaled to effectively unlimited makeup styles by sampling shapes, colors, finishes, and transparencies, which the paper argues is impossible with manual or VLM-based style sets.
If the reported metrics hold, the approach would be the first decoupled pipeline with simultaneous high fidelity, identity preservation, and real-time video rates on this benchmark.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial extension: the same mask-decoupling idea could be tested for other semi-transparent appearance changes (nail polish, body paint, AR overlays) where an RGBA layer is composited over varying base tones.
Editorial extension: the fairness claim would be more directly supported by measuring skin-tone invariance per subgroup—e.g., whether the extracted mask changes when the same makeup is applied to a light and dark face—rather than by an aggregate identity FID.
Editorial extension: because the evaluation generates test targets with the same renderer used for training, a stronger check would be human raters or real before/after pairs; the paper's current numbers may overstate real-world generalization.
Editorial extension: the cosine-similarity alpha formula could be independently calibrated on physical makeup swatches; if real opacity depends on layer thickness in a different way, Eq. (1) would need a data-driven replacement.

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

Colleague,

The most useful thing I can tell you about this paper is that the architecture is interesting and the ablation study is honest, but the headline quantitative claim is not trustworthy. The authors train on pseudo-GT masks produced by their graphics renderer and then evaluate by generating test targets with that same renderer. That makes Table 1 a self-consistency check, not a comparison of real makeup-transfer fidelity. I agree with the stress-test note on this; the enormous FID/PSNR margins are exactly what you'd expect from in-distribution test data.

What's genuinely new: the decoupled two-stage pipeline (extract an RGBA mask once, then apply it in real time), the graphics-rendered pseudo-GT generation with controllable shape/color/transparency, the k-means-based eye mask estimator, and the alpha-weighted reconstruction plus lip-color-regressor losses. The video inference idea is clean and the temporal consistency argument is sound. The ablation in Figure 7 gives useful evidence that each component does something.

The soft spots, in rough order of severity: (1) the circular evaluation, which I already mentioned, is load-bearing. (2) There is no runtime measurement anywhere, despite 'real-time' being in the title. (3) The identity-preservation metric FID(I) compares distributions of original vs. transferred faces without paired labels or error bars; it cannot isolate identity preservation. (4) No code is released, so auditing the exact renderer and evaluation protocol is impossible. (5) The alpha heuristic in Eq. 1 (inverse cosine similarity to skin tone) is a pragmatic guess, and the paper doesn't validate it against any real transparency data. The limitation section does concede that opaque and special-effect makeup are out of scope, which is fine, but it doesn't fix the benchmark.

Who should read this: anyone working on makeup transfer or virtual try-on will find the data-generation and loss ideas worth thinking about. It's also a good teaching example for why synthetic evaluation loops can be misleading. As it stands, I wouldn't cite it for a SOTA claim, but I'd send it to peer review so the authors are pushed to add real before/after evaluation, an independent renderer, runtime numbers, and code. The core idea is worth serious referee time.

Recommendation: engage with it as a work-in-progress, not as a proven result.

Referee Report

4 major / 5 minor

Summary. The paper proposes a decoupled virtual makeup try-on pipeline: a model extracts a transparent RGBA makeup mask from a reference image, and the mask is then warped and alpha-composited onto target frames in real time. Training data are pseudo-ground-truth pairs produced by (i) a graphics-based rendering pipeline that composites synthetic makeup masks onto FFHQ faces and (ii) a k-means-based extraction from real makeup images, where per-pixel alpha is estimated by cosine similarity to a skin-tone estimate (Eq. 1). The training objectives include alpha-weighted reconstruction, alpha L1 loss, a lip-color regressor loss, and adversarial losses. The authors compare against six baselines on FID, LPIPS, PSNR, and an identity-preservation FID(I) metric, and provide qualitative video results and ablations.

Significance. If the quantitative evaluation were valid, the main contributions - the real-time decoupled inference design, the controllable graphics-based pseudo-GT generation, and the transparency-aware losses - would be practically significant for makeup transfer, especially for video and diverse skin tones. The qualitative results and the ablation study are suggestive, and the authors are candid about the non-opaque-makeup limitation. However, the central quantitative claims rest on a circular evaluation protocol and on an identity metric that does not measure identity preservation. As presented, the paper does not demonstrate state-of-the-art performance on real makeup transfer.

major comments (4)

[Experiment, Quantitative Comparison; Table 1] The headline quantitative result is circular. The paper trains on pseudo-GT generated by the graphics rendering pipeline (Method, Graphics-Rendered Pseudo Ground Truth) and then evaluates by generating test after-makeup images 'using the graphic-based approach described in section .' - i.e., the same renderer. The proposed model is therefore tested in-distribution with respect to the exact mask shapes, alpha blending, warping, and compositing operations it was trained to invert, while all six baselines are tested out-of-distribution. The very large margins in Table 1 (FID 9.7/11.2 vs. 26.8-63.0; PSNR 40.6/41.0 vs. ~29-31) are exactly the signature of this train/eval coupling. The claim 'outperforms all baselines across all metrics' is not supported. The authors should evaluate on real before/after pairs, or at minimum on synthetic data from an independent renderer, and should also report
[Experiment, Quantitative Comparison; FID(I)] The identity-preservation metric FID(I) is not a measure of identity preservation. FID is a distribution-level distance between two sets of images; a low FID between original faces and transferred results does not imply that each individual's identity is preserved. No paired identity labels are used, and no error bars or statistical tests are reported. To support the 'identity-preserving' claim, the authors should report a paired identity metric (e.g., face-recognition cosine similarity between source and transferred images) with confidence intervals. In its current form, the FID(I) row in Table 1 does not substantiate the identity-preservation superiority claimed in the text.
[Method, Eq. (1); pseudo-label generation] The alpha estimation in Eq. (1) is a load-bearing assumption: it is used to create eye-makeup pseudo-labels from real images and to define the alpha-weighted objectives. The equation assumes a single per-region skin-tone estimate and assumes that makeup alpha is a monotone decreasing function of color dissimilarity to that skin tone. No validation is provided against ground-truth alpha composites. This assumption may fail for dark skin tones, for makeup colors close to the skin tone, or for specular/glossy products. The authors should validate Eq. (1) on synthetic composites with known alpha across a range of skin tones, or show real before/after pairs where the extracted alpha matches a manually labeled or physically measured mask.
[Video Inference Pipeline and Experiment] Temporal consistency is a headline contribution ('robust temporal consistency', 'temporally consistent video inference'), but the only evidence is qualitative video frames. No quantitative temporal metric (e.g., warping error, optical-flow consistency, or temporal FID) is reported. Since the claim is that the decoupled design guarantees temporal consistency, the authors should measure it directly on video benchmarks and compare with the baselines. This is a necessary complement to the image-quality evaluation.

minor comments (5)

[Throughout] Several cross-references are missing or rendered as 'Section .' (e.g., after Eq. (1) and in the Experiment's quantitative-evaluation paragraph). The final manuscript must fix these LaTeX cross-reference issues.
[Abstract and Introduction] Typos: 'identify features' should be 'identity features' (Abstract); 'especially or video-based' should be 'especially for video-based' (Introduction).
[Figure 4 caption and Experimental Setting] Typo: 'makup' should be 'makeup'. In the Experimental Setting, 'stragtegies' and 'surpervised' should be 'strategies' and 'supervised'.
[Eq. (2) and ablation section] The loss notation is inconsistent: Eq. (2) uses lambda with a superscript 'p' for the adversarial term but the L_Adv terms do not carry the same superscript, and 'lambda_p Adv' is typeset ambiguously. The ablation study is purely qualitative; adding quantitative values (e.g., FID or mask IoU for each ablated variant) would strengthen the claims.
[Reproducibility] No code or pretrained models are released, despite links to metric implementations. Releasing the makeup extraction models, or at least a demo, would improve reproducibility and allow independent verification of the real-time and fidelity claims.

Circularity Check

1 steps flagged · score 7.0 of 10

Table 1's SOTA claim rests on same-renderer train/test evaluation: the model is trained on graphics-rendered pseudo-GT and tested on graphics-rendered targets, so the comparison measures self-reconstruction rather than real makeup transfer.

fitted input called prediction [Experiment, Quantitative Comparison (synthetic evaluation strategy; Table 1)]
"Due to the lack of real before-and-after makeup image pairs, we adopt a synthetic evaluation strategy. We randomly sample two no-makeup faces at a time and apply makeup to both using the graphic-based approach described in section . Makeup is then transferred from the after-makeup image of the first face to the before-makeup image of the second face. The transferred result is compared to the after-makeup image of the second face. This process is repeated across all faces in the dataset. The evaluation results, reported in Table 1, show that our method outperforms all baselines across all metri"

The 'graphic-based approach' used to create the test targets is the same graphics-rendering pipeline used to build the training pseudo-ground-truth, as stated in the Method: 'we sample a synthesized mask for each natural face image and apply it using 2D thin-plate spline warping, followed by alpha blending... The resulting dataset serves as pseudo ground truth for supervised training.' Thus the Table 1 test targets are samples from the same renderer distribution the model was trained to invert. High PSNR (40.6/41.0 vs ~29–31) and low FID (9.7/11.2 vs 26.8–63.0) reflect in-distribution self-reconstruction, not superior real makeup transfer. The 'ground-truth' after-makeup image is defined by the same compositing operator that generated the training labels; the model is measured on reproduci

full rationale

The paper's core engineering contribution—a decoupled makeup-mask extraction and rendering pipeline—is not inherently circular: the architecture, losses, and real-image qualitative comparisons have independent content. The alpha heuristic in Eq. 1 is an assumption about transparency estimation, not a circular derivation. The self-citation to Chau et al. 2024 for face tracking is a component choice and is not load-bearing for the main claim. However, the strongest quantitative claim, 'our method outperforms all baselines across all metrics,' is supported only by a synthetic evaluation in which the ground-truth after-makeup images are generated with 'the graphic-based approach described in section [Method]'—the same renderer used to synthesize the training pseudo-ground-truth masks. This makes the evaluation a self-consistency check: the model is asked to reproduce the output of the very pipeline that produced its training labels, while all six baselines are evaluated out-of-distribution with respect to that pipeline. The identity-preservation FID on real references is weaker evidence (no paired identity labels, no error bars) but is not itself circular. The absence of released code further prevents independent audit, but the circularity finding rests on the quoted train/test coupling. Overall score: 7, because the central SOTA claim reduces to same-renderer train/test coupling for external validity.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The method's central results depend on the fidelity of a synthetic rendering pipeline, a hand-set clustering configuration, and a set of loss weights. These are treated as ground truth or hyperparameters without independent verification. The evaluation protocol uses the same rendering pipeline, so the quantitative claim is not independent of these settings.

free parameters (6)

k (k-means clusters) = 6
Hand-set in the k-means skin-tone estimation for eye makeup pseudo labels; affects alpha estimates.
s (top color clusters for skin tone) = 2
Hand-set; used in weighted average of most frequent colors to estimate skin tone.
lambda_Recon = 100
Hand-tuned loss weight in Eq. 2.
lambda_Adv = 10
Hand-tuned loss weight in Eq. 2.
lambda_Alpha = 100
Hand-tuned loss weight in Eq. 2.
lambda_LipCol = 50
Hand-tuned loss weight in Eq. 2.

assumptions (6)

domain assumption Graphics-rendered pseudo-GT is a faithful proxy for real makeup transparency.
The method and evaluation both rely on synthetic masks composited onto FFHQ faces; if these do not match real makeup optics, the model won't generalize. Invoked in 'Graphics-Rendered Pseudo Ground Truth'.
ad hoc to paper Makeup alpha can be estimated as max(0, 1 - CosSim(x_LAB, skinTone_LAB)) (Eq. 1).
Used to create eye-makeup pseudo labels from real images; no derivation or validation is provided.
domain assumption A per-pixel RGBA mask is a sufficient representation to capture makeup style and disentangle it from identity.
The entire pipeline is built on this representation; if real makeup cannot be represented this way (e.g., due to lighting interactions), the method fails.
ad hoc to paper Synthetic evaluation on data from the same rendering pipeline as training measures real-world performance.
The Experiment section uses the authors' own graphics-based approach to generate test targets, making the comparison self-referential.
standard math Thin-plate spline warping and k-means clustering are valid for aligning and labeling makeup regions.
Standard computational geometry and clustering methods; not specific to this paper.
domain assumption The authors' prior face alignment model (Chau, Yu, Jiang 2024) is real-time and accurate.
The inference pipeline depends on this model for alignment; no details are provided in this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards High-Fidelity, Identity-Preserving Real-Time Makeup Transfer: Decoupling Style Generation." pith.science (2026). https://pith.science/paper/AED4ZSKU

@misc{pith2026250902445,
  author       = {Pith},
  title        = {Pith review of: Towards High-Fidelity, Identity-Preserving Real-Time Makeup Transfer: Decoupling Style Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AED4ZSKU}},
  note         = {Machine review of arXiv:2509.02445}
}

read the original abstract

We present a novel framework for real-time virtual makeup try-on that achieves high-fidelity, identity-preserving cosmetic transfer with robust temporal consistency. In live makeup transfer applications, it is critical to synthesize temporally coherent results that accurately replicate fine-grained makeup and preserve user's identity. However, existing methods often struggle to disentangle semitransparent cosmetics from skin tones and other identify features, causing identity shifts and raising fairness concerns. Furthermore, current methods lack real-time capabilities and fail to maintain temporal consistency, limiting practical adoption. To address these challenges, we decouple makeup transfer into two steps: transparent makeup mask extraction and graphics-based mask rendering. After the makeup extraction step, the makeup rendering can be performed in real time, enabling live makeup try-on. Our makeup extraction model trained on pseudo-ground-truth data generated via two complementary methods: a graphics-based rendering pipeline and an unsupervised k-means clustering approach. To further enhance transparency estimation and color fidelity, we propose specialized training objectives, including alpha-weighted reconstruction and lip color losses. Our method achieves robust makeup transfer across diverse poses, expressions, and skin tones while preserving temporal smoothness. Extensive experiments demonstrate that our approach outperforms existing baselines in capturing fine details, maintaining temporal stability, and preserving identity integrity.

Figures

Figures reproduced from arXiv: 2509.02445 by the authors.

**Figure 1.** Our method extracts high-fidelity RGBA masks that disentangle makeup from skin tone and identity features, enabling [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Comparison of pseudo ground truth generation [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Graphic-rendering pseudo ground truth pipeline [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Pipeline for generating an eye makeup mask us [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 5.** Figure 5: Left: Model architecture during inference. Right: Training architecture of the lip model. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 6.** Figure 6: Qualitative comparison. The makeup references are from Makeup-Wild dataset(Jiang et al. 2020) and the natural faces are from makeup transfer dataset(Li et al. 2018). Zoom in to see the makeup details. transfer across diverse skin tones. Additional qualitative results …

**Figure 7.** Figure 7: Ablation study on Makeup-Wild dataset(Jiang et al. 2020). (a) Comparison between equal-weighted and alpha [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Illustration of limitations. Our method relies on some visibility of the underlying skin tone to estimate makeup [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: Additional qualitative comparison with state of the art methods. The makeup references are from Makeup-Wild [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

**Figure 10.** Figure 10: Additional qualitative comparison with state of the art methods. The makeup references and faces are both from [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

**Figure 11.** Figure 11: Additional video testing result. The makeup reference images are from Makeup-Wild dataset (Jiang et al. 2020) and [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 32 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Bookstein, F. L. 1989. Principal warps: Thin-plate splines and the decomposition of deformations. IEEE Transactions on pattern analysis and machine intelligence, 11(6): 567--585

work page 1989
[4]

Chang, H.; Lu, J.; Yu, F.; and Finkelstein, A. 2018. Pairedcyclegan: Asymmetric style transfer for applying and removing makeup. In Proceedings of the IEEE conference on computer vision and pattern recognition, 40--48

work page 2018
[5]

Chau, K. C. L.; Yu, Z.; and Jiang, R. 2024. Occlusion-Aware Real-Time Tiny Facial Alignment Model for Makeup Virtual Try-On. In 2024 International Symposium on Multimedia (ISM), 221--224. IEEE

work page 2024
[6]

Deng, H.; Han, C.; Cai, H.; Han, G.; and He, S. 2021. Spatially-invariant style-codes controlled makeup transfer. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 6549--6557

work page 2021
[7]

H.; Chechik, G.; and Cohen-Or, D

Gal, R.; Alaluf, Y.; Atzmon, Y.; Patashnik, O.; Bermano, A. H.; Chechik, G.; and Cohen-Or, D. 2022. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618

arXiv 2022
[8]

Grishchenko, I.; Ablavatski, A.; Kartynnik, Y.; Raveendran, K.; and Grundmann, M. 2020. Attention mesh: High-fidelity face mesh prediction in real-time. arXiv preprint arXiv:2006.10962

arXiv 2020

Show all 41 references

[9]

T.; Tai, Y.-W.; and Tang, C.-K

Gu, Q.; Wang, G.; Chiu, M. T.; Tai, Y.-W.; and Tang, C.-K. 2019. Ladn: Local adversarial disentangling network for facial makeup and de-makeup. In Proceedings of the IEEE/CVF International conference on computer vision, 10481--10490

2019
[10]

Guo, D.; and Sim, T. 2009. Digital face makeup by example. In 2009 IEEE conference on computer vision and pattern recognition, 73--79. IEEE

2009
[11]

Heusel, M.; Ramsauer, H.; Unterthiner, T.; Nessler, B.; and Hochreiter, S. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30

2017
[12]

Isola, P.; Zhu, J.-Y.; Zhou, T.; and Efros, A. A. 2017. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1125--1134

2017
[13]

Jiang, W.; Liu, S.; Gao, C.; Cao, J.; He, R.; Feng, J.; and Yan, S. 2020. Psgan: Pose and expression robust spatial-aware gan for customizable makeup transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5194--5202

2020
[14]

Jin, Q.; Chen, X.; Jin, M.; Chen, Y.; Shi, R.; Zheng, Y.; Zhu, Y.; and Ni, B. 2024. Toward Tiny and High-Quality Facial Makeup with Data Amplify Learning. In European Conference on Computer Vision, 340--356. Springer

2024
[15]

Johnson, J.; Alahi, A.; and Fei-Fei, L. 2016. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, 694--711. Springer

2016
[16]

Karras, T.; Laine, S.; and Aila, T. 2019. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4401--4410

2019
[17]

P.; and Ba, J

Kingma, D. P.; and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

2014 arXiv
[18]

Kubelka, P. 1931. Ein beitrag zur optik der farbanstriche. Z. tech. Phys, 12: 593--601

1931
[19]

Li, C.; Zhou, K.; and Lin, S. 2015. Simulating makeup through physics-based manipulation of intrinsic image layers. In Proceedings of the IEEE Conference on computer vision and pattern recognition, 4621--4629

2015
[20]

Li, T.; Qian, R.; Dong, C.; Liu, S.; Yan, Q.; Zhu, W.; and Lin, L. 2018. Beautygan: Instance-level facial makeup transfer with deep generative adversarial network. In Proceedings of the 26th ACM international conference on Multimedia, 645--653

2018
[21]

Mirza, M.; and Osindero, S. 2014. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784

2014 arXiv
[22]

T.; and Hoai, M

Nguyen, T.; Tran, A. T.; and Hoai, M. 2021. Lipstick ain't enough: beyond color matching for in-the-wild makeup transfer. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 13305--13314

2021
[23]

Ning, X.; Duan, P.; Li, W.; and Zhang, S. 2020. Real-time 3D face alignment using an encoder-decoder network with an efficient deconvolution layer. IEEE Signal Processing Letters, 27: 1944--1948

2020
[24]

Ruiz, N.; Li, Y.; Jampani, V.; Pritch, Y.; Rubinstein, M.; and Aberman, K. 2023. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22500--22510

2023
[25]

Seitzer, M. 2020. pytorch-fid: FID Score for PyTorch . https://github.com/mseitzer/pytorch-fid. Version 0.3.0

2020
[26]

Sun, Z.; Chen, Y.; and Xiong, S. 2022. Ssat: A symmetric semantic-aware transformer network for makeup transfer and removal. In Proceedings of the AAAI Conference on artificial intelligence, volume 36, 2325--2334

2022
[27]

Sun, Z.; Xiong, S.; Chen, Y.; Du, F.; Chen, W.; Wang, F.; and Rong, Y. 2024 a . SHMT: Self-supervised Hierarchical Makeup Transfer via Latent Diffusion Models. Advances in neural information processing systems

2024
[28]

Sun, Z.; Xiong, S.; Chen, Y.; and Rong, Y. 2024 b . Content-style decoupling for unsupervised makeup transfer without generating pseudo ground truth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7601--7610

2024
[29]

S.; and Xu, Y.-Q

Tong, W.-S.; Tang, C.-K.; Brown, M. S.; and Xu, Y.-Q. 2007. Example-based cosmetic transfer. In 15th Pacific Conference on Computer Graphics and Applications (PG'07), 211--218. IEEE

2007
[30]

Wan, Z.; Chen, H.; An, J.; Jiang, W.; Yao, C.; and Luo, J. 2022. Facial attribute transformers for precise and robust makeup transfer. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, 1717--1726

2022
[31]

Xu, L.; Du, Y.; and Zhang, Y. 2013. An automatic framework for example-based virtual makeup. In 2013 IEEE International Conference on Image Processing, 3206--3210. IEEE

2013
[32]

C.; and Li, C

Yan, Q.; Guo, C.; Zhao, J.; Dai, Y.; Loy, C. C.; and Li, C. 2023. Beautyrec: Robust, efficient, and component-specific makeup transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1102--1110

2023
[33]

Yang, C.; He, W.; Xu, Y.; and Gao, Y. 2022. Elegant: Exquisite and locally editable gan for makeup transfer. In European Conference on Computer Vision, 737--754. Springer

2022
[34]

C.; Cai, W.; and Wu, W

Yu, J.; Zhu, H.; Jiang, L.; Loy, C. C.; Cai, W.; and Wu, W. 2023. CelebV-Text : A Large-Scale Facial Text-Video Dataset. In CVPR

2023
[35]

Zhang, H.; Chen, W.; He, H.; and Jin, Y. 2019. Disentangled makeup transfer with generative adversarial network. arXiv preprint arXiv:1907.01144

2019 arXiv
[36]

Zhang, J.; Shan, S.; Kan, M.; and Chen, X. 2014. Coarse-to-fine auto-encoder networks (cfan) for real-time face alignment. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, 1--16. Springer

2014
[37]

A.; Shechtman, E.; and Wang, O

Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, 586--595

2018
[38]

Zhang, Y.; Wei, L.; Zhang, Q.; Song, Y.; Liu, J.; Li, H.; Tang, X.; Hu, Y.; and Zhao, H. 2024. Stable-makeup: When real-world makeup transfer meets diffusion model. arXiv preprint arXiv:2403.07764

2024 arXiv
[39]

Zhao, Y.; Liu, Y.; Shen, C.; Gao, Y.; and Xiong, S. 2020. MobileFAN: Transferring deep hidden representation for face alignment. Pattern Recognition, 100: 107114

2020
[40]

Zhu, J.-Y.; Park, T.; Isola, P.; and Efros, A. A. 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2223--2232

2017
[41]

Zhu, M.; Yi, Y.; Wang, N.; Wang, X.; and Gao, X. 2022. Semi-parametric makeup transfer via semantic-aware correspondence. arXiv preprint arXiv:2203.02286

2022 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Bookstein, F. L. 1989. Principal warps: Thin-plate splines and the decomposition of deformations. IEEE Transactions on pattern analysis and machine intelligence, 11(6): 567--585

work page 1989

[4] [4]

Chang, H.; Lu, J.; Yu, F.; and Finkelstein, A. 2018. Pairedcyclegan: Asymmetric style transfer for applying and removing makeup. In Proceedings of the IEEE conference on computer vision and pattern recognition, 40--48

work page 2018

[5] [5]

Chau, K. C. L.; Yu, Z.; and Jiang, R. 2024. Occlusion-Aware Real-Time Tiny Facial Alignment Model for Makeup Virtual Try-On. In 2024 International Symposium on Multimedia (ISM), 221--224. IEEE

work page 2024

[6] [6]

Deng, H.; Han, C.; Cai, H.; Han, G.; and He, S. 2021. Spatially-invariant style-codes controlled makeup transfer. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 6549--6557

work page 2021

[7] [7]

H.; Chechik, G.; and Cohen-Or, D

Gal, R.; Alaluf, Y.; Atzmon, Y.; Patashnik, O.; Bermano, A. H.; Chechik, G.; and Cohen-Or, D. 2022. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618

arXiv 2022

[8] [8]

Grishchenko, I.; Ablavatski, A.; Kartynnik, Y.; Raveendran, K.; and Grundmann, M. 2020. Attention mesh: High-fidelity face mesh prediction in real-time. arXiv preprint arXiv:2006.10962

arXiv 2020