Pith. sign in

REVIEW 5 major objections 5 minor 69 references

HairShifter: Consistent and High-Fidelity Video Hair Transfer via Anchor-Guided Animation

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read HairShifter claims to be the first dedicated video hairstyle transfer framework, producing temporally consistent, identity-preserving results by animating a single high-quality anchor frame.

desk verdict A solid first dedicated video hair transfer system with strong empirical results, but the decoupling assumption needs a direct test and the evaluation has a few unpolished edges. read the letter →

arxiv 2507.12758 v1 pith:XA3U3WNS submitted 2025-07-17 cs.CV

Wangzheng Shi , Yinglin Zheng , Yuxin Lin , Jianmin Bao , Ming Zeng , Dong Chen This is my paper
classification cs.CV
keywords videohairtransferimageanimationeditinganchorframetemporalconsistencygatedfusionSPADEdecoder
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

Video hairstyle transfer has lagged behind single-image hair transfer because frame-by-frame editing flickers and holistic animation warps non-hair regions. HairShifter tries to establish that the two goals can be met at once by splitting the problem into an anchor stage and an animation stage: an image hair transfer module produces one high-quality edited frame, and a separate network animates that frame using motion and non-hair context from the driving video. The paper's central claim is that a deliberately wrong pseudo driving frame during training forces the network to take hairstyle only from the anchor and everything else from the video, yielding state-of-the-art visual quality, temporal consistency, non-hair fidelity, and scalability. If this holds, video hair transfer becomes practical for social media, gaming, and film editing without user-provided masks or per-video optimization.

What carries the argument

The load-bearing mechanism is an Anchor Frame + Animation pipeline whose heart is the pseudo driving frame training strategy: during training, the image hair transfer module generates a driving frame with deliberately wrong hair, so the animation network must learn hair appearance from the source anchor and non-hair dynamics from the driving input. The other central component is the Multi-Scale Gated SPADE Decoder, a dual-pathway decoder that fuses warped hair features, non-hair context features, and a hair mask through sigmoid-gated spatial blending at multiple scales, producing seamless integration while preserving identity and background.

What would settle it

At inference, take a real driving frame and use the same image hair transfer module to replace its hair with a different reference style while keeping pose, expression, and background identical; if the network's output hair visibly changes, the network is still reading hair cues from the driving frame, falsifying the claimed decoupling. Alternatively, retrain the model without the pseudo-frame strategy and show that video quality and non-hair fidelity do not degrade.

Watch

Extended reading notes

Core claim

The paper claims that video hair transfer can be achieved without choosing between temporal stability and non-hair preservation. HairShifter first creates a static anchor image with the target hairstyle using an adapted image hair transfer module, then animates this anchor frame-by-frame with a network trained under a decoupling objective. During training, the driving frame's hair is replaced with a random incorrect hairstyle, producing a pseudo driving frame; the network must reconstruct the ground-truth frame from the correct hairstyle in the anchor and the non-hair content of the pseudo frame. This forces the hair appearance stream and the non-hair context stream to separate. A Multi-Scale Gated SPADE Decoder then fuses the two streams with a learned spatial gate modulated by a hair mask, blending synthesized hair onto preserved identity, expression, and background. The paper reports state-of-the-art scores on video quality, non-hair fidelity, and temporal-consistency benchmarks, plus user-study preference, and positions the framework as the first designed specifically for video hairstyle transfer.

Load-bearing premise

The whole method rests on the assumption that pseudo driving frames with deliberately wrong hair teach the network to ignore the driving frame's hair completely and take hairstyle only from the anchor, and that this separation still holds when real driving frames are used at inference.

Editorial extensions

If this is right

  • Frame-wise image hair transfer models can be replaced by the anchor-plus-animation design to remove flicker without per-video tuning.
  • Because only the anchor frame requires the expensive diffusion-based image hair transfer, longer videos amortize the per-frame cost to a small constant, enabling real-time editing of long sequences.
  • Non-hair regions such as identity, expression, and background can be preserved exactly, making the method suitable for video editing tasks where accurate masks are unavailable.
  • The decoupling training makes the network robust to pose differences between the reference hair image and the driving video, a common failure of static hair transfer models.
  • The framework establishes a baseline for video hairstyle transfer that later work can compare against directly.

Reading between the lines

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

  • If the decoupling is as clean as claimed, the same pseudo-frame trick could be applied to other region-specific video edits, such as clothing, glasses, or tattoos, wherever a strong image editor exists to produce wrong-region pseudo frames.
  • A testable extension would measure whether the network's output hair changes when the driving frame's real hair is swapped at inference; if it does, the claimed decoupling is incomplete.
  • The scalability claim implies that per-frame cost approaches a fixed small constant, so the method should handle arbitrarily long video, but only if the anchor frame quality remains stable for extreme poses.
  • The acknowledged long-to-short hair limitation suggests a natural follow-up that combines anchor animation with video inpainting for the newly uncovered scalp or background regions.
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.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes HairShifter, an 'Anchor Frame + Animation' framework for video hairstyle transfer. At inference, an image hair transfer (IHT) module creates a static anchor frame with the target hairstyle, and an animation network G synthesizes each output frame by combining hair appearance from the anchor with motion and non-hair context from the driving video. At training time, a 'pseudo driving frame' is generated by IHT with a random wrong hairstyle, and G is trained to reconstruct the original frame from the anchor and the pseudo frame, forcing hair/non-hair decoupling. A dedicated non-hair context encoder and a Multi-Scale Gated SPADE (MSG-SPADE) decoder fuse the disentangled streams. Quantitative comparisons on 200 held-out CelebV-HQ videos, an ablation study, and a user study are reported, alongside claims of state-of-the-art quality, temporal consistency, and real-time scalability.

Significance. If the reported results hold, HairShifter would be a meaningful step in a previously under-served task: video hair transfer with explicit temporal consistency and non-hair preservation. The framework is well-motivated, and the held-out evaluation against several external baselines is a strength. The design of the pseudo-driving-frame training signal is creative, and the ablation study shows monotonic gains from the proposed components. However, the central claim that the network truly decouples hair from non-hair at inference is not directly verified, and the quantitative evidence lacks uncertainty quantification. The paper also contains a clear internal contradiction in the ablation analysis. With additional targeted experiments and corrected reporting, the contribution could be solid; in its current form, the evidence is promising but not fully conclusive.

major comments (5)
  1. [Sec. 3.2 and Table 2] The decoupling training assumes that training G(I_s, I_d') with pseudo frames I_d'=IHT(I_d, R_random) teaches the network to take hair appearance exclusively from I_s and to ignore real driving-frame hair at inference. Because I_d' differs from real I_d precisely in the hair region, and because the same IHT module is used for both pseudo-frame generation and anchor generation, this generalization is load-bearing but untested. The paper provides no metric that measures whether the output hair matches the reference rather than the driving frame. Please add a direct hair-source fidelity evaluation, e.g., report hair-region FID/LPIPS between the output and I_s (and between the output and I_d), or run a controlled experiment in which the driving frame's hair is replaced by a known different style and measure the output's hair-source attribution.
  2. [Sec. 4.3, Table 2] The claim that HairShifter achieves 'state-of-the-art performance across the board' is not supported by Table 2: HairShifter is second on PSNR (32.577 vs. 33.745) and second on VBench-TF (99.032 vs. 99.191). More importantly, no error bars, confidence intervals, or significance tests are reported for any metric, so the many close scores (e.g., SSIM 0.959 vs. 0.953, L1 0.025 vs. 0.028) may not reflect a reliable advantage. Please report per-video standard deviations or confidence intervals and, where relevant, pairwise significance tests; otherwise the state-of-the-art claim is overstated.
  3. [Sec. 4.4, Table 4] The ablation text states that 'Naive Pixel Blending (Setting 2) ... degrades overall quality (worse FID/FVD)', but Table 4 shows Setting 2 improves both FID-VID (15.011 vs. 21.007) and FVD (183.667 vs. 242.575) relative to Setting 1. This is a direct numerical contradiction. Please correct the text or the table and re-interpret the ablation result accordingly, since the current discussion misrepresents the effect of naive blending.
  4. [Sec. 4.2.1 and Appendix A.2] The VideoGrain baseline is confounded in two ways: it is text-driven with prompts generated by an MLLM from reference images, while all other baselines use direct image references or the same IHT anchor, and its editing region is set to the entire image. The manuscript acknowledges that text provides less precise visual control, but this baseline is still used in Table 2 and the user study as representative of the video-editing paradigm. Please either replace it with a reference-image-driven video editing baseline, or clearly present VideoGrain as an exploratory text-driven comparison and exclude it from claims of superiority over the video-editing paradigm.
  5. [Sec. 3.2 and Sec. 4.1] The training protocol does not specify how the anchor frame I_s is constructed relative to the pseudo-driving frame I_d'. If I_s and I_d' are generated from the same frame (and therefore share the same pose), the warping and gating modules are trained only with trivial source-driving pose differences, whereas inference (Sec. 3.5.2) must handle an anchor frame whose pose differs from each driving frame. Please specify whether I_s comes from the same timestamp or a different timestamp during training, and if poses are matched, evaluate performance across varying anchor-to-driving pose gaps or augment training with cross-pose pairs.
minor comments (5)
  1. [Abstract] The phrase 'a Image Hair Transfer' is a grammatical error and should read 'an Image Hair Transfer'.
  2. [Figure 2 caption] The caption references 'Sec. 3.3', 'Sec. 3.4', and 'Sec. 3.5' for pseudo-frame generation, disentangled feature encoding, and multi-scale fusion, but the corresponding subsections in the text are numbered 3.2, 3.3, and 3.4; please align these references.
  3. [Figure 2 caption] The word 'Distangled' should be 'Disentangled'.
  4. [Sec. 3.4, Eq. (2)] The notation for the synthesis-pathway activation is inconsistent: Eq. (2) uses h_w^k while the surrounding text and Figure 3 describe it as h_s^k; please unify the notation.
  5. [Figure 7 caption] The caption contains the typo 'input posV Ce', which appears to be a corrupted version of 'input pose'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: HairShifter's SOTA claim rests on held-out benchmark comparisons, not on a quantity equivalent to its own fitted inputs.

full rationale

The derivation chain is: IHT (fine-tuned Stable-Hair) generates the anchor I_s and pseudo frames I_d'; animation network G is trained with reconstruction losses to map (I_s, I_d') to I_d; at inference G animates I_s with real driving frames I_d. No equation in the paper defines a predicted quantity in terms of the same quantity, and no fitted parameter is renamed as a prediction. The training target I_d is the original frame, while the claimed capability (transferring R_target's hair via I_s) is evaluated on 200 held-out CelebV-HQ videos with random CelebA-HQ references against external baselines, plus a user study and ablations. The same IHT is used for training-time pseudo frames and inference-time anchors (Secs. 3.1 and 3.2), which aligns the artifact distribution rather than making the output equal to the input by construction. The self-cited face parser [61] supplies masks for losses and non-hair metrics; it is a tool, not the load-bearing result. Sec. 4.5 openly lists limitations (long-to-short hair transfer, limited secondary hair dynamics, hand-hair interaction artifacts), further indicating the method is not forced by its own construction. The decoupling assumption that invariance learned on IHT pseudo frames transfers to real driving frames is an empirical generalization risk, not a circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method's central claim rests on several domain assumptions (mask accuracy, motion model validity, pseudo-frame decoupling) and hand-set training hyperparameters, but introduces no new physical or mathematical entities. The main unproven premise is the pseudo-frame decoupling assumption.

free parameters (3)
  • Loss weights (lambda_adv, lambda_p, lambda_rec, lambda_hair, lambda_face) = all 1.0
    Chosen by hand to balance adversarial, perceptual, reconstruction, hair, and face losses; no sensitivity study reported (Sec 3.5.1).
  • Training schedule for animation network = 100 epochs, batch size 4, learning rate 2e-5
    Set heuristically; no grid search reported (Sec 4.1).
  • IHT fine-tuning schedule = 100,000 steps, batch size 8, learning rate 2e-5
    Set heuristically for pose-invariant image hair transfer; no ablation of this schedule (Appendix A.1).
assumptions (5)
  • domain assumption The face parsing network [61] gives accurate binary hair masks for every training and evaluation frame.
    Used for pseudo-frame generation, mask guidance in Eq. 1, and non-hair region metrics (Sec 3.3, 3.4, 4.2.2).
  • domain assumption LivePortrait's motion estimator and warping correctly capture pose and expression, and stay valid under the new decoder.
    Frozen motion components are reused from LivePortrait (Sec 3.3); any failure in motion estimation propagates to output.
  • ad hoc to paper IHT-generated pseudo driving frames with random incorrect hairstyles force complete hair/non-hair decoupling.
    This is the central training premise (Sec 3.2), not independently verified.
  • domain assumption The CelebV-HQ split for evaluation is clean and IHT fine-tuning did not use evaluation identities.
    Paper states evaluation is outside the training set but IHT fine-tuning data overlap is not specified (Sec 4.1, 4.3, App A.1).
  • domain assumption Random reference hairs from CelebA-HQ provide enough diversity for effective decoupling training.
    Used to construct pseudo frames (Sec 4.2); no analysis of coverage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HairShifter: Consistent and High-Fidelity Video Hair Transfer via Anchor-Guided Animation." pith.science (2026). https://pith.science/paper/XA3U3WNS

@misc{pith2026250712758,
  author       = {Pith},
  title        = {Pith review of: HairShifter: Consistent and High-Fidelity Video Hair Transfer via Anchor-Guided Animation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XA3U3WNS}},
  note         = {Machine review of arXiv:2507.12758}
}
read the original abstract

Hair transfer is increasingly valuable across domains such as social media, gaming, advertising, and entertainment. While significant progress has been made in single-image hair transfer, video-based hair transfer remains challenging due to the need for temporal consistency, spatial fidelity, and dynamic adaptability. In this work, we propose HairShifter, a novel "Anchor Frame + Animation" framework that unifies high-quality image hair transfer with smooth and coherent video animation. At its core, HairShifter integrates a Image Hair Transfer (IHT) module for precise per-frame transformation and a Multi-Scale Gated SPADE Decoder to ensure seamless spatial blending and temporal coherence. Our method maintains hairstyle fidelity across frames while preserving non-hair regions. Extensive experiments demonstrate that HairShifter achieves state-of-the-art performance in video hairstyle transfer, combining superior visual quality, temporal consistency, and scalability. The code will be publicly available. We believe this work will open new avenues for video-based hairstyle transfer and establish a robust baseline in this field.

Figures

Figures reproduced from arXiv: 2507.12758 by the authors.

Figure 1
Figure 1. High-Fidelity Video Hairstyle Transfer with HairShifter. We successfully transfers the reference hairstyle while [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the HairShifter (a) During training, we enforces hair/non-hair decoupling using Pseudo Driving Frame [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The Multi-Scale Gated SPADE (MSG-SPADE) Decoder. (a) Features dual pathways for synthesis ( [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative Comparison with other methods. HairShifter achieves more refined and stable hairstyle transfer while [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison for the Ablation Study. Com [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Example limitation showing artifacts when trans [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of our adapted Image Hair Trans [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 35 canonical work pages

  1. [1]

    Jianhong Bai, Tianyu He, Yuchi Wang, Junliang Guo, Haoji Hu, Zuozhu Liu, and Jiang Bian. 2024. Uniedit: A unified tuning-free framework for video motion and appearance editing. arXiv preprint arXiv:2402.13185 (2024)

    arXiv 2024

  2. [2]

    Joao Carreira and Andrew Zisserman. 2017. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 6299–6308

    2017

  3. [3]

    Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. 2023. Pix2video: Video edit- ing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 23206–23217

    work page 2023

  4. [4]

    Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. 2023. Stablevideo: Text- driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 23040–23050

    work page 2023

  5. [5]

    Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. 2023. Flatten: optical flow-guided attention for consistent text-to-video editing. arXiv preprint arXiv:2310.05922 (2023)

    arXiv 2023

  6. [6]

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 4690–4699

    2019

  7. [7]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34 (2021), 8780–8794

    2021

  8. [8]

    Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. 2024. Emoportraits: Emotion- enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 8498–8507

    work page 2024

Show all 69 references
  1. [9]

    Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Vic- tor Lempitsky, and Egor Zakharov. 2022. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia. 2663–2671

    2022

  2. [10]

    Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. 2023. Tokenflow: Consis- tent diffusion features for consistent video editing.arXiv preprint arXiv:2307.10373 (2023)

    2023 arXiv

  3. [11]

    Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang

  4. [12]

    Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. 2024. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168 (2024)

    2024 arXiv

  5. [13]

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. 2023. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725 (2023)

    2023 arXiv

  6. [14]

    Yue Han, Junwei Zhu, Keke He, Xu Chen, Yanhao Ge, Wei Li, Xiangtai Li, Jiangn- ing Zhang, Chengjie Wang, and Yong Liu. 2024. Face-Adapter for Pre-trained Diffusion Models with Fine-Grained ID and Attribute Control. In European Con- ference on Computer Vision . Springer, 20–36

    2024

  7. [15]

    Thorsten Hempel, Ahmed A Abdelrahman, and Ayoub Al-Hamadi. 2022. 6d rotation representation for unconstrained head pose estimation. In 2022 IEEE International Conference on Image Processing (ICIP) . IEEE, 2496–2500

    2022

  8. [16]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)

    2017

  9. [17]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

    2020

  10. [18]

    Fa-Ting Hong and Dan Xu. 2023. Implicit identity representation conditioned memory compensation network for talking head video generation. InProceedings of the IEEE/CVF International Conference on Computer Vision . 23062–23072

    2023

  11. [19]

    Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. 2022. Depth-aware genera- tive adversarial network for talking head video generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 3397–3406

    2022

  12. [20]

    Huaibo Huang, Ran He, Zhenan Sun, Tieniu Tan, et al. 2018. Introvae: Introspec- tive variational autoencoders for photographic image synthesis. Advances in neural information processing systems 31 (2018)

    2018

  13. [21]

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuan- han Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. 2024. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision...

    2024

  14. [22]

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. 2017. Image-to- image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition . 1125–1134

    2017

  15. [23]

    Hyeonho Jeong and Jong Chul Ye. 2023. Ground-a-video: Zero-shot grounded video editing using text-to-image diffusion models. arXiv preprint arXiv:2310.01107 (2023)

    2023 arXiv

  16. [24]

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. 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, Proceed- ings, Part II 14 . Springer, 694–711

    2016

  17. [25]

    Tero Karras, Samuli Laine, and Timo Aila. 2019. A style-based generator ar- chitecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 4401–4410

    2019

  18. [26]

    Sasikarn Khwanmuang, Pakkapon Phongthawee, Patsorn Sangkloy, and Supasorn Suwajanakorn. 2023. StyleGAN Salon: multi-view latent optimization for pose- invariant hairstyle transfer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 8609–8618

    2023

  19. [27]

    Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. 2022. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. In European Conference on Computer Vision. Springer, 188–203

    2022

  20. [28]

    Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. 2022. Implicit warping for animation with image sets. Advances in Neural Information Processing Systems 35 (2022), 22438–22450

    2022

  21. [29]

    Maxim Nikolaev, Mikhail Kuznetsov, Dmitry P Vetrov, and Aibek Alanov. 2024. Hairfastgan: Realistic and robust hair transfer with a fast encoder-based approach. Advances in Neural Information Processing Systems 37 (2024), 45600–45635

    2024

  22. [30]

    Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. 2019. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 2337–2346

    2019

  23. [31]

    Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. 2023. Fatezero: Fusing attentions for zero-shot text-based video editing. InProceedings of the IEEE/CVF International Conference on Computer Vision. 15932–15942

    2023

  24. [32]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  25. [33]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  26. [34]

    Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. 2021. Loho: Latent optimization of hairstyles via orthogonalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1984–1993

    2021

  27. [35]

    Hamid R Sheikh, Muhammad F Sabir, and Alan C Bovik. 2006. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on image processing 15, 11 (2006), 3440–3451

    2006

  28. [36]

    Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. 2019. First order motion model for image animation. Advances in neural information processing systems 32 (2019)

    2019

  29. [37]

    Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    2014 arXiv

  30. [38]

    Zhentao Tan, Menglei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. 2020. Michigan: multi-input-conditioned hair image generation for portrait editing. arXiv preprint arXiv:2010.16417 (2020)

    2020 arXiv

  31. [39]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    2023 arXiv

  32. [40]

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. 2018. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717 (2018)

    2018 arXiv

  33. [41]

    Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. 2021. One-shot free-view neu- ral talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10039–10049

    2021

  34. [42]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing 13, 4 (2004), 600–612

    2004

  35. [43]

    Huawei Wei, Zejun Yang, and Zhisheng Wang. 2024. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694 (2024). MM ’25, October 27–31, 2025, Dublin, Ireland Wangzheng Shi, Yinglin Zheng et al

    2024 arXiv

  36. [44]

    Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. 2022. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 18072–18081

    2022

  37. [45]

    Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Gang Hua, and Nenghai Yu. 2023. Hairclipv2: Unifying hair editing via proxy feature blending. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 23589–23599

    2023

  38. [46]

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. 2023. Tune-a- video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Con...

    2023

  39. [47]

    Yiqian Wu, Yong-Liang Yang, and Xiaogang Jin. 2022. Hairmapper: Removing hair from portraits using gans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 4227–4236

    2022

  40. [48]

    You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. 2024. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers . 1–11

    2024

  41. [49]

    Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. 2024. Easyanimate: A high-performance long video genera- tion method based on transformer architecture. arXiv preprint arXiv:2405.18991 (2024)

    2024

  42. [50]

    Shiyuan Yang, Xiaodong Chen, and Jing Liao. 2023. Uni-paint: A unified frame- work for multimodal image inpainting with pretrained diffusion model. In Pro- ceedings of the 31st ACM International Conference on Multimedia . 3190–3199

    2023

  43. [51]

    Shurong Yang, Huadong Li, Juhao Wu, Minhao Jing, Linze Li, Renhe Ji, Jiajun Liang, and Haoqiang Fan. 2024. Megactor: Harness the power of raw video for vivid portrait animation. arXiv preprint arXiv:2405.20851 (2024)

    2024 arXiv

  44. [52]

    Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. 2023. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers. 1–11

    2023

  45. [53]

    Xiangpeng Yang, Linchao Zhu, Hehe Fan, and Yi Yang. 2025. VideoGrain: Mod- ulating Space-Time Attention for Multi-grained Video Editing. arXiv preprint arXiv:2502.17258 (2025)

    2025 arXiv

  46. [54]

    Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. 2023. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2023

  47. [55]

    Bohan Zeng, Xuhui Liu, Sicheng Gao, Boyu Liu, Hong Li, Jianzhuang Liu, and Baochang Zhang. 2023. Face animation with an attribute-guided diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. 628–637

    2023

  48. [56]

    Yu Zeng, Yang Zhang, Liu Jiachen, Linlin Shen, Kaijun Deng, Weizhao He, and Jinbao Wang. 2024. HairDiffusion: Vivid Multi-Colored Hair Editing via Latent Diffusion. Advances in Neural Information Processing Systems 37 (2024), 5048– 5073

    2024

  49. [57]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional con- trol to text-to-image diffusion models. InProceedings of the IEEE/CVF international conference on computer vision . 3836–3847

    2023

  50. [58]

    Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. 2023. Controlvideo: Training-free controllable text-to-video genera- tion. arXiv preprint arXiv:2305.13077 (2023)

    2023 arXiv

  51. [59]

    Yuxuan Zhang, Qing Zhang, Yiren Song, Jichao Zhang, Hao Tang, and Jiaming Liu. 2024. Stable-hair: Real-world hair transfer via diffusion model.arXiv preprint arXiv:2407.14078

    2024 arXiv

  52. [60]

    Jian Zhao and Hui Zhang. 2022. Thin-plate spline motion model for image animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 3657–3666

    2022

  53. [61]

    Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. 2022. General facial rep- resentation learning in a visual-linguistic manner. In Proceedings of the IEEE/CVF conference on computer vision and pattern re...

    2022

  54. [62]

    Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. 2022. CelebV-HQ: A large-scale video facial attributes dataset. In European conference on computer vision . Springer, 650–667

    2022

  55. [63]

    Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. 2021. Barber- shop: Gan-based image compositing using segmentation masks. arXiv preprint arXiv:2106.01505 (2021)

    2021 arXiv

  56. [64]

    Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. 2024. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. In European Conference on Computer Vision. Springer, 195–211. HairShifter: Consistent and High-Fidelity Vid...

    2024

  57. [66]

    Gender - Clearly specify whether the person appears to be male, female, or another gender identity

  58. [67]

    Hair Color - Describe the hair color, such as black, blonde, brown, red, etc

  59. [68]

    Our fine- tuned IHT demonstrates improved robustness, generating a hairstyle more consistent with the input posV Ce compared to the baseline

    Hair Shape - Specify the shape or style of the Stable-Hair IHT Ref Hair Input Images Figure 7: Comparison of our adapted Image Hair Trans- fer (IHT) module with the baseline Stable-Hair. Our fine- tuned IHT demonstrates improved robustness, generating a hairstyle more consiste...

  60. [69]

    Avoid mentioning the background, environment, or any elements unrelated to the person’s facial features

    Hair Structure - Describe the hair texture and structure, including volume, layering, smoothness, or other defining characteristics. Avoid mentioning the background, environment, or any elements unrelated to the person’s facial features. Focus solely on the person’s facial and...

    2007

  61. [2024]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 7621–7630

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.