Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

GeoAvatar: Adaptive Geometrical Gaussian Splatting for 3D Head Avatar

T0 review · 2 major / 6 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 GeoAvatar claims that head avatar quality is limited by uniform Gaussian regularization and shows that region-adaptive regularization, an anatomically extended mouth mesh, and part-wise deformation achieve state-of-the-art reconstruction…

desk verdict Solid engineering paper with consistent reconstruction gains; the unvalidated T=0 inference shortcut is the main weakness in the animation claims. read the letter →

arxiv 2507.18155 v1 pith:EHZC5Z2V submitted 2025-07-24 cs.GR cs.CVcs.LG

SeungJun Moon , Hah Min Lew , Seungeun Lee , Ji-Su Kang , Gyeong-Moon Park This is my paper
classification cs.GRcs.CVcs.LG
keywords 3DheadavatarGaussiansplattingFLAMErigidandflexiblesetspart-wisedeformationmouthmodelingmonocularvideoDynamicFace
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

GeoAvatar argues that the reason monocular head avatars trade off identity preservation against animation quality is a mismatch between the 3D face prior and local geometry: some facial regions, such as the forehead and lips, fit the FLAME mesh tightly, while others, such as scalp, ears, and hair, deviate strongly, yet existing methods regularize all Gaussians uniformly. The paper proposes to classify Gaussians per facial part as rigid or flexible before training, using an unsupervised Adaptive Pre-allocation Stage, and to apply tighter radial constraints to rigid parts and looser constraints to flexible parts. It adds an anatomically extended mouth mesh, covering frontal and molar teeth, palate, and floor, with part-wise upper and lower deformation networks that move each mouth half as a unit. A new angular regularization term is introduced to keep far-travelling Gaussians assigned to their own triangles. The paper reports state-of-the-art reconstruction and animation results on monocular benchmarks and on multi-view data using a single training view, and releases a new expressive monocular dataset, DynamicFace.

What carries the argument

The load-bearing object is the Adaptive Pre-allocation Stage (APS), a preprocessing criterion that turns a FLAME mesh into three Gaussian sets: rigid $\mathcal{F}_r$, flexible $\mathcal{F}_f$, and mouth $\mathcal{F}_m$. After a warm-up, it computes, for each of $n$ facial parts, the average distance $\mathrm{Distance}(k) = \frac{1}{|p_k|}\sum_{f_i \in p_k}\frac{1}{|G_i|}\sum_{j \in G_i}\|\mu_{i,j}\|$ and thresholds parts by the mean of these distances, assigning well-fitted parts to the rigid set and poorly fitted parts to the flexible set. The mouth set is produced by a geometric construction: the FLAME lip-ring teeth trajectory is treated as a circular arc, reflected across a computed pseudo-center to extend it into molars, then shifted back to form palate and floor; two deformation networks $\Phi_{\mathrm{upper}}$ and $\Phi_{\mathrm{lower}}$ apply one offset per half-mouth. The regularization scheme carries the rigging argument: radial losses $\mathcal{L}_p(r) = \mathrm{ReLU}(r - \tau_p)$ with set-specific thresholds $\tau_r = 0.1$, $\tau_f = 2.0$, and $\tau_m = \tau_r$, plus the angular term $\mathcal{L}_{\mathrm{angle}}(\phi)$ that acts only when $r > \tau_r$. APS thus allocates flexibility where FLAME is wrong, the mouth mesh supplies geometry FLAME lacks, part-wise deformation keeps mouth halves coherent, and the angular loss attempts precise Gaussian-to-triangle rigging.

What would settle it

For a trained GeoAvatar model, compute, for every Gaussian with $r > \tau_r$ and $\phi < \tau_\phi$, which mesh triangle is nearest to its position; if a non-negligible fraction are closer to a neighboring triangle than to their rigged face, the angular loss is not delivering the rigging it claims. A cleaner experiment is to train the same model with regularization extended to the full 3D direction, constraining both $\phi$ and $\theta$: if the extra azimuthal term changes neither quantitative results nor animation artifacts, the phi-only loss is not the active mechanism.

Watch

Extended reading notes

Core claim

The central claim is that a single, region-independent regularization strategy is the bottleneck in Gaussian-splatting head avatars: it either lets Gaussians drift too far from the rigged mesh in regions where the mesh is accurate, causing animation artifacts, or pins them down in regions where the mesh is wrong, blurring identity details. GeoAvatar's solution is geometric adaptation: APS measures the mean local offset of Gaussians per FLAME part and assigns parts with small offsets to the rigid set, with tight $\tau_r = 0.1$ radial regularization, and parts with large offsets to the flexible set, with loose $\tau_f = 2.0$, so each region gets the freedom it needs. For the mouth, where FLAME has no geometry, the mesh is extended to molars, palate, and floor, and deformation is applied per part rather than per Gaussian, preserving structural consistency between the upper teeth and palate and the lower teeth and floor. Rigging is enforced by radial losses per set plus the angular loss $L_{\mathrm{angle}}(\phi) = \mathbf{1}_{r > \tau_r}\mathrm{ReLU}(\phi - \tau_\phi)$ with $\tau_\phi = 45^\circ$, intended to keep high-radius Gaussians closer to their own triangle than to a neighbour. With these components, the paper reports the best self-reenactment numbers on SplattingAvatar and DynamicFace, with PSNR 32.635 and 32.760 respectively, and, trained on one view, comparable PSNR and SSIM and better LPIPS than a 16-view baseline on NeRSemble.

Load-bearing premise

The whole rigging scheme depends on the assumption that limiting only one angular component of each Gaussian's offset, the polar angle, keeps the Gaussian attached to its own mesh triangle; the other angular component is unconstrained, so a Gaussian could still drift toward a neighboring region.

Editorial extensions

If this is right

  • If GeoAvatar is right, region-adaptive offset regularization should replace uniform thresholds in Gaussian-splatting avatars: rigid parts get tight rigging while flexible parts keep identity details.
  • Modeling mouth parts as coherent units means upper teeth with palate and lower teeth with floor move together, which should reduce artifacts in extreme expressions such as shouts and laughs.
  • A few minutes of monocular video can produce avatars whose perceptual quality, measured by LPIPS, matches or beats a 16-view baseline, provided the interior mouth geometry is modeled explicitly.
  • The DynamicFace benchmark invites comparing how methods degrade on highly expressive facial motions rather than only neutral speech, and the paper's ablation shows each component adds a measurable gain in LPIPS.
  • The proposed regularization and mesh modification are separable components whose gains stack, so each can be adopted independently by other avatar systems.

Reading between the lines

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

  • Beyond the paper: APS classifies entire FLAME parts at once using the mean of per-part distances, so a part containing both well-fitted and poorly fitted triangles is treated wholesale; a per-triangle refinement is a natural untested extension.
  • The phi-only angular loss suggests a testable fix: regularize the full 3D direction of the offset. If the claimed rigging guarantee is the real source of the animation gains, such a loss should further reduce artifacts without hurting reconstruction; if not, the gains come from APS and the mouth mesh.
  • Part-wise deformation is a broader design idea: any facial structure that moves as a unit, such as the tongue, lips, or eyelids, could be deformed with shared offsets instead of per-Gaussian offsets, a direction the paper leaves implicit.
  • DynamicFace could support cross-dataset generalization studies: train on DynamicFace and reenact onto SplattingAvatar subjects, which the paper's cross-dataset reenactment examples suggest is feasible.
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

2 major / 6 minor

Summary. GeoAvatar is a monocular-video 3D head avatar method built on Gaussian Splatting with FLAME rigging. The paper makes three main proposals: an Adaptive Pre-allocation Stage (APS) that divides FLAME faces into rigid and flexible sets based on the part-wise mean of Gaussian local offsets; an anatomically extended mouth mesh (molars, palate, floor) with a part-wise deformation network that deforms upper and lower mouth parts with separate offsets; and a regularization loss combining a radial position loss with an angular loss L_angle(φ) intended to keep each Gaussian closer to its rigged triangle than to neighbors. The paper also introduces DynamicFace, a dataset of 10 monocular videos with expressive facial motions. The method is evaluated on the SplattingAvatar dataset, DynamicFace, and NeRSemble (multi-view), with self-reenactment, cross-reenactment, and novel-view synthesis comparisons against six baselines. Reported results show consistent improvements, e.g., PSNR 32.635 vs 29.686 for the strongest single-view baseline on SplattingAvatar.

Significance. If the results are reproducible, GeoAvatar is a meaningful step forward in balancing reconstruction fidelity and animation robustness for head avatars. The APS idea of spatially varying offset regularization is well motivated by the observed distributions of local means, the anatomical mouth modeling is a practical improvement over prior frontal-teeth-only meshes, and the released DynamicFace dataset with expressive motions is a useful community resource. The empirical evaluation is fairly comprehensive, with three datasets, six baselines, module ablations, and per-subject tables in the appendix. However, two load-bearing points need strengthening: the timestep-conditioned mouth deformation is used at inference with an unvalidated T=0 shortcut, and the angular loss's theoretical justification ignores the azimuthal angle. Both concerns are addressable with additional experiments and analysis.

major comments (2)
  1. [§3.2, Eq. (6)] The mouth deformation network Φ is trained with timestep T as an input, yet at inference T is set to 0 with the one-sentence justification that FLAME correction is not needed. This is a train/test mismatch that is never validated: the self-reenactment evaluation in Tables 1, 2, and 4 uses the last 350 frames of the same video as test frames, so a T-conditioned network can memorize or interpolate per-frame corrections rather than generalizing in expression/pose space; in cross-reenactment, driving FLAME parameters come from a different actor and are combined with T=0, a pair never seen during training. No ablation in Section 4.4 or the appendix varies T, removes the T input, or measures sensitivity to T. The animation gains could therefore be overstated if Φ relies on T as a frame-index memory. Please provide an experiment that trains without T (or with T disabled at test time across a range of values) and reports the same metrics.
  2. [§3.3, Eq. (8), Figure 5] L_angle regularizes only the polar angle φ and not the azimuthal angle θ, even though both are defined in the same local coordinate system. The 2D motivation in Figure 5 does not extend to 3D: the distance from a Gaussian to a neighboring triangle depends on θ and on the triangle geometry, so the claim that φ < τ_φ keeps the Gaussian closer to its rigged triangle than to neighbors is not established. Moreover, Table 3 shows that adding L_angle (E vs D) improves SSIM and LPIPS but slightly degrades MSE and PSNR (0.748 vs 0.733, 32.697 vs 32.751), so the empirical support for the 'precise rigging' benefit is mixed. Please provide a 3D argument or measure directly whether the loss decreases the fraction of Gaussians closer to a non-rigged triangle.
minor comments (6)
  1. [Table 1] The header 'MonoGaussainAvatar' is a typo for 'MonoGaussianAvatar'; also 'GaussianAvatars0' is used in the table and text but not defined at first use in Section 4.2.
  2. [§3, polar-coordinate definition] The polar-coordinate definition states that θ and φ are the angles between (x,y,z) and the positive x-axis and z-axis, respectively; this is not the standard spherical coordinate system, and the relation of φ to Eq. (8) should be stated explicitly.
  3. [Table 2] The 'Gap (%)' column has no formula, and the sign convention appears inconsistent across metrics (a positive gap is good for SSIM/LPIPS but bad for MSE/PSNR in the usual interpretation). Please report the formula used.
  4. [§3.1] The rigid/flexible assignment rule uses strict inequalities (Distance(k) < τ_part and Distance(k) > τ_part); the equality case is unspecified.
  5. [§4.4, Table 3] The text says adding L_angle 'yielded more accurate results', but the MSE and PSNR of configuration E are slightly worse than those of D; this trade-off should be acknowledged.
  6. [Figure 16 caption] The caption contains a typo: 'SplattinAvatar' should be 'SplattingAvatar'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the method's components are trained on training data and evaluated on held-out frames, with no load-bearing reduction to the paper's own inputs.

full rationale

This is an empirical method paper, not a derivation paper, and no load-bearing step reduces to its own inputs by construction. The APS rigid/flexible split is computed from part-wise mean distances of Gaussians after a warm-up phase (Eq. 3), and the resulting fixed thresholds tau_r and tau_f are then used in the regularization loss; the split is a training-time module, and all reported metrics use held-out test frames, so no fitted quantity is renamed as a prediction. The mouth structure is constructed explicitly from FLAME lip rings by circular reflection, and the part-wise deformation network in Eq. (6) is a learned module whose inference-time choice T=0 is a train/test distribution-shift concern, not a circular reduction. Equation (8) constraining only phi and not theta is an effectiveness limitation of the regularization, not a circular step. The only self-citation in the paper, reference [26] for CHANGER in the application demos, is not load-bearing for the central reconstruction or animation claims. No uniqueness theorem or prior-work ansatz is imported from the authors' own work to force the method's choices. The ablations in Tables 3, 5, and 6 isolate each proposed component against a common baseline and against held-out data, which is consistent with a self-contained empirical evaluation. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities; the mouth structure is a mesh modification, not a physical object. The main reliance is on hand-set thresholds and heuristics for region splitting and mouth mesh construction, which are the key free assumptions.

free parameters (4)
  • tau_r = 0.1
    Threshold for rigid-set radial regularization in Eq. (7); hand-set in Section 3.3 and Table 9.
  • tau_f = 2.0
    Threshold for flexible-set radial regularization in Eq. (7); hand-set in Section 3.3 and Table 9.
  • tau_phi = 0.78 rad (45 degrees)
    Threshold for angular regularization in Eq. (8); hand-set in Section 3.3 and Table 9.
  • N (iterations before APS) = 100000
    Number of training steps before computing part-wise distances; set in Table 9.
assumptions (5)
  • domain assumption The part-wise mean of Gaussian local mean norms, Distance(k), is a reliable indicator of whether a facial part is geometrically well-fitted by FLAME (rigid) or poorly fitted (flexible).
    APS classification in Section 3.1 relies on this heuristic; the paper shows qualitative examples but no quantitative validation of the classification.
  • ad hoc to paper The FLAME local coordinate frame is such that the radial distance r and polar angle phi are sufficient to ensure a Gaussian remains closer to its rigged triangle than to neighbors; the azimuthal angle theta can be ignored.
    The angular loss in Eq. (8) only regularizes phi, while theta is defined but unused; Figure 5 illustrates only a 2D case.
  • ad hoc to paper At inference, the mouth deformation network can be evaluated with a fixed timestep T=0 and still produce correct mouth geometry, despite being trained with varying T.
    Section 3.2 states 'we simply set T = 0' without supporting evidence; this causes a train-inference input distribution shift.
  • domain assumption The teeth trajectory of the FLAME lip ring approximates a circular arc, so reflecting vertices about a pseudo-center yields anatomically plausible molar teeth, palate, and floor.
    Appendix 8.7 constructs the mouth mesh using this geometric assumption; if false, the added mesh is a misleading prior.
  • ad hoc to paper The fixed thresholds tau_r, tau_f, tau_phi generalize across subjects and datasets.
    These values are set once in Table 9 and used for all experiments; the paper does not analyze their sensitivity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GeoAvatar: Adaptive Geometrical Gaussian Splatting for 3D Head Avatar." pith.science (2026). https://pith.science/paper/EHZC5Z2V

@misc{pith2026250718155,
  author       = {Pith},
  title        = {Pith review of: GeoAvatar: Adaptive Geometrical Gaussian Splatting for 3D Head Avatar},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHZC5Z2V}},
  note         = {Machine review of arXiv:2507.18155}
}
read the original abstract

Despite recent progress in 3D head avatar generation, balancing identity preservation, i.e., reconstruction, with novel poses and expressions, i.e., animation, remains a challenge. Existing methods struggle to adapt Gaussians to varying geometrical deviations across facial regions, resulting in suboptimal quality. To address this, we propose GeoAvatar, a framework for adaptive geometrical Gaussian Splatting. GeoAvatar leverages Adaptive Pre-allocation Stage (APS), an unsupervised method that segments Gaussians into rigid and flexible sets for adaptive offset regularization. Then, based on mouth anatomy and dynamics, we introduce a novel mouth structure and the part-wise deformation strategy to enhance the animation fidelity of the mouth. Finally, we propose a regularization loss for precise rigging between Gaussians and 3DMM faces. Moreover, we release DynamicFace, a video dataset with highly expressive facial motions. Extensive experiments show the superiority of GeoAvatar compared to state-of-the-art methods in reconstruction and novel animation scenarios.

Figures

Figures reproduced from arXiv: 2507.18155 by the authors.

Figure 1
Figure 1. Overview of GeoAvatar. We propose GeoAvatar, a novel adaptive geometrical Gaussian Splatting framework, and release a new monocular video dataset, DynamicFace. Given an input novel animation, GeoAvatar generates robust and high textured 3D head avatars. Abstract Despite recent progress in 3D head avatar generation, bal￾ancing identity preservation, i.e., reconstruction, with novel poses and expressions, i.e., animat… view at source ↗
Figure 2
Figure 2. Different distributions of the local mean per each facial region. (a) We plot the mean value of local means of Gaussians in each facial region. Indeed, regions where the FLAME mesh cannot reconstruct the ground truth geometry, e.g., scalp, neck, and ear regions with long hair, show high mean values. (b) However, existing models, e.g., SplattingAvatar and FlashAvatar, do not consider the aforementioned distribution d… view at source ↗
Figure 3
Figure 3. Structure of GeoAvatar. Our method segments 3DMM faces into three sets: rigid set, flexible set, and mouth structure set. Adaptive Pre-allocation Stage segments rigid and flexible sets in an unsupervised way, by utilizing part-wise mean distance. For the mouth structure set, we apply the deformation to its vertices part-wisely, e.g., upper and lower parts. After deformation, we place local Gaussians into global coor… view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Visualization of mouth structure. We compare our mouth structure with GaussianAvatars [40] with the widely open￾mouth frame. The structure of GaussianAvatars only covers the frontal teeth and not enough to cover the molar teeth. On the con￾trary, ours has not only the …
Figure 5
Figure 5. Figure 5: Strategy for rigging Gaussians correctly. (a) With high values of r and φ, Gaussians may be positioned far from the rigged triangle, i.e., denoted as red, increasing the likelihood of representing a different part from the one assigned to the red triangle. (b) While ma…
Figure 6
Figure 6. Figure 6: Self- and cross-reenactment and novel-view synthesis results. We compared our method with baselines on various scenarios, i.e., self- and cross-reenactment and novel-view synthesis. In self-reenactment, ours showed robust generation results while maintaining high textu…
Figure 7
Figure 7. Figure 7: Qualitative comparison on the multi-view dataset. We compared the result between ours trained on a single view, and GaussianAvatars trained on 16 views. Though we utilized only a single view, ours showed remarkably better reproduction on teeth, while improving the text…
Figure 8
Figure 8. Figure 8: Qualitative ablation results. We compared each ablation configuration qualitatively. With applying APS, i.e., B, the resolution of output improved notably. With FLAME teeth and mouth deformation, i.e., C and D, the teeth generation quality improved remarkably. Finally,…
Figure 9
Figure 9. Figure 9: Cross-reenactment and novel-view synthesis. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Additional qualitative ablation results. We show qualitative results for additional ablation models for thorough eval￾uations. Models that only utilizes the rigid set,i.e., F and H, show notably blurred results for the region where the FLAME mesh can￾not cover the gro…
Figure 11
Figure 11. Figure 11: (a) Rendered results with the mesh, (b) Distribution of [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Comparison of each mask. Background Matting yields a noisy mask, i.e., containing non-human parts, while BiSeNet yields an over-smoothed mask. We intersect two masks and obtain the final mask for training. is sitting in. On the other hand, BiSeNet can distinguish huma…
Figure 14
Figure 14. Figure 14: Comparisons on a one-shot-based baseline (GAGA [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: Additional self-reenactment synthesis results. We show additional self-reenactment results of ours, compared to baselines, on SplattingAvatar and DynamicFace datasets. Our shows high-resolution results not only on regions where FLAME geometry is accurate, e.g., eyes, …
Figure 16
Figure 16. Figure 16: Additional cross-reenactment synthesis results. We show additional cross-reenactment results of ours, compared to baselines, on SplattingAvatar and DynamicFace datasets. To evaluate models thoroughly, we utilize the source and target actors from different datasets, e.…
Figure 17
Figure 17. Figure 17: Additional novel-view synthesis results. We show additional novel-view synthesis results of ours, compared to baselines, on SplattingAvatar and DynamicFace datasets. Ours shows clean and robust results not only on the facial region, e.g., the first, third, and fifth r…
Figure 18
Figure 18. Figure 18: Examples of DynamicFace sequences. We show our DynamicFace example sequences for all subjects. Our DynamicFace has diverse highly expressive facial motions [PITH_FULL_IMAGE:figures/full_fig_p022_18.png]
Figure 19
Figure 19. Figure 19: Application examples. Our proposed GeoAvatar framework shows its versatility across multiple applications. (a) Real￾time Interactable Digital Human Demo: By integrating a large language model (LLM), text-to-speech (TTS), and NVIDIA Audio2Face (A2F) module, GeoAvatar e…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. S-Avatar: Diffusion-Guided Gaussian Head Avatars from a Single Image

    cs.CV 2026-07 conditional novelty 5.0 of 10

    A three-stage pipeline generates animatable 3D Gaussian head avatars from one image by diffusion-based splat synthesis, FLAME fitting, and inverse-distance binding with scale adaptation.

Reference graph

Works this paper leans on

68 extracted references · 57 canonical work pages · cited by 1 Pith paper

  1. [1]

    https://build.nvidia.com/nvidia/ audio2face, 2023

    Omniverse audio2face: Generate expressive facial anima- tion from just an audio source with nvidia’s deep learning ai technology. https://build.nvidia.com/nvidia/ audio2face, 2023. 16

    work page 2023

  2. [2]

    Geneavatar: Generic expression-aware volu- metric head avatar editing from a single image

    Chong Bao, Yinda Zhang, Yuan Li, Xiyu Zhang, Bang- bang Yang, Hujun Bao, Marc Pollefeys, Guofeng Zhang, and Zhaopeng Cui. Geneavatar: Generic expression-aware volu- metric head avatar editing from a single image. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8952–8963, 2024. 1

    work page 2024

  3. [3]

    Dynaslam: Tracking, mapping, and inpainting in dynamic scenes

    Berta Bescos, Jos ´e M F ´acil, Javier Civera, and Jos ´e Neira. Dynaslam: Tracking, mapping, and inpainting in dynamic scenes. IEEE Robotics and Automation Letters , 3(4):4076– 4083, 2018. 3

    work page 2018

  4. [4]

    Hiface: High-fidelity 3d face recon- struction by learning static and dynamic details

    Zenghao Chai, Tianke Zhang, Tianyu He, Xu Tan, Tadas Baltrusaitis, HsiangTao Wu, Runnan Li, Sheng Zhao, Chun Yuan, and Jiang Bian. Hiface: High-fidelity 3d face recon- struction by learning static and dynamic details. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 9087–9098, 2023. 1

    work page 2023

  5. [5]

    Efficient geometry-aware 3d generative adversarial networks

    Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123–16133, 2022

    work page 2022

  6. [6]

    Monogaus- sianavatar: Monocular gaussian point-based head avatar

    Yufan Chen, Lizhen Wang, Qijing Li, Hongjiang Xiao, Shengping Zhang, Hongxun Yao, and Yebin Liu. Monogaus- sianavatar: Monocular gaussian point-based head avatar. In ACM SIGGRAPH 2024 Conference Papers, pages 1–9, 2024. 5, 7

    work page 2024

  7. [7]

    Gaus- siantalker: Real-time high-fidelity talking head synthesis with audio-driven 3d gaussian splatting

    Kyusun Cho, Joungbin Lee, Heeji Yoon, Yeobin Hong, Jaehoon Ko, Sangjun Ahn, and Seungryong Kim. Gaus- siantalker: Real-time high-fidelity talking head synthesis with audio-driven 3d gaussian splatting. arXiv preprint arXiv:2404.16012, 2024. 1, 4

    arXiv 2024

  8. [8]

    Generalizable and ani- matable gaussian head avatar

    Xuangeng Chu and Tatsuya Harada. Generalizable and ani- matable gaussian head avatar. Advances in Neural Informa- tion Processing Systems, 37:57642–57670, 2024. 15, 16

    work page 2024

Show all 68 references
  1. [9]

    Capture, learning, and synthe- sis of 3d speaking styles

    Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthe- sis of 3d speaking styles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10101–10111, 2019. 4

    2019

  2. [10]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 12

    2019

  3. [11]

    Headgas: Real-time animatable head avatars via 3d gaus- sian splatting

    Helisa Dhamo, Yinyu Nie, Arthur Moreau, Jifei Song, Richard Shaw, Yiren Zhou, and Eduardo P ´erez-Pellitero. Headgas: Real-time animatable head avatars via 3d gaus- sian splatting. In European Conference on Computer Vision, pages 459–476. Springer, 2025. 1, 2, 3

    2025

  4. [12]

    Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray trac- ing

    Abdallah Dib, Cedric Thebault, Junghyun Ahn, Philippe- Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray trac- ing. In Proceedings of the IEEE/CVF Internati...

    2021

  5. [13]

    Efros, and Xiaolong Wang

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 20796– 20805, 2024. 15

    2024

  6. [14]

    Efros, and Xiaolong Wang

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 20796– 20805, 2024. 3

    2024

  7. [15]

    Dynamic neural radiance fields for monocular 4d facial avatar reconstruction

    Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 8649–8658, 2021. 1, 5

    2021

  8. [16]

    Morphable face models-an open framework

    Thomas Gerig, Andreas Morel-Forster, Clemens Blumer, Bernhard Egger, Marcel Luthi, Sandro Sch ¨onborn, and Thomas Vetter. Morphable face models-an open framework. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018) , pages 75–82. IEEE,

    2018

  9. [17]

    Learning neural parametric head models

    Simon Giebenhain, Tobias Kirschstein, Markos Georgopou- los, Martin R ¨unz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003–21012, 2023. 1

    2023

  10. [18]

    Neural head avatars from monocular rgb videos

    Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 1, 5

    2022

  11. [19]

    Diffrelight: Diffusion- based facial performance relighting

    Mingming He, Pascal Clausen, Ahmet Levent Tas ¸el, Li Ma, Oliver Pilarski, Wenqi Xian, Laszlo Rikker, Xueming Yu, Ryan Burgert, Ning Yu, et al. Diffrelight: Diffusion- based facial performance relighting. arXiv preprint arXiv:2410.08188, 2024. 8

    arXiv

  12. [20]

    2d gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 1

    2024

  13. [21]

    3d face re- construction from a single image using a single reference face shape

    Ira Kemelmacher-Shlizerman and Ronen Basri. 3d face re- construction from a single image using a single reference face shape. IEEE transactions on pattern analysis and ma- chine intelligence, 33(2):394–405, 2010. 3

    2010

  14. [22]

    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,

  15. [23]

    Nersemble: Multi-view radi- ance field reconstruction of human heads.ACM Transactions on Graphics (TOG), 42(4):1–14, 2023

    Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radi- ance field reconstruction of human heads.ACM Transactions on Graphics (TOG), 42(4):1–14, 2023. 1, 5

    2023

  16. [24]

    3d morphable face models and their applications

    Josef Kittler, Patrik Huber, Zhen-Hua Feng, Guosheng Hu, and William Christmas. 3d morphable face models and their applications. In Articulated Motion and Deformable Ob- jects: 9th International Conference, AMDO 2016, Palma de Mallorca, Spain, July 13-15, 2016, Proceedings 9, ...

    2016

  17. [25]

    Deblurring 3d gaussian splatting, 2024

    Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting, 2024. 3

    2024

  18. [26]

    Towards high-fidelity head blending with chroma keying for industrial applications

    Hah Min Lew, Sahng-Min Yoo, Hyunwoo Kang, and Gyeong-Moon Park. Towards high-fidelity head blending with chroma keying for industrial applications. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6188–6196. IEEE, 2025. 16

    2025

  19. [27]

    Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting

    Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Talkinggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. arXiv preprint arXiv:2404.15264, 2024. 1, 14

    2024 arXiv

  20. [28]

    Learning a model of facial shape and expression from 4d scans

    Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194–1, 2017. 1, 2, 3, 14

    2017

  21. [29]

    Generalizable one-shot 3d neural head avatar

    Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot 3d neural head avatar. Advances in Neural Information Processing Systems, 36, 2024. 1

    2024

  22. [30]

    Real-time high-resolution background matting

    Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher- Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762–8771, 2021. 13

    2021

  23. [31]

    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. InACM SIGGRAPH 2024 Con- ference Papers, pages 1–11, 2024. 8

    2024

  24. [32]

    Smpl: A skinned multi- person linear model

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi- person linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851–866. 2023. 1

    2023

  25. [33]

    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. 8

    2024 arXiv

  26. [34]

    3d gaussian blendshapes for head avatar animation

    Shengjie Ma, Yanlin Weng, Tianjia Shao, and Kun Zhou. 3d gaussian blendshapes for head avatar animation. In ACM SIGGRAPH 2024 Conference Papers, pages 1–10, 2024. 1

    2024

  27. [35]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    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. Communications of the ACM, 65(1):99–106, 2021. 1, 5, 15

    2021

  28. [36]

    Survey on 3d face reconstruction from uncalibrated images

    Araceli Morales, Gemma Piella, and Federico M Sukno. Survey on 3d face reconstruction from uncalibrated images. Computer Science Review, 40:100400, 2021. 3

    2021

  29. [37]

    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 transactions on graphics (TOG), 41(4):1–15, 2022. 1, 15

    2022

  30. [38]

    Npms: Neural parametric models for 3d de- formable shapes

    Pablo Palafox, Alja ˇz Boˇziˇc, Justus Thies, Matthias Nießner, and Angela Dai. Npms: Neural parametric models for 3d de- formable shapes. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 12695–12705,

  31. [39]

    Nerfies: Deformable neural radiance fields

    Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 3

    2021

  32. [40]

    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 Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 2...

  33. [41]

    Gaussian garments: Reconstruct- ing simulation-ready clothing with photorealistic appearance from multi-view video

    Boxiang Rong, Artur Grigorev, Wenbo Wang, Michael J Black, Bernhard Thomaszewski, Christina Tsalicoglou, and Otmar Hilliges. Gaussian garments: Reconstruct- ing simulation-ready clothing with photorealistic appearance from multi-view video. arXiv preprint arXiv:2409.08189 ,

    arXiv

  34. [42]

    Visual slam and structure from motion in dynamic environments: A survey

    Muhamad Risqi U Saputra, Andrew Markham, and Niki Trigoni. Visual slam and structure from motion in dynamic environments: A survey. ACM Computing Surveys (CSUR), 51(2):1–36, 2018. 3

    2018

  35. [43]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 4104–4113, 2016. 3, 15

    2016

  36. [44]

    Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting

    Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2024

  37. [45]

    Next3d: Gener- ative neural texture rasterization for 3d-aware head avatars

    Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Gener- ative neural texture rasterization for 3d-aware head avatars. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 20991–21002, 2023. 1

    2023

  38. [46]

    Correspondence-guided sfm-free 3d gaussian splatting for nvs

    Wei Sun, Xiaosong Zhang, Fang Wan, Yanzhao Zhou, Yuan Li, Qixiang Ye, and Jianbin Jiao. Correspondence-guided sfm-free 3d gaussian splatting for nvs. arXiv preprint arXiv:2408.08723, 2024. 15

    2024 arXiv

  39. [47]

    The interpretation of structure from mo- tion

    Shimon Ullman. The interpretation of structure from mo- tion. Proceedings of the Royal Society of London. Series B. Biological Sciences, 203(1153):405–426, 1979. 3

    1979

  40. [48]

    Attention is all you need

    A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5

    2017

  41. [49]

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

    Cong Wang, Di Kang, He-Yi Sun, Shen-Han Qian, Zi-Xuan Wang, Linchao Bao, and Song-Hai Zhang. Mega: Hybrid mesh-gaussian head avatar for high-fidelity rendering and head editing. arXiv preprint arXiv:2404.19026, 2024. 1

    2024

  42. [50]

    Gaussianhead: Impressive head avatars with learnable gaussian diffusion

    Jie Wang, Jiu-Cheng Xie, Xianyan Li, Feng Xu, Chi- Man Pun, and Hao Gao. Gaussianhead: Impressive head avatars with learnable gaussian diffusion. arXiv preprint arXiv:2312.01632, 2023. 2, 4

    2023 arXiv

  43. [51]

    Flashavatar: High-fidelity digital avatar rendering at 300fps

    Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity digital avatar rendering at 300fps. arXiv preprint arXiv:2312.02214, 2023. 1, 2, 3, 4, 5, 7, 13, 14

    2023 arXiv

  44. [52]

    Omniavatar: Geometry-guided controllable 3d head synthesis

    Hongyi Xu, Guoxian Song, Zihang Jiang, Jianfeng Zhang, Yichun Shi, Jing Liu, Wanchun Ma, Jiashi Feng, and Lin- jie Luo. Omniavatar: Geometry-guided controllable 3d head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1281...

    2023

  45. [53]

    Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels

    Yuelang Xu, Lizhen Wang, Xiaochen Zhao, Hongwen Zhang, and Yebin Liu. Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels. In ACM SIGGRAPH 2023 Conference Proceedings , pages 1–10, 2023

    2023

  46. [54]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931–1941, 2024. 4

    1931

  47. [55]

    Faghead: Fully an- imate gaussian head from monocular videos

    Yixin Xuan, Xinyang Li, Gongxin Yao, Shiwei Zhou, Donghui Sun, Xiaoxin Chen, and Yu Pan. Faghead: Fully an- imate gaussian head from monocular videos. arXiv preprint arXiv:2406.19070, 2024. 1

    2024 arXiv

  48. [56]

    Bisenet: Bilateral segmentation network for real-time semantic segmentation

    Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceed- ings of the European conference on computer vision (ECCV), pages 325–341, 2018. 13

    2018

  49. [57]

    Gavatar: Animatable 3d gaussian avatars with implicit mesh learning

    Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 896–905, 2024. 1

    2024

  50. [58]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 7

    2018

  51. [59]

    Psavatar: A point-based morphable shape model for real-time head avatar creation with 3d gaussian splatting

    Zhongyuan Zhao, Zhenyu Bao, Qing Li, Guoping Qiu, and Kanglin Liu. Psavatar: A point-based morphable shape model for real-time head avatar creation with 3d gaussian splatting. arXiv preprint arXiv:2401.12900, 2024. 1

    2024 arXiv

  52. [60]

    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. 1, 2, 3

    2024 arXiv

  53. [61]

    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. 5

    2022

  54. [62]

    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

    2023

  55. [63]

    Headstu- dio: Text to animatable head avatars with 3d gaussian splat- ting

    Zhenglin Zhou, Fan Ma, Hehe Fan, and Yi Yang. Headstu- dio: Text to animatable head avatars with 3d gaussian splat- ting. arXiv preprint arXiv:2402.06149, 2024. 1

    2024 arXiv

  56. [64]

    Towards metrical reconstruction of human faces

    Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European confer- ence on computer vision, pages 250–269. Springer, 2022. 3, 14

    2022

  57. [65]

    Instant volumetric head avatars

    Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4574–4584, 2023. 1, 5, 7, 13

    2023

  58. [67]

    ID preservation

    Appendix In the Appendix, we show additional extensive experimen- tal results. First, we show details of the quantitative com- parison of both datasets: DynamicFace and SplattingAvatar [44]. Second, we depict the qualitative results on various self- and cross-reenactment and n...

  59. [68]

    𝑣"#𝑣"$𝑣! 𝑣

    and BiSeNet [56]. Though Background Matting can distinguish the foreground objects from the background, it still contains non-human parts, e.g., chairs that the subject InputBackground MattingBiSeNetFinal Mask Figure 12. Comparison of each mask. Background Matting yields a noi...

    2023

  60. [2024]

    1, 2, 3, 4, 5, 6, 7, 8, 13, 14

Pith tools

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