Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

HairCUP: Hair Compositional Universal Prior for 3D Gaussian Avatars

T0 review · 3 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 HairCUP claims that separating face and hair into independent latent spaces, supervised by synthetic hairless images, yields a universal 3D avatar prior for seamless hairstyle transfer and few-shot personalization.

desk verdict A credible compositional Gaussian avatar prior with a genuinely new synthetic-bald-data trick, but the load-bearing assumption is unvalidated and the evaluation is narrower than the claims; worth a serious referee, not a desk reject. read the letter →

arxiv 2507.19481 v1 pith:WGQNERNS submitted 2025-07-25 cs.CV

Byungjun Kim , Shunsuke Saito , Giljoo Nam , Tomas Simon , Jason Saragih , Hanbyul Joo , Junxuan Li This is my paper
classification cs.CV
keywords 3DGaussianavatarsface-hairdisentanglementcompositionalpriorsyntheticbalddatahairstyletransferuniversalavatarmodelfew-shotpersonalizationrelightable
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 claims that the right way to build a universal prior for 3D head avatars is to model the face and hair as two separate, composable parts rather than as a single holistic entity, and that this separation is learnable if the model sees paired examples of the same person with and without hair. Because real hairless captures are impractical at scale, the authors synthesize them: they fit a bald mesh to each subject, optimize a scalp texture with a diffusion inpainting prior, and composite the rendered mesh into hair-masked training images. These synthetic pairs supervise separate identity-conditioned hypernetworks and Gaussian decoders, one for face and one for hair, and a boundary-free segmentation loss keeps the two Gaussian sets apart. If the claim holds, the result is a universal avatar prior that supports seamless hairstyle swapping between identities, few-shot personalization from a short monocular video, and relightable avatars whose face and hair can be controlled independently.

What carries the argument

The load-bearing object is a paired synthetic-bald dataset plus a compositional Gaussian decoder architecture: separate identity-conditioned hypernetworks for face and hair produce bias maps from mean albedo and geometry UV maps; separate decoders emit relightable 3D Gaussians for each component; a bald-geometry predictor supplies the scalp mesh that anchors hair Gaussians; and a boundary-free segmentation loss, relaxed near the face-hair boundary, assigns each Gaussian to one component. The paired bald images are what make the separation learnable, because they force the face Gaussians to reconstruct a hairless version of the subject instead of hiding hair appearance inside the face code.

What would settle it

Collect real hairless multi-view captures for a few subjects, run the same face-only and compositional rendering losses against those real images, and measure scalp-region L1 and LPIPS for the synthetic-bald-trained model versus a version fine-tuned on the real bald captures; if the synthetic model is worse, the diffusion-inpainted targets carry bias.

Watch

Extended reading notes

Core claim

The paper's central claim is that explicit compositionality—independent latent codes, hypernetworks, and Gaussian decoders for face and hair, with hair Gaussians anchored to a predicted bald scalp—yields a universal prior whose face and hair spaces are disentangled enough for identity-preserving transfer, whereas holistic priors trained on the same data remain entangled. The key supervision is the synthetic bald image: the face-only rendering is trained against a hairless composite of the subject, so facial Gaussians cannot encode hair-dependent appearance, while the hair model is trained on the original images and anchored to the bald mesh so it adapts to any target head shape. The paper further claims that this prior supports few-shot fine-tuning on unseen subjects from monocular captures, producing high-fidelity compositional avatars with independent face and hair control.

Load-bearing premise

The load-bearing premise is that the diffusion-generated bald images are faithful enough to serve as ground truth for how each subject looks without hair; if the hallucinated scalp color, geometry, or occlusion handling is biased, the face and hair latent spaces will remain entangled and transfer poorly across identities.

Editorial extensions

If this is right

  • Hairstyle transfer becomes a feed-forward operation: combining one identity's face hypernetwork with another's hair hypernetwork transfers the hairstyle without additional scaling or alignment, because hair Gaussians are defined relative to the predicted bald scalp.
  • A short monocular head-rotation video is enough to fine-tune the prior into a personalized relightable avatar, and the face and hair of that avatar can later be recombined with any other identity.
  • The face-only rendering path reconstructs plausible appearance under hair occlusion, effectively dehairing an avatar and hallucinating the scalp region.
  • Avatar interpolation improves: with face and hair manifolds separated, interpolating between subjects no longer morphs ears into hair as the holistic baseline does.
  • Relighting remains consistent across a transferred hairstyle because both components share the same relightable Gaussian appearance model, and the paper demonstrates relightable hairstyle transfer.

Reading between the lines

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

  • The same recipe—mask out an occluding attribute, inpaint its absence with a diffusion prior, and train a separate decoder on the paired data—should extend to glasses, beards, hats, and other headwear, as long as the attribute can be segmented and the inpainted target stays multi-view consistent.
  • A direct test the paper does not run: if the face and hair codes are truly disentangled, a classifier trained on face latent codes to predict the source identity's hairstyle should perform at chance; showing otherwise would reveal residual entanglement.
  • Because the scalp texture is optimized from a single neutral frame and a diffusion prior, factors like skin color under different expressions or lighting are extrapolated; even a handful of real hairless multi-view captures of new subjects would calibrate how much bias the synthetic targets introduce.
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

3 major / 5 minor

Summary. The paper introduces HairCUP, a universal relightable 3D Gaussian avatar prior with explicit compositionality of face and hair. Face and hair are modeled by separate identity-conditioned hypernetworks and Gaussian decoders, and both components are anchored on a predicted bald head mesh. The key training signal is a synthetic hairless image (Eq. 22) formed by compositing the original photo with a rendered bald mesh whose texture is optimized from a single neutral, fully-lit frame using a reconstruction loss and Score Distillation Sampling with a diffusion prior (Eq. 20). Training combines compositional and face-only reconstruction losses with a Gaussian segmentation loss. The paper claims seamless face/hair transfer between avatars and few-shot monocular personalization, and evaluates on 8 unseen identities against DELTA, with additional qualitative interpolation and ablation results.

Significance. If the synthetic hairless supervision is trustworthy, the explicit compositional prior is a timely and practically useful extension of URAvatar, enabling controllable hairstyle transfer and few-shot personalization with relightable 3D Gaussians. Strengths of the paper include a detailed method section, a large studio capture setup (252 pretraining / 8 held-out subjects), and genuine held-out fine-tuning evaluation. The main weakness is that the validity of the synthetic bald targets is not established, and the quantitative evaluation is too narrow to fully support the central claims; with those points addressed, the contribution would be solid.

major comments (3)
  1. [Section 3.3, Eq. (22); Section 3.4; Limitations] The face-only reconstruction loss Lface_rec is supervised by synthetic bald images Ibald, whose occluded scalp region is hallucinated by a diffusion prior through an SDS loss and optimized from a single neutral, fully-lit frame. The central disentanglement claim therefore depends on this hallucinated scalp being an accurate per-subject ground truth. The paper provides no comparison against real hairless or tight-cap captures and no quantitative identity-retention metric for hair-swapped avatars. Since the Limitations paragraph itself concedes color discrepancy and poor relighting of the occluded scalp, these are exactly the failure modes that would entangle face identity with the diffusion prior. Please validate the synthetic bald data on subjects with real haircut or cap captures, and quantify how bald-texture error propagates to face transfer and personalization quality.
  2. [Section 4.1, Table 1; Section 4.2] The quantitative evaluation for few-shot personalization compares only DELTA on 8 unseen subjects and reports average metrics without per-subject variance, error bars, or statistical significance. Since URAvatar and LUCAS are universal prior baselines using the same relightable Gaussian codec setup, a comparison with them under the same fine-tuning protocol is needed to substantiate the universal-prior claim. In addition, the headline application of hairstyle transfer is evaluated only qualitatively (Figs. 5, 11, 12); please report transfer metrics such as face identity preservation, hair fidelity (e.g., LPIPS or FID on swapped compositions), and boundary artifact measures.
  3. [Section 4.2, Figs. 7-8] The ablations for hair offset modeling and boundary-free segmentation loss are qualitative only, despite these components being central to the method. Without quantitative ablation on the same 8-subject held-out protocol, it is hard to attribute the reported improvements to the compositional prior, the segmentation loss, or the synthetic data pipeline. Please provide quantitative results for each ablated design choice, including the use of synthetic bald targets versus no such supervision.
minor comments (5)
  1. [Eq. (15)] In Eq. (15), n_f^k is written as t_b^k + delta t_f^k, which appears to be a typo; it should presumably read n_b^k + delta n_f^k, using the normal from the bald geometry map rather than the texel position.
  2. [Section 3.3, Eq. (21)] The term Icond is called a 'bald image prompt' but is described as generated by a pretrained text-to-image inpainting model; please define precisely how Icond is constructed and whether it is an image, a text prompt, or both.
  3. [Section 3.5] The sentence 'we only update the last layer of the hypernetworks' is underspecified: it should state which hypernetworks (face, hair, eyes) are updated, and what 'last layer' means for a UNet-based hypernetwork.
  4. [Supplementary B.2, Fig. 10] The main text says the bald texture is optimized from a single neutral expression frame, but the supplementary material describes an additional capture with hair tied back to reduce occlusion; please state this auxiliary capture in the main method for reproducibility.
  5. [Table 1] The numerical formatting in Table 1 contains stray spaces (e.g., '23 .775', '0 .790') and would benefit from standard formatting; more importantly, per-subject results should be reported given the small held-out set.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the face-hair disentanglement is trained on synthetic bald targets and evaluated on held-out identities, so the central claim is empirically contingent rather than true by construction.

full rationale

The paper's central claim is that a compositional universal prior can disentangle face and hair and enable identity-preserving transfer. The disentanglement is trained via an explicit face-only reconstruction loss against synthetic bald images defined by Eq. (22), Ibald = Mtilde_hair * Iorig + (1 - Mtilde_hair) * R(Mbald), plus a segmentation loss and separate face/hair decoders and hypernetworks. These are architectural inductive biases and supervised training targets, not a derivation in which the conclusion is assumed by the inputs. The face and hair latent spaces are not defined in terms of the transfer/identity-preservation claim; rather, the transfer behavior is an empirical consequence evaluated qualitatively and against the external baseline DELTA [10] on unseen identities. Self-citations to URAvatar [26] and Codec Avatar Studio [33] provide the base architecture, the relightable Gaussian representation, and the capture data; they are prior work and real data, not uniqueness theorems or ansatz smuggled in via citation, and no load-bearing argument reduces to a self-citation. The Limitations paragraph concedes color discrepancies and poor relighting of the occluded scalp; these are data-fidelity risks that could bias the learned disentanglement, but they are not circular reductions. No fitted parameter is renamed as a prediction, and no equation is equivalent to its own input by construction. The main assessment risk is external validity of the synthetic hairless supervision, not logical circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a chain of learned components: a diffusion-based bald texture generator, a bald geometry predictor, and separate identity-conditioned hypernetworks, all trained on a proprietary 260-subject capture dataset. None of these components is independently verified outside the paper; the most fragile is the synthetic hairless supervision, which is generated by the same authors' prior ecosystem and never checked against real bald captures. The free parameters listed are the hand-chosen settings that control how much the diffusion prior influences the training signal.

free parameters (6)
  • SDS loss weight lambda_sds_bald = 1e-6
    Hand-chosen in Supp. B.2 to balance diffusion prior vs reconstruction during bald texture optimization; central to synthetic scalp hallucination.
  • Reconstruction loss weight lambda_rec_bald = 1
    Set alongside SDS in Supp. B.2; controls visible face region fidelity in bald texture optimization.
  • Bald texture optimization schedule = 2500 iterations (1500 rec, 1000 SDS)
    Hand-tuned two-stage optimization in Supp. B.2; determines quality of the synthetic hairless data.
  • Boundary-free segmentation loss schedule = unspecified iteration count
    The paper says the segmentation loss is disabled near the face-hair boundary 'after a certain number of iterations' without giving the number or criterion.
  • Learning rates = 5e-4 pretrain, 1e-5 fine-tune
    Standard choices reported in Section 4; affect convergence and final quality.
  • Pretraining/fine-tuning subject split = 252 pretrain, 8 fine-tune
    Data split chosen in Section 4; evaluation is limited to 8 unseen subjects.
assumptions (5)
  • domain assumption Diffusion-inpainted synthetic bald images are a faithful proxy for real hairless captures
    Used as the face-only rendering supervision target (Eq. (22), Section 3.3); if the scalp hallucination is biased, disentanglement is compromised.
  • domain assumption Predefined UV face/hair masks correctly separate face from hair across all subjects
    The expression/hair motion encoders use Mf and Mh = 1 - Mf (Section 3.2, Eq. (4)); mask errors would leak hair into face codes.
  • domain assumption A single neutral-frame bald mesh with shared topology is a consistent anchor for hair across identities and expressions
    Gaussian positions are defined relative to the bald mesh (Eq. (14)); this enables transfer but assumes the bald geometry predictor generalizes to unseen head shapes.
  • domain assumption The diffusion prior and ControlNet trained on dome-captured human images are representative enough to inpaint occluded scalps
    SDS loss in Eq. (21) relies on this prior; the authors acknowledge possible color discrepancy but do not validate it.
  • standard math Standard Gaussian splatting and the relightable Gaussian appearance model are valid rendering assumptions
    The method inherits Eqs. (1)-(3) from Saito et al. and Kerbl et al.; no new physical assumptions are introduced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HairCUP: Hair Compositional Universal Prior for 3D Gaussian Avatars." pith.science (2026). https://pith.science/paper/WGQNERNS

@misc{pith2026250719481,
  author       = {Pith},
  title        = {Pith review of: HairCUP: Hair Compositional Universal Prior for 3D Gaussian Avatars},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGQNERNS}},
  note         = {Machine review of arXiv:2507.19481}
}
read the original abstract

We present a universal prior model for 3D head avatars with explicit hair compositionality. Existing approaches to build generalizable priors for 3D head avatars often adopt a holistic modeling approach, treating the face and hair as an inseparable entity. This overlooks the inherent compositionality of the human head, making it difficult for the model to naturally disentangle face and hair representations, especially when the dataset is limited. Furthermore, such holistic models struggle to support applications like 3D face and hairstyle swapping in a flexible and controllable manner. To address these challenges, we introduce a prior model that explicitly accounts for the compositionality of face and hair, learning their latent spaces separately. A key enabler of this approach is our synthetic hairless data creation pipeline, which removes hair from studio-captured datasets using estimated hairless geometry and texture derived from a diffusion prior. By leveraging a paired dataset of hair and hairless captures, we train disentangled prior models for face and hair, incorporating compositionality as an inductive bias to facilitate effective separation. Our model's inherent compositionality enables seamless transfer of face and hair components between avatars while preserving identity. Additionally, we demonstrate that our model can be fine-tuned in a few-shot manner using monocular captures to create high-fidelity, hair-compositional 3D head avatars for unseen subjects. These capabilities highlight the practical applicability of our approach in real-world scenarios, paving the way for flexible and expressive 3D avatar generation.

Figures

Figures reproduced from arXiv: 2507.19481 by the authors.

Figure 1
Figure 1. HairCUP. HairCUP is a compositional universal avatar model that generates relightable Gaussian codec avatars for multiple subjects using a single model. It separately models hair and face, enabling seamless hairstyle transfer between avatars without requiring additional scale optimization or suffering from unnatural hair boundary artifacts. Abstract We present a universal prior model for 3D head avatars with explici… view at source ↗
Figure 2
Figure 2. Model overview. HairCUP comprises ID-conditioned face/hair hypernetworks and a compositional avatar model. (a) The hypernetworks take UV-unwrapped mean albedo and geometry maps as input, generating multi-scale bias maps as ID conditions, added to each layer of the face/hair Gaussian decoders. (b) The compositional avatar model consists of a hair motion and face expression encoder E{f,h} and Gaussian decoders D {f,h}… view at source ↗
Figure 3
Figure 3. Synthetic bald image. We get synthetic bald images by compositing (a) the original capture with (c) the rendered bald mesh. Hair masks (b) remove the hair region, allowing the bald mesh to be rendered into (d) the final composited image. input, as it does not influence the hair state. The predicted delta translations δt {f,h} k are then added to the correspond￾ing texel t b k of the bald geometry map Gb: t {f,h} k =… view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison of hairstyle transfer. Our method achieves more natural hairstyle transfer with seamless face-hair blending, while DELTA creates a rigid boundary. Ad￾ditionally, our results offer greater visual fidelity in face and hair. exhibits limitations in …
Figure 7
Figure 7. Figure 7: Ablation on anchor geometry. Anchoring 3D Gaus￾sians to a haired mesh leads to unnatural transfers due to mis￾aligned source geometry. Using a bald mesh enables more natural hairstyle transfer by providing a consistent scalp anchor. Full segmentation loss Boundary-free…
Figure 8
Figure 8. Figure 8: Ablation study on boundary-free segmentation loss. Boundary-free segmentation loss yields smoother face-hair transi￾tions, while full-region loss creates sharp boundaries. 5. Conclusion We introduced a compositional universal avatar model that independently represents …
Figure 9
Figure 9. Figure 9: Synthetic bald image pairs. Each pair shows (left) the original image and (right) the synthetic bald image generated using our compositing pipeline. The synthetic bald images preserve facial identity while removing hair occlusion, enabling effective supervision for fac…
Figure 10
Figure 10. Figure 10: Auxiliary capture for bald texture optimization. To minimize occlusion from certain hairstyles, we capture an addi￾tional image with the subject’s hair tied back (b). This ensures that the optimized bald texture (c) maintains consistent skin color, even when the targe…
Figure 11
Figure 11. Figure 11: Hairstyle Transfer: Single Face, Multiple Hairs. This figure demonstrates transferring various hairstyles onto a single facial identity. The consistent facial features and expres￾sions highlight the model’s ability to seamlessly integrate different hairstyles while pr…
Figure 13
Figure 13. Figure 13: Relighting with hairstyle transfer. The leftmost col￾umn shows face and expression reference images captured from a real subject (Face/Exp. ID), with expression changing across frames. The second column shows the hair identity image (Hair ID) used for hair transfer. T…
Figure 14
Figure 14. Figure 14: Compositional 3D avatars of the training subjects. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: Zero-shot and Fine-tuned Compositional Avatars. Our model generates a plausible 3D avatar for a novel identity without fine-tuning (Zero-shot Avatar, middle column), reenacting the facial expression shown in the reference image (GT, left column). We visualize the comp…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Durian: Dual Reference Image-Guided Portrait Animation with Attribute Transfer

    cs.CV 2025-09 conditional novelty 7.0 of 10

    Durian introduces a dual-reference diffusion model trained via self-reconstruction on video frames to enable cross-identity attribute transfer in portrait animations, supporting multi-attribute composition and interpolation.

  2. PhysHead: Simulation-Ready Gaussian Head Avatars

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    PhysHead builds simulation-ready head avatars by layering 3D Gaussians on a head mesh and physics-simulatable hair strands, enabling wind-blown and expression-driven hair motion from video data.

Reference graph

Works this paper leans on

82 extracted references · 73 canonical work pages · cited by 2 Pith papers

  1. [1]

    Learning personal- ized high quality volumetric head avatars from monocular rgb videos

    Ziqian Bai, Feitong Tan, Zeng Huang, Kripasindhu Sarkar, Danhang Tang, Di Qiu, Abhimitra Meka, Ruofei Du, Ming- song Dou, Sergio Orts-Escolano, et al. Learning personal- ized high quality volumetric head avatars from monocular rgb videos. In Proc. CVPR, 2023. 2

    2023

  2. [2]

    Effi- cient 3d implicit head avatar with mesh-anchored hash table blendshapes

    Ziqian Bai, Feitong Tan, Sean Fanello, Rohit Pandey, Ming- song Dou, Shichen Liu, Ping Tan, and Yinda Zhang. Effi- cient 3d implicit head avatar with mesh-anchored hash table blendshapes. In Proc. CVPR, 2024. 2

    2024

  3. [3]

    A morphable model for the synthesis of 3d faces

    V olker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proc. ACM SIGGRAPH, 1999. 2

    work page 1999

  4. [4]

    Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis

    Marcel C B ¨uhler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Ser- gio Orts-Escolano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, et al. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. InProc. ICCV,

    work page

  5. [5]

    Facewarehouse: A 3d facial expression database for visual computing

    Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Trans. Vis. Comput. Graph., 20(3): 413–425, 2013. 2

    work page 2013

  6. [6]

    Authentic volumetric avatars from a phone scan

    Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shoou-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41(4), 2022. 2, 3, 12, 14

    work page 2022

  7. [7]

    Pega- sus: Personalized generative 3d avatars with composable at- tributes

    Hyunsoo Cha, Byungjun Kim, and Hanbyul Joo. Pega- sus: Personalized generative 3d avatars with composable at- tributes. In Proc. CVPR, 2024. 2, 5

    work page 2024

  8. [8]

    Perse: Person- alized 3d generative avatars from a single portrait

    Hyunsoo Cha, Inhee Lee, and Hanbyul Joo. Perse: Person- alized 3d generative avatars from a single portrait. arXiv preprint arXiv:2412.21206, 2024. 2

    arXiv 2024

Show all 82 references
  1. [9]

    Black, and Timo Bolkart

    Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J. Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In Proc. ACM SIGGRAPH Asia, 2022. 2

    2022

  2. [10]

    Learning disentangled avatars with hybrid 3d representations

    Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J Black. Learning disentangled avatars with hybrid 3d representations. arXiv preprint arXiv:2309.06441, 2023. 2, 3, 5, 7

    2023 arXiv

  3. [11]

    Dynamic neural radiance fields for monocular 4d facial avatar reconstruction

    Guy Gafni, Justus Thies, Michael Zollh ¨ofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proc. CVPR, 2021. 2

    2021

  4. [12]

    Reconstructing personalized se- mantic facial nerf models from monocular video

    Xuan Gao, Chenglai Zhong, Jun Xiang, Yang Hong, Yudong Guo, and Juyong Zhang. Reconstructing personalized se- mantic facial nerf models from monocular video. ACM Trans. Graph., 41(6):1–12, 2022. 2

    2022

  5. [13]

    Npga: Neural parametric gaussian avatars

    Simon Giebenhain, Tobias Kirschstein, Martin R ¨unz, Lour- des Agapito, and Matthias Nießner. Npga: Neural parametric gaussian avatars. In Proc. ACM SIGGRAPH Asia, 2024. 2

    2024

  6. [14]

    Neu- ral head avatars from monocular rgb videos

    Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neu- ral head avatars from monocular rgb videos. In Proc. CVPR,

  7. [15]

    Vid2avatar-pro: Authentic avatar from videos in the wild via universal prior

    Chen Guo, Junxuan Li, Yash Kant, Yaser Sheikh, Shunsuke Saito, and Chen Cao. Vid2avatar-pro: Authentic avatar from videos in the wild via universal prior. In Proc. CVPR, 2025. 2

    2025

  8. [16]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. NeurIPS, 2020. 3, 5

    2020

  9. [17]

    Single- view hair modeling using a hairstyle database

    Liwen Hu, Chongyang Ma, Linjie Luo, and Hao Li. Single- view hair modeling using a hairstyle database. ACM Trans. Graph., 34(4):1–9, 2015. 6

    2015

  10. [18]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  11. [19]

    Realistic one-shot mesh-based head avatars

    Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proc. ECCV, 2022. 2

    2022

  12. [20]

    Ncho: Un- supervised learning for neural 3d composition of humans and objects

    Taeksoo Kim, Shunsuke Saito, and Hanbyul Joo. Ncho: Un- supervised learning for neural 3d composition of humans and objects. In Proc. ICCV, 2023. 2

    2023

  13. [21]

    Gala: Generating animatable layered assets from a sin- gle scan

    Taeksoo Kim, Byungjun Kim, Shunsuke Saito, and Hanbyul Joo. Gala: Generating animatable layered assets from a sin- gle scan. In Proc. CVPR, 2024. 2, 5, 6

    2024

  14. [22]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 7

    2014 arXiv

  15. [23]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding vari- ational bayes. In Proc. ICLR, 2014. 3, 6, 12

    2014

  16. [24]

    Guess the unseen: Dynamic 3d scene reconstruction from partial 2d glimpses

    Inhee Lee, Byungjun Kim, and Hanbyul Joo. Guess the unseen: Dynamic 3d scene reconstruction from partial 2d glimpses. In Proc. CVPR, 2024. 5

    2024

  17. [25]

    Megane: Morphable eyeglass and avatar network

    Junxuan Li, Shunsuke Saito, Tomas Simon, Stephen Lom- bardi, Hongdong Li, and Jason Saragih. Megane: Morphable eyeglass and avatar network. In Proc. CVPR, 2023. 2

    2023

  18. [26]

    Uravatar: Universal relightable gaussian codec avatars

    Junxuan Li, Chen Cao, Gabriel Schwartz, Rawal Khirodkar, Christian Richardt, Tomas Simon, Yaser Sheikh, and Shun- suke Saito. Uravatar: Universal relightable gaussian codec avatars. In Proc. ACM SIGGRAPH Asia, 2024. 2, 3, 4, 5, 6, 8, 12, 14

    2024

  19. [27]

    Black, Hao Li, and Javier Romero

    Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and ex- pression from 4d scans. ACM Trans. Graph., 36(6), 2017. 2, 3

    2017

  20. [28]

    Layga: Layered gaussian avatars for animatable clothing transfer

    Siyou Lin, Zhe Li, Zhaoqi Su, Zerong Zheng, Hongwen Zhang, and Yebin Liu. Layga: Layered gaussian avatars for animatable clothing transfer. In Proc. ACM SIGGRAPH,

  21. [29]

    Lucas: Layered universal codec avatars

    Di Liu, Teng Deng, Giljoo Nam, Yu Rong, Stanislav Pid- horskyi, Junxuan Li, Jason Saragih, Dimitris N Metaxas, and Chen Cao. Lucas: Layered universal codec avatars. In Proc. CVPR, 2025. 2 9

    2025

  22. [30]

    Mix- ture of volumetric primitives for efficient neural rendering

    Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mix- ture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4):1–13, 2021. 2

    2021

  23. [31]

    Gaussianhair: Hair modeling and rendering with light-aware gaussians

    Haimin Luo, Min Ouyang, Zijun Zhao, Suyi Jiang, Longwen Zhang, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Gaussianhair: Hair modeling and rendering with light-aware gaussians. arXiv preprint arXiv:2402.10483, 2024. 3

    2024 arXiv

  24. [32]

    Pixel codec avatars

    Shugao Ma, Tomas Simon, Jason Saragih, Dawei Wang, Yuecheng Li, Fernando De La Torre, and Yaser Sheikh. Pixel codec avatars. In Proc. CVPR, 2021. 2

    2021

  25. [33]

    Jewett, Simon Ven- shtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mo- hamed Ezzeldin A

    Julieta Martinez, Emily Kim, Javier Romero, Timur Bagaut- dinov, Shunsuke Saito, Shoou-I Yu, Stuart Anderson, Michael Zollh ¨ofer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani...

    2024

  26. [34]

    Occupancy net- works: Learning 3d reconstruction in function space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Se- bastian Nowozin, and Andreas Geiger. Occupancy net- works: Learning 3d reconstruction in function space. In Proc. CVPR, 2019. 2

    2019

  27. [35]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. In Proc. ECCV, 2020. 2

    2020

  28. [36]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM Trans. Graph., 41(4):1–15,

  29. [37]

    Strand-accurate multi-view hair capture

    Giljoo Nam, Chenglei Wu, Min H Kim, and Yaser Sheikh. Strand-accurate multi-view hair capture. In Proc. CVPR ,

  30. [38]

    Rsgan: face swapping and editing using face and hair repre- sentation in latent spaces

    Ryota Natsume, Tatsuya Yatagawa, and Shigeo Morishima. Rsgan: face swapping and editing using face and hair repre- sentation in latent spaces. In Proc. ACM SIGGRAPH, 2018. 2

    2018

  31. [39]

    Capture of hair geometry from multiple images.ACM Trans

    Sylvain Paris, Hector M Briceno, and Franc ¸ois X Sillion. Capture of hair geometry from multiple images.ACM Trans. Graph., 23(3):712–719, 2004. 3

    2004

  32. [40]

    Deepsdf: Learning con- tinuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning con- tinuous signed distance functions for shape representation. In Proc. CVPR, 2019. 14

    2019

  33. [41]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proc. CVPR, 2019. 2

    2019

  34. [42]

    Rasterized edge gradients: Handling discontinuities differentiably

    Stanislav Pidhorskyi, Tomas Simon, Gabriel Schwartz, He Wen, Yaser Sheikh, and Jason Saragih. Rasterized edge gradients: Handling discontinuities differentiably. In Proc. ECCV, 2024. 5

    2024

  35. [43]

    Barron, and Ben Milden- hall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. In Proc. ICLR, 2023. 5, 12

    2023

  36. [44]

    Gaus- sianavatars: Photorealistic head avatars with rigged 3d gaus- sians

    Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaus- sianavatars: Photorealistic head avatars with rigged 3d gaus- sians. In Proc. CVPR, 2024. 2, 3, 8

    2024

  37. [45]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In Proc. CVPR, 2022. 5, 13

    2022

  38. [46]

    Ronneberger, P.Fischer, and T

    O. Ronneberger, P.Fischer, and T. Brox. U-net: Convolu- tional networks for biomedical image segmentation. InProc. MICCAI, 2015. 6

    2015

  39. [47]

    Neural strands: Learning hair geometry and appearance from multi-view im- ages

    Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chen- glei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view im- ages. In Proc. ECCV, 2022. 3

    2022

  40. [48]

    Radu Alexandru Rosu, Keyu Wu, Yao Feng, Youyi Zheng, and Michael J. Black. DiffLocks: Generating 3d hair from a single image using diffusion models. In Proc. CVPR, 2025. 3

    2025

  41. [49]

    3d hair synthesis using volumetric variational autoencoders

    Shunsuke Saito, Liwen Hu, Chongyang Ma, Hikaru Ibayashi, Linjie Luo, and Hao Li. 3d hair synthesis using volumetric variational autoencoders. ACM Trans. Graph. , 37(6):1–12, 2018. 3

    2018

  42. [50]

    Relightable gaussian codec avatars

    Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In Proc. CVPR, 2024. 2, 3, 4, 6, 8, 12, 14

    2024

  43. [51]

    Namboodiri, and Ben- jamin E Lundell

    Jack Saunders, Charlie Hewitt, Yanan Jian, Marek Kowal- ski, Tadas Baltruˇsaitis, Yiye Chen, Darren Cosker, Virginia Estellers, Nicholas Gyde, Vinay P. Namboodiri, and Ben- jamin E Lundell. GASP: Gaussian avatars with synthetic priors, 2024. 2

    2024

  44. [52]

    Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing

    Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In Proc. CVPR,

  45. [53]

    Deep- voxels: Learning persistent 3d feature embeddings

    Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deep- voxels: Learning persistent 3d feature embeddings. In Proc. CVPR, 2019. 2

    2019

  46. [54]

    Martel, Alexander W

    Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In NeurIPS, 2020. 5 10

    2020

  47. [55]

    Neural haircut: Prior-guided strand-based hair reconstruction

    Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempitsky, and Egor Zakharov. Neural haircut: Prior-guided strand-based hair reconstruction. In Proc. CVPR, 2023. 3

    2023

  48. [56]

    Black, and Justus Thies

    Vanessa Sklyarova, Egor Zakharov, Otmar Hilliges, Michael J. Black, and Justus Thies. Text-conditioned gen- erative model of 3d strand-based human hairstyles. In Proc. CVPR, 2024. 3

    2024

  49. [57]

    De- ferred neural rendering: Image synthesis using neural tex- tures

    Justus Thies, Michael Zollh ¨ofer, and Matthias Nießner. De- ferred neural rendering: Image synthesis using neural tex- tures. ACM Trans. Graph., 38(4):1–12, 2019. 2

    2019

  50. [58]

    Nonlinear 3d face morphable model

    Luan Tran and Xiaoming Liu. Nonlinear 3d face morphable model. In Proc. CVPR, 2018. 2

    2018

  51. [59]

    Mega: Hybrid mesh-gaussian head avatar for high-fidelity rendering and head editing

    Cong Wang, Di Kang, Heyi Sun, Shenhan Qian, Zixuan Wang, Linchao Bao, and Song-Hai Zhang. Mega: Hybrid mesh-gaussian head avatar for high-fidelity rendering and head editing. In Proc. CVPR, 2025. 2, 3, 6

    2025

  52. [60]

    All-frequency rendering of dynamic, spatially- varying reflectance

    Jiaping Wang, Peiran Ren, Minmin Gong, John Snyder, and Baining Guo. All-frequency rendering of dynamic, spatially- varying reflectance. In Proc. ACM SIGGRAPH Asia. 2009. 3

    2009

  53. [61]

    Neuwigs: A neural dynamic model for volumetric hair capture and animation

    Ziyan Wang, Giljoo Nam, Tuur Stuyck, Stephen Lombardi, Chen Cao, Jason Saragih, Michael Zollh ¨ofer, Jessica Hod- gins, and Christoph Lassner. Neuwigs: A neural dynamic model for volumetric hair capture and animation. In Proc. CVPR, 2023. 3

    2023

  54. [62]

    Nerfiller: Completing scenes via generative 3d inpainting

    Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. In Proc. CVPR, 2024. 5

    2024

  55. [63]

    Reconfusion: 3d reconstruction with diffusion priors

    Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proc. CVPR ,

  56. [64]

    Hairmap- per: Removing hair from portraits using gans

    Yiqian Wu, Yong-Liang Yang, and Xiaogang Jin. Hairmap- per: Removing hair from portraits using gans. In Proc. CVPR, 2022. 5

    2022

  57. [65]

    Modeling clothing as a separate layer for an animatable hu- man avatar

    Donglai Xiang, Fabian Prada, Timur Bagautdinov, Weipeng Xu, Yuan Dong, He Wen, Jessica Hodgins, and Chenglei Wu. Modeling clothing as a separate layer for an animatable hu- man avatar. ACM Trans. Graph., 40(6):1–15, 2021. 2

    2021

  58. [66]

    Flashavatar: High-fidelity head avatar with efficient gaussian embedding

    Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In Proc. CVPR, 2024. 2, 3

    2024

  59. [67]

    Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians

    Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In Proc. CVPR, 2024. 2, 3

    2024

  60. [68]

    Facescape: A large- scale high quality 3d face dataset and detailed riggable 3d face prediction

    Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: A large- scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proc. CVPR, 2020. 2

    2020

  61. [69]

    Towards practical capture of high-fidelity relightable avatars

    Haotian Yang, Mingwu Zheng, Wanquan Feng, Haibin Huang, Yu-Kun Lai, Pengfei Wan, Zhongyuan Wang, and Chongyang Ma. Towards practical capture of high-fidelity relightable avatars. In Proc. ACM SIGGRAPH Asia, 2023. 2

    2023

  62. [70]

    Vrmm: A volumetric re- lightable morphable head model

    Haotian Yang, Mingwu Zheng, Chongyang Ma, Yu-Kun Lai, Pengfei Wan, and Haibin Huang. Vrmm: A volumetric re- lightable morphable head model. In Proc. ACM SIGGRAPH,

  63. [71]

    One2avatar: Generative implicit head avatar for few-shot user adaptation, 2024

    Zhixuan Yu, Ziqian Bai, Abhimitra Meka, Feitong Tan, Qiangeng Xu, Rohit Pandey, Sean Fanello, Hyun Soo Park, and Yinda Zhang. One2avatar: Generative implicit head avatar for few-shot user adaptation, 2024. 2

    2024

  64. [72]

    Human hair recon- struction with strand-aligned 3d gaussians

    Egor Zakharov, Vanessa Sklyarova, Michael Black, Giljoo Nam, Justus Thies, and Otmar Hilliges. Human hair recon- struction with strand-aligned 3d gaussians. In Proc. ECCV,

  65. [73]

    Hao Zhang, Yao Feng, Peter Kulits, Yandong Wen, Justus Thies, and Michael J. Black. TECA: Text-Guided Genera- tion and Editing of Compositional 3D Avatars . InProc. Intl. Conf. on 3D Vision (3DV), 2024. 2, 3

    2024

  66. [74]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proc. ICCV, 2023. 5, 12

    2023

  67. [75]

    Stable-hair: Real-world hair transfer via diffusion model

    Yuxuan Zhang, Qing Zhang, Yiren Song, and Jiaming Liu. Stable-hair: Real-world hair transfer via diffusion model. arXiv preprint arXiv:2407.14078, 2024. 5

    2024 arXiv

  68. [76]

    Headgap: Few-shot 3d head avatar via generalizable gaussian priors

    Xiaozheng Zheng, Chao Wen, Zhaohu Li, Weiyi Zhang, Zhuo Su, Xu Chang, Yang Zhao, Zheng Lv, Xiaoyuan Zhang, Yongjie Zhang, et al. Headgap: Few-shot 3d head avatar via generalizable gaussian priors. arXiv preprint arXiv:2408.06019, 2024. 2, 14

    2024 arXiv

  69. [77]

    Im avatar: Implicit morphable head avatars from videos

    Yufeng Zheng, Victoria Fern ´andez Abrevaya, Marcel C B¨uhler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545–13555, 2022. 2

    2022

  70. [78]

    Pointavatar: Deformable point-based head avatars from videos

    Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057–21067, 2023. 2

    2023

  71. [79]

    Mofanerf: Morphable facial neural radiance field

    Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. InProc. ECCV, 2022. 2

    2022

  72. [80]

    Instant volumetric head avatars

    Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proc. CVPR, 2023. 2

    2023

  73. [81]

    Garbin, Alexandros Lattas, George Kopanas, Paulo Gotardo, Thabo Beeler, Justus Thies, and Timo Bolkart

    Wojciech Zielonka, Stephan J. Garbin, Alexandros Lattas, George Kopanas, Paulo Gotardo, Thabo Beeler, Justus Thies, and Timo Bolkart. Synthetic prior for few-shot drivable head avatar inversion. arXiv:2501.06903, 2025. 2

    2025 arXiv

  74. [82]

    Orig. Light- ing

    Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Trans. Vis. Comput. Graph., 8(3):223–238, 2002. 12 11 HairCUP: Hair Compositional Universal Prior for 3D Gaussian Avatars Supplementary Material A. Preliminaries: URAvatar [26] Our meth...

    2002

Pith tools

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