Anchoring on Reality: Breaking the Pseudo-Target Ceiling in Makeup Transfer

Bo Wei; Hong Gu; Jiachen Yang; Wangmeng Zuo; Xianhui Lin; Xiaoming Li; Xing Liu; Yi Dong; Zhongzhong Li; Zirui Wang

REVIEW 3 major objections 6 minor 54 references

Anchoring on Reality: Breaking the Pseudo-Target Ceiling in Makeup Transfer

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Makeup transfer can break its 'pseudo-target ceiling' by anchoring final supervision to the real reference image instead of the synthetic targets it was trained on.

desk verdict A genuinely new training strategy for makeup transfer—reconstructing the real reference through a differentiable carrier—with strong ablations, but the evaluation has a few soft spots that revisions should address. read the letter →

arxiv 2606.31089 v2 pith:MGVR5Q5P submitted 2026-06-30 cs.CV

Bo Wei , Xianhui Lin , Yi Dong , Zhongzhong Li , Zonghui Li , Zirui Wang , Jiachen Yang , Xing Liu

show 3 more authors

Hong Gu Xiaoming Li Wangmeng Zuo

This is my paper

classification cs.CV

keywords makeuptransferpseudo-targetceilingreality-anchoredrefinementdifferentiablecarrierdiffusiontransformerflowmatchingFaces2K2K-resolutionportraitdataset

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

ART argues that makeup transfer is bottlenecked not by the generator but by the targets used to supervise it: models trained to imitate synthetic pseudo-targets inherit their artifacts and can never surpass them. The paper's proposal is to initialize on pseudo-targets for global placement, then switch supervision to the real reference, reconstructing it from its bare-skin counterpart through a differentiable cycle. Because the transferred makeup stays in latent space as a differentiable carrier, the reconstruction loss can push gradients back into the initial prediction, penalizing any missed cosmetic detail and overriding synthetic flaws. The paper also contributes MakeupFaces2K, an 8,573-image 2K-resolution dataset. If the mechanism holds, it would explain the reported gains in fine-grained makeup fidelity, identity preservation, and background stability, and would offer a general training recipe for DiT-based image editing.

What carries the argument

The load-bearing object is the differentiable makeup carrier z-hat, a latent prediction produced by a one-step Euler update from a noise-perturbed pseudo-target. Instead of decoding z-hat to pixels, the model feeds it, together with the reference's bare-skin latent, into the same DiT to reconstruct the real reference latent; because z-hat remains in the computational graph, the reconstruction loss backpropagates through the DiT into the velocity field that produced z-hat. A controlled-noise bottleneck at sigma_tr = 0.6 modulates how much pseudo-target structure is retained, and a bottleneck loss anchors the carrier to coarse makeup placement, preventing the degenerate shortcut of copying the

What would settle it

Run Stage II with the makeup carrier detached from the computational graph (stop-gradient at z-hat) while keeping everything else identical; if outputs still improve as much as the full model, the claim that refinement gradients flow through the carrier is falsified. A complementary check: degrade the bare-skin counterpart by blending in increasing amounts of real reference makeup, and verify that makeup similarity and identity scores collapse only when the anchor approaches the full reference.

Watch

Extended reading notes

Core claim

The central claim is that the real reference image, not a synthetic approximation, is the correct supervisory anchor for makeup transfer. In Stage I the model learns coarse semantic alignment by regressing to pseudo-targets generated by a large editing model. In Stage II, the model first predicts a transfer output, keeps it as a differentiable latent makeup carrier, and then is asked to reconstruct the real reference from its extracted bare-skin counterpart using only that carrier as the makeup condition. The reconstruction loss therefore measures whether the carrier contains every cosmetic cue needed to rebuild the reference, and its gradient flows back through the carrier into the initial

Load-bearing premise

The auxiliary makeup remover, trained only on synthetic (pseudo-target, source) pairs, must produce a bare-skin counterpart of real reference images that retains genuine skin texture and leaves no makeup residue; if it fails on real photos, Stage II's reconstruction loss supervises the wrong quantity and the 'reality anchor' is not actually real.

Editorial extensions

If this is right

The final output can be better than the pseudo-targets it was initialized on; the paper shows corrections even when running on exact training pairs.
Complex cosmetics such as glitter, dense stickers, and face paint transfer with spatial details intact, while identity and background stay close to the source.
The advantage persists when weaker pseudo-targets are used for initialization, implying the method is not merely capping at pseudo-target quality.
Training can scale to 2048x2048 resolution using the new MakeupFaces2K dataset, making high-fidelity 2K makeup transfer feasible.
The same reality-anchored cycle is proposed as a general training paradigm for DiT-based image editing beyond makeup.

Reading between the lines

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

A direct extension of the logic: any edit task with a reliable 'content remover' (e.g., style transfer, relighting, expression editing) could replace pseudo-target regression with reconstruction of the real edited image from its content-only counterpart, using the edit as a differentiable carrier.
If the mechanism is right, the practical ceiling shifts from generator capacity to the makeup remover's fidelity; investing in better bare-skin extraction — real paired data, physics-based makeup removal — should yield further gains in transfer quality.
The paper's critique of global CLIP/DINO similarity suggests future makeup benchmarks should prefer rubric-based local evaluation; otherwise methods that match global color palettes but drop local decals would be over-scored.
A testable consequence: stopping gradients at the carrier should collapse the Stage II gain; if outputs still improve without gradient flow through z-hat, the proposed mechanism is not the actual cause.

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

The core idea here is real: instead of just regressing to pseudo-targets, ART trains the transfer model to reconstruct the actual reference image from its bare-skin counterpart, conditioning on a differentiable makeup carrier. That is a meaningful departure from the existing pseudo-target regression and cycle-consistency literature, and the paper does a good job showing it works. The two-stage setup (imitate pseudo-targets first, then refine against the real reference) is well motivated, and the controlled-noise bottleneck is a sensible way to keep the refinement stable. The MF2K dataset is also a useful contribution—2K-resolution in-the-wild makeup portraits with demographic diversity and artistic styles are genuinely missing from the field.

The experimental support is substantially stronger than average for this area. The ablations isolate the contribution of each design choice: Stage I alone vs. Stage II alone vs. both; varying σtr; varying λbot; different pseudo-target sources; and the inclusion of unpaired data. They also report a user study, multiple VLM judges, and a 1,000-pair LADN set to confirm the rankings hold at larger scale. The supplementary is honest about a known failure mode (lighting mismatch) and about the auxiliary remover's imperfections, even arguing that slight blurriness or residue acts as useful regularization. That level of self-scrutiny deserves credit.

The soft spots are real but not fatal. First, the headline metric is a self-defined VLM rubric without error bars or significance tests. The consistency across four VLM judges helps, but the paper would be stronger with per-pair variance or a simple significance statement. Second, σtr and λbot were tuned on the same MF2K evaluation set that later supports the SOTA claim. The tuning is coarse (a few discrete values), so this is a minor leak, but it should be noted. Third, the makeup remover R is trained only on synthetic pairs, and the robustness test only injects residue, not blurring or selective removal. The paper's argument that blurriness would force the carrier to encode identity and thereby trigger a penalty is plausible, but it is not directly validated. Finally, no code or data is released, which limits reproducibility.

None of this breaks the central claim. The mechanism is sensible, the ablations are clean, and the qualitative results—especially the training-set inference showing the model overrides pseudo-target artifacts—are compelling. I'd be happy to see this go to peer review. A serious referee should ask for error bars, a clearer statement about hyperparameter selection, and a deeper analysis of the remover's failure modes, but the paper deserves that referee time.

Referee Report

3 major / 6 minor

Summary. The paper proposes ART, a two-stage Diffusion Transformer framework for makeup transfer. Stage I initializes the model on synthetic pseudo-targets generated by a commercial editor; Stage II introduces a 'reality-anchored refinement cycle' in which the model first predicts a differentiable makeup carrier and then reconstructs the real reference latent from a bare-skin counterpart extracted by an auxiliary makeup remover. A controlled-noise bottleneck and a structural regularizer are used to stabilize training. The paper also introduces MakeupFaces2K (MF2K), a 2K-resolution in-the-wild makeup dataset with 8,573 images, and reports state-of-the-art results on MT, LADN, MT-Wild, and MF2K, along with ablations, a user study, and additional VLM judges.

Significance. If the claims hold, the reality-anchored refinement is a conceptually interesting way to escape the artifacts of synthetic pseudo-targets by using the real reference as a supervision anchor. The proposed MF2K dataset is a potentially valuable community resource. The paper is thorough in its ablations, includes multi-judge and human evaluations, and provides a larger-scale LADN robustness check. However, the central mechanism rests on the assumption that the auxiliary makeup remover generalizes from synthetic pairs to real, often artistic, references, and the evaluation has statistical and methodological gaps. These issues are load-bearing for the paper's main claims, so the result is not yet fully established.

major comments (3)

[Sec. 3.2, Eq. (3); Suppl. A.5] The reality anchor is only as real as the auxiliary remover R. R is trained exclusively on synthetic (Ipseudo, Isrc) pairs, yet it is applied to real reference images, including the artistic subset of MF2K (face paint, stickers, glitter). If R leaves residue or removes genuine skin texture, L_refine in Eq. (5) supervises the wrong target. The robustness test in Suppl. A.5 blends I_bare_ref with I_ref, which preserves high-frequency detail and does not mimic R's actual failure modes such as blurring or selective removal. The assertion that blurriness/residues act as 'implicit regularization' is post-hoc and is not validated by swapping R for a stronger remover or by measuring the information content of the residual. Please provide a controlled experiment with a cleaner remover, or real before/after makeup pairs, to establish that I_bare_ref is a faithful bare-skin counterpart for in-the-w
[Sec. 4.2, Table 1; Sec. 4.4] The primary metric MSimG is a VLM rubric proposed by the authors, and no error bars or significance tests are reported for any metric. More concerning, the two key hyperparameters σtr and λbot are selected from ablations performed on the MF2K evaluation set (Fig. 6 and Suppl. Fig. I), and the same set is then used for the final SOTA claims in Table 1. This is a form of test-set tuning and can inflate the reported advantage. Please use a validation split for hyperparameter selection and report bootstrap confidence intervals or paired significance tests for MSim, ID, and L2-M. Also, FID computed on only 100 generated images is noisy; please either use a larger sample or report the uncertainty.
[Sec. 3.2, Eqs. (6)-(7); Suppl. B.2] The paper claims to 'break the pseudo-target ceiling,' but the structural regularizer L_bottleneck explicitly regresses the velocity at σtr toward the pseudo-target latent. The method therefore remains partially dependent on pseudo-target supervision, and the 'ceiling' concept is not operationally defined. The evidence in Table 3 shows that Stage I&II beats Stage I alone, which is encouraging, but it does not quantify how much of the final output inherits pseudo-target structure versus reference-derived detail. Please provide a quantitative decomposition or an experiment on a synthetic task where the pseudo-target is known to be wrong, so that the ceiling-breaking claim can be tested more directly than by visual inspection of training-set pairs.

minor comments (6)

[Fig. 2] The labels 'Frozen' and 'Trainable' are ambiguous: the remover is trainable in Stage I but frozen in Stage II, while the transfer model is trainable in both. Please clarify the schematic.
[Sec. 4.1] LoRA rank and scaling factor are both reported as 32; please specify the alpha convention and whether rank applies to all linear layers of the DiT.
[Table 1] MAD is missing L2-M values on all sets. Please state why it is omitted (e.g., background editing or no mask support).
[Sec. 4.4 / Fig. 6] The shaded horizontal bands for continuous σtr ranges are not formally defined. Please state the sampling distribution and whether the same random seed is used across variants.
[Sec. 3.3] The claim of being the 'first 2K-resolution in-the-wild makeup portrait dataset' should be qualified by comparing against the resolution and content of existing high-resolution portrait benchmarks, or by stating the search scope.
[Sec. 4.6 / Suppl. D] Only one illumination-mismatch failure case is shown. Since the limitation is acknowledged, additional systematic analysis (e.g., varying illumination directions) would strengthen the failure-mode discussion.

Circularity Check

1 steps flagged · score 4.0 of 10

MF2K state-of-the-art is partially fitted: key Stage-II hyperparameters are tuned on the MF2K evaluation set and then reported as SOTA on that same set; the core refinement objective itself is not circular.

fitted input called prediction [Sec. 4.4 (Controlled-noise bottleneck) and Eq. (7); Table 1 MF2K Test Set; Suppl. B.2]
"We analyze the sensitivity of the controlled noise level σtr in the Stage II refinement cycle on the MF2K evaluation set. Fig. 6 reveals a clear trade-off. ... We therefore set σtr = 0.6 by default, which provides a strong balance between cosmetic fidelity and identity preservation while achieving the best background stability. Furthermore, to justify our empirical choice of the penalty weight λbot = 0.2 (defined in Eq. 7), we conduct a comprehensive visual ablation across varying values, as illustrated in Fig. I."

The two load-bearing Stage-II hyperparameters (σtr=0.6, λbot=0.2) are selected by ablating on the MF2K evaluation set, and the same MF2K Test Set is then used in Table 1 to report the headline result (Ours MSimG 9.22 vs 8.43). The reported MF2K advantage is therefore not an independent prediction: the configuration was chosen to maximize performance on that very benchmark, so the 'ceiling-breaking' claim is partly an artifact of test-set fitting rather than an out-of-sample result. Non-tuned datasets (MT, LADN, MT-Wild) provide partial independent evidence, but the primary complex-makeup benchmark is the one used for tuning.

full rationale

No self-referential equation collapse was found: Linit (Eq. 2), Lrefine (Eq. 5), and Lbottleneck (Eq. 6) are distinct training objectives; Lrefine targets the external real-reference latent z_ref, not the pseudo-target z_pseudo, and the paper demonstrates (Fig. H) that removing Lbottleneck collapses to copying the reference, so the cycle is not a definitional tautology. The auxiliary remover R is trained on synthetic (Ipseudo, Isrc) pairs, which is a domain-gap weakness for real complex makeup rather than circularity. The main circularity burden is evaluation: Sec. 4.4 explicitly tunes σtr on the MF2K evaluation set, and Suppl. B.2 tunes λbot by ablation under the same protocol, after which Table 1 reports SOTA on that same set. Because the central complex-makeup benchmark is the one used for tuning, the empirical claim is partially fitted. On non-tuned datasets the method remains competitive, supporting independent content, hence score 4 rather than higher.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim rests on two fitted hyperparameters (σtr, λbot), a synthetic-data-trained makeup remover, and proprietary pseudo-targets. The invented 'makeup carrier' is an internal latent without independent evidence. These are the main items the reader 'did not pay for upstream'.

free parameters (4)

σtr (controlled noise level) = 0.6
Selected by ablation on the MF2K evaluation set (Fig. 6) to balance MSim, ID, and L2-M; directly controls how much pseudo-target prior survives into the makeup carrier and is central to the Stage II objective.
λbot (structural regularizer weight) = 0.2
Chosen by visual ablation (Fig. I) on MF2K; weak values blur, strong values bind the output to the pseudo-target. It sets the trade-off in Eq. (7).
λrefine (refinement loss weight) = 1
Fixed weight for the refinement loss in Eq. (7); not ablated but determines the supervision strength of the reality-anchored cycle.
Remover loss weights (λlpips, λid, λlmk) = 1, 2, 10^4
Set in Suppl. B.3 to balance perceptual, identity, and landmark terms when training the auxiliary makeup remover R; these affect the quality of the bare-skin anchor that Stage II relies on.

assumptions (5)

standard math Flow matching objective (Eq. 1) is a valid training target for the conditional DiT.
Borrowed from Lipman et al. [24]; unproved background used throughout the paper.
domain assumption The pretrained FLUX.1-Kontext-dev and VAE latents preserve enough identity and makeup detail for transfer.
Backbone choice in Sec. 4.1; if the latent space discards fine texture, the refinement cycle cannot recover it.
domain assumption Pseudo-targets generated by Nano Banana Pro provide approximate global makeup placement and source identity.
Stage I and L_bottleneck rely on this; if pseudo-targets are grossly misaligned, the carrier initialization is uninformative.
ad hoc to paper The makeup remover R trained on synthetic (Ipseudo, Isrc) pairs generalizes to real reference images.
Stage II anchors on R(Iref) as the bare-skin counterpart; no real paired data validates this. Entered at Eq. (3), Sec. 3.2.
domain assumption VLM rubric scores reflect human judgment of makeup fidelity.
Primary metric MSimG/MSimQ; the authors provide reasoning logs but no rigorous validation against a large human study (Sec. 4.2, Suppl. C.2).

invented entities (1)

Differentiable makeup carrier (ẑ in Eq. 4)
purpose: A latent representation that carries reference makeup details from the transfer prediction into the reference-reconstruction cycle; gradients flow through it to the transfer model.
A proposed internal mechanism; no external falsifiable handle beyond the paper's own ablations and qualitative results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Anchoring on Reality: Breaking the Pseudo-Target Ceiling in Makeup Transfer." pith.science (2026). https://pith.science/paper/MGVR5Q5P

@misc{pith2026260631089,
  author       = {Pith},
  title        = {Pith review of: Anchoring on Reality: Breaking the Pseudo-Target Ceiling in Makeup Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MGVR5Q5P}},
  note         = {Machine review of arXiv:2606.31089}
}

read the original abstract

Makeup transfer applies a reference cosmetic style to a source face while preserving its identity and geometry. However, this task is severely hindered by the lack of real paired training data. Current methods rely on either weak priors or synthetic pseudo-targets from large-scale editing models. These paradigms provide suboptimal guidance, often leading to degraded fine-grained details, synthetic artifacts, and identity drift. To this end, we propose Anchoring on Reality Makeup Transfer (ART), a two-stage framework with a reality-anchored refinement cycle. In Stage I, the model is initialized with pseudo-targets to establish basic semantic alignment and global makeup placement. Crucially, Stage II shifts supervision from pseudo-targets to the real reference, reconstructing it from its bare-skin counterpart through a differentiable cycle that penalizes any omitted detail and overrides synthetic artifacts. Furthermore, we introduce MakeupFaces2K (MF2K), the first 2K-resolution in-the-wild makeup portrait dataset comprising 8,573 images. Extensive experiments demonstrate that our method achieves superior makeup fidelity, strong background stability, and robust identity preservation, especially for complex makeup styles.

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 9 linked inside Pith

[1]

Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report (2025)

2025
[2]

1 kontext: Flow matching for in-context image generation and editing in latent space

Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., En- glish, J., English, Z., Esser, P., Kulal, S., et al.: Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space. arXiv e-prints pp. arXiv–2506 (2025)

2025
[3]

In: Proceedings of the IEEE international conference on computer vision

Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face align- ment problem?(and a dataset of 230,000 3d facial landmarks). In: Proceedings of the IEEE international conference on computer vision. pp. 1021–1030 (2017)

2017
[4]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Chang, H., Lu, J., Yu, F., Finkelstein, A.: Pairedcyclegan: Asymmetric style trans- fer for applying and removing makeup. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 40–48 (2018) 16 B. Wei et al

2018
[5]

IEEE Transactions on Image Processing33, 2404–2418 (2024)

Chen, C., Mo, J., Hou, J., Wu, H., Liao, L., Sun, W., Yan, Q., Lin, W.: Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing33, 2404–2418 (2024)

2024
[6]

In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition

Chen,H.J.,Hui,K.M.,Wang,S.Y.,Tsao,L.W.,Shuai,H.H.,Cheng,W.H.:Beauty- glow: On-demand makeup transfer framework with reversible generative network. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10042–10050 (2019)

2019
[7]

Deng, H., Han, C., Cai, H., Han, G., He, S.: Spatially-invariant style-codes con- trolledmakeuptransfer.In:ProceedingsoftheIEEE/CVFConferenceoncomputer vision and pattern recognition. pp. 6549–6557 (2021)

2021
[8]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690–4699 (2019)

2019

Show all 54 references

[9]

Communications of the ACM63(11), 139–144 (2020)

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM63(11), 139–144 (2020)

2020
[10]

Google DeepMind: Gemini 3 Pro.https://deepmind.google/models/gemini/ pro/(2025), accessed: March 5, 2026

2025
[11]

Google DeepMind: Gemini 3 Pro Image.https://deepmind.google/models/ gemini-image/pro/(2025), accessed: March 5, 2026

2025
[12]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)

Gu, Q., Wang, G., Chiu, M.T., Tai, Y.W., Tang, C.K.: Ladn: Local adversarial disentangling network for facial makeup and de-makeup. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)

2019
[13]

In: European Conference on Computer Vision

Guo, J., Zhu, X., Yang, Y., Yang, F., Lei, Z., Li, S.Z.: Towards fast, accurate and stable 3d dense face alignment. In: European Conference on Computer Vision. pp. 152–168. Springer (2020)

2020
[14]

Information Fusion123, 103241 (2025)

He, F., Li, H., Ning, X., Li, Q.: Beautydiffusion: Generative latent decomposition for makeup transfer via diffusion models. Information Fusion123, 103241 (2025)

2025
[15]

Advances in neural information processing systems30(2017)

Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems30(2017)

2017
[16]

Advances in neural information processing systems33, 6840–6851 (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

2020
[17]

ICLR1(2), 3 (2022)

Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022)

2022
[18]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

Jayasumana, S., Ramalingam, S., Veit, A., Glasner, D., Chakrabarti, A., Kumar, S.: Rethinking fid: Towards a better evaluation metric for image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 9307–9315 (2024)

2024
[19]

In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

Jiang, W., Liu, S., Gao, C., Cao, J., He, R., Feng, J., Yan, S.: Psgan: Pose and expression robust spatial-aware gan for customizable makeup transfer. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5194–5202 (2020)

2020
[20]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

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

2019
[21]

arXiv preprint arXiv:1312.6114 (2013)

Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

2013 arXiv
[22]

In: European conference on computer vision

Kips, R., Gori, P., Perrot, M., Bloch, I.: Ca-gan: Weakly supervised color aware gan for controllable makeup transfer. In: European conference on computer vision. pp. 280–296. Springer (2020) Reality-Anchored Makeup Transfer 17

2020
[23]

In: Proceed- ings of the 26th ACM international conference on Multimedia

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

2018
[24]

arXiv preprint arXiv:2210.02747 (2022)

Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022)

2022 arXiv
[25]

IEEE Transactions on Pattern Analysis and Machine Intelligence44(11), 8538–8551 (2021)

Liu, S., Jiang, W., Gao, C., He, R., Feng, J., Li, B., Yan, S.: Psgan++: Robust detail-preserving makeup transfer and removal. IEEE Transactions on Pattern Analysis and Machine Intelligence44(11), 8538–8551 (2021)

2021
[26]

arXiv preprint arXiv:2306.06101 (2023)

Mishchenko, K., Defazio, A.: Prodigy: An expeditiously adaptive parameter-free learner. arXiv preprint arXiv:2306.06101 (2023)

2023 arXiv
[27]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Nguyen, T., Tran, A.T., Hoai, M.: 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 (CVPR). pp. 13305–13314 (June 2021)

2021
[28]

OpenAI: GPT Image 1.5 Model.https://developers.openai.com/api/docs/ models/gpt-image-1.5(2025), accessed: March 5, 2026

2025
[29]

arXiv preprint arXiv:2304.07193 (2023)

Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

2023 arXiv
[30]

Peebles,W.,Xie,S.:Scalablediffusionmodelswithtransformers.In:Proceedingsof the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023)

2023
[31]

In: International conference on machine learning

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

2021
[32]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

2022
[33]

In: Proceedings of the Computer Vision and Pattern Recognition Con- ference

Ruan, B.K., Shuai, H.H.: Mad: Makeup all-in-one with cross-domain diffusion model. In: Proceedings of the Computer Vision and Pattern Recognition Con- ference. pp. 749–758 (2025)

2025
[34]

arXiv preprint arXiv:2508.10104 (2025)

Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khali- dov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv preprint arXiv:2508.10104 (2025)

2025 arXiv
[35]

arXiv preprint arXiv:2010.02502 (2020)

Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

2010 arXiv
[36]

Advances in Neural Information Processing Systems36, 3732–3784 (2023)

Stein, G., Cresswell, J., Hosseinzadeh, R., Sui, Y., Ross, B., Villecroze, V., Liu, Z., Caterini, A.L., Taylor, E., Loaiza-Ganem, G.: Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. Advances in Neural Information Processing...

2023
[37]

In: Proceedings of the AAAI Conference on artificial intelligence

Sun, Z., Chen, Y., Xiong, S.: Ssat: A symmetric semantic-aware transformer net- work for makeup transfer and removal. In: Proceedings of the AAAI Conference on artificial intelligence. vol. 36, pp. 2325–2334 (2022)

2022
[38]

IEEE Transactions on Neural Networks and Learning Systems36(1), 1287–1301 (2023)

Sun, Z., Chen, Y., Xiong, S.: Ssat++: A semantic-aware and versatile makeup transfer network with local color consistency constraint. IEEE Transactions on Neural Networks and Learning Systems36(1), 1287–1301 (2023)

2023
[39]

Advances in Neural Information Processing Systems37, 16016–16042 (2024)

Sun, Z., Xiong, S., Chen, Y., Du, F., Chen, W., Wang, F., Rong, Y.: Shmt: Self- supervised hierarchical makeup transfer via latent diffusion models. Advances in Neural Information Processing Systems37, 16016–16042 (2024)

2024
[40]

In: Proceedings of the 18 B

Sun, Z., Xiong, S., Chen, Y., Rong, Y.: Content-style decoupling for unsupervised makeup transfer without generating pseudo ground truth. In: Proceedings of the 18 B. Wei et al. IEEE/CVF conference on computer vision and pattern recognition. pp. 7601–7610 (2024)

2024
[41]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Tan, Z., Liu, S., Yang, X., Xue, Q., Wang, X.: Ominicontrol: Minimal and universal control for diffusion transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14940–14950 (2025)

2025
[42]

arXiv preprint arXiv:2503.08280 (2025)

Tan, Z., Xue, Q., Yang, X., Liu, S., Wang, X.: Ominicontrol2: Efficient conditioning for diffusion transformers. arXiv preprint arXiv:2503.08280 (2025)

2025 arXiv
[43]

In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

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

2022
[44]

arXiv preprint arXiv:2508.05994 (2025)

Wu, H., Fu, Y., Li, Y., Gao, Y., Du, K.: Evomakeup: High-fidelity and controllable makeup editing with makeupquad. arXiv preprint arXiv:2508.05994 (2025)

2025 arXiv
[45]

In: European conference on computer vision

Xiang,J.,Chen,J.,Liu,W.,Hou,X.,Shen,L.:Ramgan:Regionattentivemorphing gan for region-level makeup transfer. In: European conference on computer vision. pp. 719–735. Springer (2022)

2022
[46]

Advances in Neural Information Processing Systems36, 10359–10384 (2023)

Xu, S., Ma, Z., Huang, Y., Lee, H., Chai, J.: Cyclenet: Rethinking cycle consistency in text-guided diffusion for image manipulation. Advances in Neural Information Processing Systems36, 10359–10384 (2023)

2023
[47]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

Yan, Q., Guo, C., Zhao, J., Dai, Y., Loy, C.C., Li, C.: Beautyrec: Robust, effi- cient, and component-specific makeup transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 1102–1110 (June 2023)

2023
[48]

In: European conference on computer vision

Yang, C., He, W., Xu, Y., Gao, Y.: Elegant: Exquisite and locally editable gan for makeup transfer. In: European conference on computer vision. pp. 737–754. Springer (2022)

2022
[49]

In: Proceedings of the Computer Vision and Pattern Recognition Conference

Zhang, J., Huang, Q., Liu, J., Guo, X., Huang, D.: Diffusion-4k: Ultra-high- resolution image synthesis with latent diffusion models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 23464–23473 (2025)

2025
[50]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

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

2018
[51]

In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers

Zhang, Y., Yuan, Y., Song, Y., Liu, J.: Stablemakeup: When real-world makeup transfer meets diffusion model. In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers. pp. 1–9 (2025)

2025
[52]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Zhang, Y., Yuan, Y., Song, Y., Wang, H., Liu, J.: Easycontrol: Adding efficient and flexible control for diffusion transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19513–19524 (2025)

2025
[53]

In: Proceedings of the IEEE interna- tional conference on computer vision

Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE interna- tional conference on computer vision. pp. 2223–2232 (2017)

2017
[54]

arXiv preprint arXiv:2203.02286 (2022)

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

2022 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report (2025)

2025

[2] [2]

1 kontext: Flow matching for in-context image generation and editing in latent space

Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., En- glish, J., English, Z., Esser, P., Kulal, S., et al.: Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space. arXiv e-prints pp. arXiv–2506 (2025)

2025

[3] [3]

In: Proceedings of the IEEE international conference on computer vision

Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face align- ment problem?(and a dataset of 230,000 3d facial landmarks). In: Proceedings of the IEEE international conference on computer vision. pp. 1021–1030 (2017)

2017

[4] [4]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Chang, H., Lu, J., Yu, F., Finkelstein, A.: Pairedcyclegan: Asymmetric style trans- fer for applying and removing makeup. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 40–48 (2018) 16 B. Wei et al

2018

[5] [5]

IEEE Transactions on Image Processing33, 2404–2418 (2024)

Chen, C., Mo, J., Hou, J., Wu, H., Liao, L., Sun, W., Yan, Q., Lin, W.: Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing33, 2404–2418 (2024)

2024

[6] [6]

In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition

Chen,H.J.,Hui,K.M.,Wang,S.Y.,Tsao,L.W.,Shuai,H.H.,Cheng,W.H.:Beauty- glow: On-demand makeup transfer framework with reversible generative network. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10042–10050 (2019)

2019

[7] [7]

Deng, H., Han, C., Cai, H., Han, G., He, S.: Spatially-invariant style-codes con- trolledmakeuptransfer.In:ProceedingsoftheIEEE/CVFConferenceoncomputer vision and pattern recognition. pp. 6549–6557 (2021)

2021

[8] [8]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4690–4699 (2019)

2019