Pith. sign in

REVIEW 3 major objections 5 minor 65 references

FlexiAvatar: Unified 3D Gaussian Human Avatars Under Arbitrary Body Visibility

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read FlexiAvatar claims that 3D Gaussian human avatars can be reconstructed from monocular video of any body visibility—full, upper, or head-only—by optimizing only regions the camera sees, one pipeline for all settings.

desk verdict A useful visibility-aware optimization for Gaussian avatars with broad experiments, but the diffusion-generated view pillar is under-validated in the partial-body settings that matter most. read the letter →

arxiv 2607.19100 v1 pith:WMQ2LKDO submitted 2026-07-21 cs.CV

Yihalem Yimolal Tiruneh , Muhammad Salman Ali , Uyoung Jeong , Muneeb A. Khan , MD Khalequzzaman Chowdhury Sayem , Allanur Bayramgeldiyev , Binod Bhattarai , Seungryul Baek This is my paper
classification cs.CV
keywords 3DGaussiansplattinghumanavatarreconstructionvisibility-awareoptimizationpartialbodyvisibilitymonocularvideoSMPL-Xdiffusion-basedtexturecompletionanimation
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

Existing avatar reconstruction methods tacitly assume the whole body appears on camera, so from head-only or upper-body clips they invent limbs, and those inventions blur the parts they got right. FlexiAvatar's fix is to measure, for every Gaussian surface point, how often it is actually seen across the video and to stop optimizing the points the camera never shows. Unseen areas such as the back are filled with diffusion-generated rotating views instead of being guessed by the optimizer. On full-body, upper-body, and head-only benchmarks the paper reports consistent gains—35.77 PSNR on full-body, 36.73 on upper-body, 33.04 on head-only—with the head-only case using roughly half the memory because unseen Gaussians are pruned. The practical claim is that one method can replace the separate pipelines the field currently maintains for each visibility setting.

What carries the argument

The machinery is the per-Gaussian visibility rate, computed by averaging binary rasterizer flags over all frames and thresholded automatically so that Gaussians without visual evidence are removed from the forward pass and from Laplacian regularization. This converts 'what has the camera seen?' into a per-Gaussian optimization mask, and it is what prevents unobserved regions from corrupting observed ones through the shared triplane feature field. Supporting it is a hybrid mesh–Gaussian representation anchored to an up-sampled SMPL-X body mesh, with a diffusion-based generator supplying a synthetic 360-degree rotation to supervise never-seen texture.

What would settle it

Run the same upper-body sequence with the visibility mask inverted—keep only the low-visibility Gaussians in the optimization and drop the high-visibility ones; if PSNR on the visible torso does not collapse, the claim that unseen Gaussians corrupt visible regions is falsified. Separately, compute silhouette overlap or multi-view reprojection error on the generated rotation views; if those views disagree with captured-view geometry, the unseen-region completion premise fails.

Watch

Extended reading notes

Core claim

FlexiAvatar's central claim is that hallucinated limbs are not just an aesthetic flaw—they actively corrupt the visible parts of an avatar. Every Gaussian surface point gets a visibility rate: the share of training frames in which it contributes to a rendered pixel. An automatic threshold then splits Gaussians into observed and unobserved sets, and the unobserved set is excluded from rasterization and regularization, stopping gradients from leaking through the shared triplane feature field into unseen regions. Around this mask, FlexiAvatar adds confidence-masked SMPL-X tracking, diffusion-generated 360-degree views to fill the back, and part-specific residual networks for face and hand detai

Load-bearing premise

The load-bearing assumption is that accumulated visibility rate cleanly separates 'seen' from 'never seen' Gaussians and that deleting the never-seen ones cannot corrupt seen regions through the shared triplane decoder; secondarily, the diffusion-generated 360-degree views must preserve the subject's identity and geometry well enough to supervise unseen texture.

Editorial extensions

If this is right

  • One pipeline covers full-body, upper-body, and head-only monocular inputs, so practitioners no longer need separate avatar systems per framing.
  • Visibility pruning directly cuts cost: in head-only tests the paper reports roughly 50% fewer Gaussians, about 50% less asset memory, and about 25% lower animation time.
  • Unseen regions such as the back are supervised by generated rotating views instead of being hallucinated by optimization, enabling more consistent novel-view renders.
  • The automatic threshold adapts to each sequence's body-coverage fraction, removing the need to hand-tune a visibility cutoff per dataset.
  • The largest single-component ablation drop on head-only data comes from removing confidence-masked SMPL-X tracking (31.07 to 33.04 PSNR), making reliable body fitting a prerequisite for the visibility mask's gains.

Reading between the lines

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

  • The visibility rate is effectively an evidence map; it could be reused for adaptive training schedules, model compression, or deciding which frames are safe to trust during fitting.
  • If the core mechanism is as general as argued, the same evidence-based masking should transfer to other avatar representations that share a latent decoder; swapping the triplane for a different shared field would be a direct test.
  • The generative completion step is verified only by face-recognition similarity, not by geometric fidelity, so silhouette or multi-view consistency of the generated rotation views would tell whether unseen-region texture can be fully trusted.
  • The paper's own conclusion notes that dynamically deforming clothing remains unsolved, which bounds the unified-pipeline claim to relatively rigid garments and motivates extending the visibility mask to adaptively refine non-rigid regions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes FlexiAvatar, a 3D Gaussian Splatting framework for reconstructing animatable human avatars from monocular video under arbitrary body visibility (full-body, upper-body, head-only). The method combines SMPL-X registration with confidence-weighted keypoint masking, a triplane-conditioned hybrid mesh–Gaussian representation, per-Gaussian visibility rates used to prune unseen Gaussians via an Otsu threshold, part-specific residual color MLPs for face and hands, and diffusion-generated auxiliary views intended to complete never-observed regions. Experiments are reported on NeuMan, ZJU-MoCap, WildAvatar, TalkShow, and INSTA, with state-of-the-art or competitive PSNR/SSIM/LPIPS numbers, plus ablations on NeuMan and INSTA and runtime/memory comparisons. The central claim is that restricting optimization to observed regions eliminates hallucinated geometry and texture drift and enables a single pipeline across the visibility spectrum.

Significance. If the claims hold, FlexiAvatar addresses a real and increasingly important limitation of 3DGS avatar methods: the assumption of full-body visibility during training. The unified treatment of full-body, upper-body, and head-only inputs is a worthwhile goal, and the reported results on multiple benchmarks, including in-the-wild data, are substantially stronger than existing baselines on several metrics. The paper also provides useful ablations, quantitative runtime/memory savings, and an explicit limitation statement. However, the load-bearing component for the partial-body regime — diffusion-generated auxiliary views — is validated almost entirely on full-body NeuMan, and the paper's 'consistently outperforms' and 'approximately 3%' statements are not supported by the reported tables as written. The core idea is plausible and the evidence is broad, but the current presentation overclaims and leaves the most novel regime under-tested.

major comments (3)
  1. [Sec. 3.2, Supp. B, Tab. 6, Tab. A2, Supp. E.2] The diffusion-generated auxiliary views are the least validated pillar of the partial-body pipeline. The only quantitative ablations involving them (Tab. 6 'w/o gen. data' and Tab. A2 synthetic-view loss weight) are on NeuMan, a full-body dataset where the 360° rotation is well supported by real full-body observations. For TalkShow and INSTA, Supp. B explicitly switches to a 'full-body-then-crop' strategy, but no ablation on those datasets varies the synthetic-view weight or removes generated views. The claim that down-weighting to 0.1 'prevents propagating errors to observed regions' (end of Sec. 4.4 and Supp. D) is therefore unverified exactly where the method's novelty is strongest. The identity check in Supp. E.2 (Tab. A3) is ArcFace only, which says nothing about body geometry or texture consistency. I ask for partial-body synthetic-view ablations and a body-level consistency check
  2. [Sec. 4.2, Tab. 3, Abstract] The claim that FlexiAvatar 'consistently outperforms' baselines is contradicted by the reported numbers. On Chemistry (Tab. 3), Ours has LPIPS 9.67, which is worse than GUAVA's 7.15 and also worse than ExAvatar's 10.29? No — ExAvatar is 10.29, so GUAVA is the counterexample. Similarly, on INSTA (Tab. 4) SSIM is tied with RGBAvatar at 0.953. The abstract's 'approximately 3%' average PSNR improvement is not derivable from the tables: relative improvements vary from about 0.98% on INSTA (33.04 vs 32.72 for RGBAvatar) to about 9.4% on NeuMan (35.77 vs 32.71 for Vid2AvatarPro). Please report the exact averaging procedure, specify the baseline used for each dataset, and qualify the 'consistently' language where individual metrics are not uniformly better.
  3. [Sec. 4.1, Tab. 1, Tab. 4] The NeuMan evaluation protocol is not fully comparable across methods. The paper states that baseline numbers are 'taken directly from the corresponding papers' (Sec. 4.1), while for TalkShow and INSTA results are obtained from official implementations. Test-time per-frame SMPL-X optimization for NeuMan is a protocol choice that can favor methods with a strong registration prior; it should be explicitly verified that all baselines used the same protocol before claiming consistent superiority. The lack of error bars or multiple seeds further weakens statements like 'consistently outperforms.' At minimum, report variance or paired per-subject comparisons, and clarify which baselines were re-run and which were copied.
minor comments (5)
  1. [Abstract / Project page] The project page link is given as 'here' with no URL in the text; please include the full URL.
  2. [Tab. 8] In the header, '5122' should be '512²'; the table would be easier to read with consistent units and column alignment.
  3. [Sec. 3.4, Eq. (6)] Please define v_f^i precisely as the rasterizer-provided binary visibility flag and state how it is aggregated over frames (e.g., only frames where the Gaussian is in frustum vs. all frames).
  4. [Fig. 3] The gradient magnitude visualization lacks a color scale and quantitative units; a marginal histogram would better support the qualitative claim.
  5. [Supp. A] In the sentence defining λ_reg, the symbol is used inconsistently with the main text (λsreg in Eq. (1) vs λreg in Supp. A). Please align notation.

Circularity Check

0 steps flagged · score 2.0 of 10

No constructional circularity; minor self-citations are non-load-bearing; synthetic-view validation is a robustness gap, not a circularity.

full rationale

FlexiAvatar's derivation chain does not reduce any reported result to its own inputs. The visibility-aware objective (Eq. 6, Sec. 3.4) uses rasterizer-accumulated visibility flags to compute per-Gaussian visibility and then masks low-evidence Gaussians; this is a training-time optimization heuristic, not a fitted parameter later reported as a prediction. The diffusion-generated auxiliary views (Sec. 3.2) are external priors, not a restatement of the target metric; the paper explicitly down-weights their loss (Supp. D, E.1) and evaluates on held-out real test frames (Sec. 4.1), so the PSNR/SSIM/LPIPS comparisons are not forced by the construction. The only weak link is the thin validation of synthetic views (Supp. E.2 checks only ArcFace face consistency, FC=0.7903, rather than body geometry), but that is a correctness/falsifiability concern, not circularity. Self-citations [6,13,17,44] appear in motivational or architectural context (e.g., 'residual appearance refinement module [13]') and are supported by independent ablations such as Tab. A7, so they are not load-bearing. No equation is equivalent by construction to its input, and no benchmark number is obtained by fitting the test set. Score 2 reflects only the presence of minor, non-load-bearing self-citations.

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

The central claim is empirical; the listed constants and hyperparameters are hand-set or tuned on evaluation benchmarks themselves. No formal derivation depends on them, but they can move PSNR by 1-2 dB (Supp. Tab. A2), so they materially affect reported gains. No new physical or conceptual entities are postulated; all components (Gaussians, triplanes, residual MLPs) are existing representations.

free parameters (6)
  • Keypoint confidence threshold tau = 0.4
    Hand-set in Sec. 3.1 'Based on empirical evidence'; controls SMPL-X registration masking and can change pose quality.
  • Warm-up iterations = 2000
    Hand-set before visibility pruning and residual activation; chosen to let coarse geometry stabilize.
  • Synthetic-view loss weight = 0.1
    Selected via ablation on NeuMan reported PSNR (Supp. Tab. A2); the alternative 1.0 lowers PSNR, so this choice materially affects the central quantitative claim.
  • Reconstruction loss weights lambda_L1, lambda_ssim, lambda_lpips, lambda_face, lambda_reg = 0.8, 0.2, 0.2, 0.8, 1.0
    Hand-tuned; specified in Supp. D; not derived from first principles.
  • Number of generated auxiliary frames = ~500 per subject
    Design choice in Supp. B; affects unseen-region coverage and training balance.
  • Otsu visibility threshold tau* = data-dependent (e.g., 0.490 on Biden)
    Automatic bimodal threshold (Sec. 3.4); not a manually fitted constant, but the binarization choice is a modeling assumption.
assumptions (5)
  • domain assumption Off-the-shelf SMPL-X and whole-body keypoint estimators provide reliable initialization across visibility settings.
    Sec. 3.1 relies on [53,55]; if pose estimates fail under heavy cropping, registration quality and thus Gaussian attachment degrade.
  • domain assumption MimicMotion-generated auxiliary views preserve subject identity and geometry well enough to serve as supervision.
    Sec. 3.2 uses synthetic views for unseen regions; only ArcFace face consistency is checked post hoc (Supp. E.2), with no geometric check.
  • domain assumption SAM foreground masks plus the ExAvatar masking strategy accurately isolate the person.
    Sec. 4.1; mask errors propagate into reconstruction losses and quantitative metrics.
  • domain assumption 3DGS rasterizer visibility flags faithfully reflect image-space evidence per Gaussian.
    Sec. 3.4 Eq. 6; the entire visibility-pruning mechanism depends on this signal.
  • standard math Standard projection, linear blend skinning, and Gaussian splatting equations are correct and differentiable as used.
    Background rendering math from [22,38]; not derived in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlexiAvatar: Unified 3D Gaussian Human Avatars Under Arbitrary Body Visibility." pith.science (2026). https://pith.science/paper/WMQ2LKDO

@misc{pith2026260719100,
  author       = {Pith},
  title        = {Pith review of: FlexiAvatar: Unified 3D Gaussian Human Avatars Under Arbitrary Body Visibility},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WMQ2LKDO}},
  note         = {Machine review of arXiv:2607.19100}
}
read the original abstract

Reconstructing animatable 3D human avatars from monocular video is a fundamental problem in computer vision with broad applications in AR/VR and digital content creation. Existing approaches typically couple parametric body models with neural rendering or 3D Gaussian splatting and optimize all body regions jointly from short videos, which often degrades fidelity in the visible areas. To overcome this limitation, we introduce FlexiAvatar, a unified framework that explicitly optimizes only the visible body regions, effectively eliminating artifacts arising from unobserved limbs. Our method integrates occlusion-robust SMPL-X tracking with part-specific residual refinement to capture high-frequency geometric and appearance details. To complete entirely unseen regions (e.g., back views), we leverage a diffusion-based approach to generate texture consistent with the observed appearance. Experiments on full-body (NeuMan, ZJU-MoCap, WildAvatar), upper/half-body (talk-show clips), and head-only (INSTA) inputs show that FlexiAvatar delivers consistently higher reconstruction quality, outperforming state-of-the-art methods by an average PSNR improvement of approximately 3% across datasets. Finally, by restricting optimization to observed regions, our method reduces the effective number of Gaussians that must be optimized and rendered, leading to reduced runtime and memory overhead in partial-visibility scenarios.

Figures

Figures reproduced from arXiv: 2607.19100 by the authors.

Figure 1
Figure 1. From monocular video under arbitrary body visibility, FlexiAvatar reconstructs animatable 3D Gaussian avatars within a single pipeline, outperforming state-of-the-art full-body, upper-body, and head-only methods. Abstract. Reconstructing animatable 3D human avatars from monocular video is a fundamental problem in computer vision with broad applications in AR/VR and digital content creation. Existing approaches typic… view at source ↗
Figure 2
Figure 2. Overview of our FlexiAvatar pipeline. From a monocular video, we perform visibility￾aware SMPL-X registration (Sec. 3.1) and generate auxiliary multi-view sequences to supervise unseen regions (Sec. 3.2). Canonical vertices are projected onto a learnable triplane whose MLPs regress static and pose-dependent Gaussian parameters (Sec. 3.3). After warm￾up, visibility-aware optimization prunes low-evidence Gaussians (Se… view at source ↗
Figure 3
Figure 3. Gradient magnitude on the shared triplane for the upper-body Oliver sequence. Unlike conventional visibility masks, vi is used to guide optimization rather than as a rendering-time filter. Since all Gaussians are decoded from the shared triplane representation T, gradients orig￾inating from observed pixels can propa￾gate to Gaussians in unobserved regions and, through the shared representation, indirectly influence … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison on the NeuMan full-body dataset. Our method recovers sharper clothing texture and more accurate hand detail [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison on the TalkShow upper-body dataset. Our method preserves facial sharpness and recovers hand detail. effectiveness of our framework in recovering detailed and expressive geometry and appearance. Full-Body Avatars on ZJU-MoCap. We further validate …
Figure 7
Figure 7. Figure 7: Ablation of the part-specific resid￾ual refinement module on the NeuMan se￾quence. FlexiAvatar recovers fine-grained facial and hand details. GT W/O SMPLX Opt. W/ SMPLX Opt [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 2 linked inside Pith

  1. [1]

    IEEE Transactions on Circuits and Systems for Video Technology (2026)

    Ali,M.S.,Zhang,C.,Cagnazzo,M.,Valenzise,G.,Tartaglione,E.,Bae,S.H.:Compression in 3d gaussian splatting: A survey of methods, trends, and future directions. IEEE Transactions on Circuits and Systems for Video Technology (2026)

    2026

  2. [2]

    In: CVPR (2019)

    Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: CVPR (2019)

    2019

  3. [3]

    In: ICML (2024)

    Chang, D., Shi, Y., Gao, Q., Xu, H., Fu, J., Song, G., Yan, Q., Zhu, Y., Yang, X., Soleymani, M.: Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In: ICML (2024)

    2024

  4. [4]

    arXiv preprint arXiv:2106.13629 (2021)

    Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021)

    arXiv 2021

  5. [5]

    In: SIGGRAPH (2024)

    Chen, Y., Wang, L., Li, Q., Xiao, H., Zhang, S., Yao, H., Liu, Y.: Monogaussianavatar: Monocular gaussian point-based head avatar. In: SIGGRAPH (2024)

    2024

  6. [6]

    IEEE Access9, 165005–165013 (2021)

    Cho, I., Shin, E., Ali, M.S., Bae, S.H.: Dynamic structured pruning with novel filter importanceandleakymaskingbasedonconvolutionandbatchnormalizationparameters. IEEE Access9, 165005–165013 (2021)

    2021

  7. [7]

    In: 3DV (2022)

    Choi,H.,Moon,G.,Armando,M.,Leroy,V.,Lee,K.M.,Rogez,G.:Mononhr:Monocular neural human renderer. In: 3DV (2022)

    2022

  8. [8]

    In: CVPR (2019)

    Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR (2019)

    2019

Show all 65 references
  1. [9]

    ToG (2021) 16 Y

    Feng, Y., Feng, H., Black, M.J., Bolkart, T.: Learning an animatable detailed 3d face model from in-the-wild images. ToG (2021) 16 Y. Tiruneh et al

    2021

  2. [10]

    In: 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR) (2023)

    Fidalgo, C.G., Sousa, M., Mendes, D., Dos Anjos, R.K., Medeiros, D., Singh, K., Jorge, J.: Magic: Manipulating avatars and gestures to improve remote collaboration. In: 2023 IEEE Conference Virtual Reality and 3D User Interfaces (VR) (2023)

    2023

  3. [11]

    In: CVPR (2023)

    Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: CVPR (2023)

    2023

  4. [12]

    In: CVPR (2025)

    Guo, C., Li, J., Kant, Y., Sheikh, Y., Saito, S., Cao, C.: Vid2avatar-pro: Authentic avatar from videos in the wild via universal prior. In: CVPR (2025)

    2025

  5. [13]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Haider,A.,Ali,M.S.,Qamar,M.,Khalil,T.,Kim,S.Y.,Oh,J.,Tartaglione,E.,Bae,S.H.: I-inr: iterative implicit neural representations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 4520–4528 (2026)

    2026

  6. [14]

    In: CVPR (2024)

    Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: Gaussianavatar: Towardsrealistichumanavatarmodelingfromasinglevideoviaanimatable3dgaussians. In: CVPR (2024)

    2024

  7. [15]

    In: CVPR (2024)

    Hu, S., Hu, T., Liu, Z.: Gauhuman: Articulated gaussian splatting from monocular human videos. In: CVPR (2024)

    2024

  8. [16]

    In: CVPR (2025)

    Huang, Z., Hu, S., Wang, G., Liu, T., Zang, Y., Cao, Z., Li, W., Liu, Z.: Wildavatar: Learning in-the-wild 3d avatars from the web. In: CVPR (2025)

    2025

  9. [17]

    In: IEEE Conf

    Jeong, U., Tiruneh, Y.Y., Chang, H.J., Baek, S., Kim, K.I.: Thom: Generating physically plausible hand-object meshes from text. In: IEEE Conf. Comput. Vis. Pattern Recog. Findings. pp. 3653–3664 (June 2026)

    2026

  10. [18]

    In: CVPR (2023)

    Jiang,T.,Chen,X.,Song,J.,Hilliges,O.:Instantavatar:Learningavatarsfrommonocular video in 60 seconds. In: CVPR (2023)

    2023

  11. [19]

    In: ECCV (2022)

    Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: ECCV (2022)

    2022

  12. [20]

    IEEE Access11, 34198–34209 (2023)

    Kang, J.H., Ali, M.S., Jeong, H.W., Choi, C.K., Kim, Y., Jeong, S.Y., Bae, S.H., Kim, H.Y.: A super-resolution-based feature map compression for machine-oriented video coding. IEEE Access11, 34198–34209 (2023)

    2023

  13. [21]

    In: ICCV (2023)

    Karras, J., Holynski, A., Wang, T.C., Kemelmacher-Shlizerman, I.: Dreampose: Fashion image-to-video synthesis via stable diffusion. In: ICCV (2023)

    2023

  14. [22]

    ToG (2023)

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ToG (2023)

    2023

  15. [23]

    In: ICLR (2015)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)

    2015

  16. [24]

    In: ICCV (2023)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: ICCV (2023)

    2023

  17. [25]

    Computer Graphics Forum (Proceedings of the Eurographics Symposium on Rendering) (2021)

    Kopanas, G., Philip, J., Leimkühler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. Computer Graphics Forum (Proceedings of the Eurographics Symposium on Rendering) (2021)

    2021

  18. [26]

    In: CVPR (2024)

    Lee, I., Kim, B., Joo, H.: Guess the unseen: Dynamic 3d scene reconstruction from partial 2d glimpses. In: CVPR (2024)

    2024

  19. [27]

    In: CVPR (2024)

    Lei, J., Wang, Y., Pavlakos, G., Liu, L., Daniilidis, K.: Gart: Gaussian articulated template models. In: CVPR (2024)

    2024

  20. [28]

    In: CVPR (2025)

    Li, J., Zhang, J., Bai, X., Zheng, J., Zhou, J., Gu, L.: Instag: Learning personalized 3d talking head from few-second video. In: CVPR (2025)

    2025

  21. [29]

    In: CVPR (2025)

    Li, L., Li, Y., Weng, Y., Zheng, Y., Zhou, K.: Rgbavatar: Reduced gaussian blendshapes for online modeling of head avatars. In: CVPR (2025)

    2025

  22. [30]

    ToG (2017)

    Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4d scans. ToG (2017)

    2017

  23. [31]

    ToG (2015)

    Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ToG (2015)

    2015

  24. [32]

    In: SIGGRAPH (2024) FlexiAvatar: Unified 3D Gaussian Human Avatars 17

    Ma,S.,Weng,Y.,Shao,T.,Zhou,K.:3dgaussianblendshapesforheadavataranimation. In: SIGGRAPH (2024) FlexiAvatar: Unified 3D Gaussian Human Avatars 17

    2024

  25. [33]

    In: CVPR (2021)

    Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: CVPR (2021)

    2021

  26. [34]

    In: ECCV (2020)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)

    2020

  27. [35]

    In: ECCV (2024)

    Moon, G., Shiratori, T., Saito, S.: Expressive whole-body 3d gaussian avatar. In: ECCV (2024)

    2024

  28. [36]

    Automatica (1979)

    Otsu, N., et al.: A threshold selection method from gray-level histograms. Automatica (1979)

    1979

  29. [37]

    In: NeurIPS (2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)

    2019

  30. [38]

    In: CVPR (2019)

    Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019)

    2019

  31. [39]

    In: CVPR (2021)

    Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)

    2021

  32. [40]

    In: Proceedings of the 2018 CHI conference on human factors in computing systems (2018)

    Piumsomboon, T., Lee, G.A., Hart, J.D., Ens, B., Lindeman, R.W., Thomas, B.H., Billinghurst, M.: Mini-me: An adaptive avatar for mixed reality remote collaboration. In: Proceedings of the 2018 CHI conference on human factors in computing systems (2018)

    2018

  33. [41]

    In: CVPR (2024)

    Qian, S., Kirschstein, T., Schoneveld, L., Davoli, D., Giebenhain, S., Nießner, M.: Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. In: CVPR (2024)

    2024

  34. [42]

    In: CVPR (2024)

    Qian,Z.,Wang,S.,Mihajlovic,M.,Geiger,A.,Tang,S.:3dgs-avatar:Animatableavatars via deformable 3d gaussian splatting. In: CVPR (2024)

    2024

  35. [43]

    In: CVPR (2025)

    Qiu, L., Zhu, S., Zuo, Q., Gu, X., Dong, Y., Zhang, J., Xu, C., Li, Z., Yuan, W., Bo, L., et al.: Anigs: Animatable gaussian avatar from a single image with inconsistent gaussian reconstruction. In: CVPR (2025)

    2025

  36. [44]

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

    Sayem, M., Chowdhury, M.T., Tiruneh, Y.Y., Khan, M.A., Ali, M.S., Bhattarai, B., Baek, S.: Handvqa: Diagnosing and improving fine-grained spatial reasoning about hands in vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2026

  37. [45]

    In: CVPR (2024)

    Shao, Z., Wang, Z., Li, Z., Wang, D., Lin, X., Zhang, Y., Fan, M., Wang, Z.: Splattin- gAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In: CVPR (2024)

    2024

  38. [46]

    In: CVPR (2023)

    Shen, K., Guo, C., Kaufmann, M., Zarate, J.J., Valentin, J., Song, J., Hilliges, O.: X-avatar: Expressive human avatars. In: CVPR (2023)

    2023

  39. [47]

    In: ICCV (2025)

    Sim, G., Moon, G.: Persona: personalized whole-body 3d avatar with pose-driven deformations from a single image. In: ICCV (2025)

    2025

  40. [48]

    In: CVPR (2025)

    Wang, C., Kang, D., Sun, H., Qian, S., Wang, Z., Bao, L., Zhang, S.H.: Mega: Hybrid mesh-gaussian head avatar for high-fidelity rendering and head editing. In: CVPR (2025)

    2025

  41. [49]

    In: The thrity-seventh asilomar conference on signals, systems & computers, 2003 (2003)

    Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The thrity-seventh asilomar conference on signals, systems & computers, 2003 (2003)

    2003

  42. [50]

    Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf:Free-viewpointrenderingofmovingpeoplefrommonocularvideo.In:CVPR (2022)

    2022

  43. [51]

    In: CVPR (2024)

    Xiang, J., Gao, X., Guo, Y., Zhang, J.: Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In: CVPR (2024)

    2024

  44. [52]

    arXiv preprint arXiv:2401.00431 (2023)

    Xiang, T., Sun, A., Delp, S., Kozuka, K., Fei-Fei, L., Adeli, E.: Wild2avatar: Rendering humans behind occlusions. arXiv preprint arXiv:2401.00431 (2023)

    2023 arXiv

  45. [53]

    In: ICCV (2023) 18 Y

    Yang,Z.,Zeng,A.,Yuan,C.,Li,Y.:Effectivewhole-bodyposeestimationwithtwo-stages distillation. In: ICCV (2023) 18 Y. Tiruneh et al

    2023

  46. [54]

    In: CVPR (2023)

    Yi,H.,Liang,H.,Liu,Y.,Cao,Q.,Wen,Y.,Bolkart,T.,Tao,D.,Black,M.J.:Generating holistic 3d human motion from speech. In: CVPR (2023)

    2023

  47. [55]

    TPAMI (2025)

    Yin, W., Cai, Z., Wang, R., Zeng, A., Wei, C., Sun, Q., Mei, H., Wang, Y., Pang, H.E., Zhang, M., et al.: Smplest-x: Ultimate scaling for expressive human pose and shape estimation. TPAMI (2025)

    2025

  48. [56]

    In: ICCV (2025)

    Zhan, Y., Zhu, Q., Niu, M., Ma, M., Zhao, J., Zhong, Z., Sun, X., Qiao, Y., Zheng, Y.: Towards explicit exoskeleton for the reconstruction of complicated 3d human avatars. In: ICCV (2025)

    2025

  49. [57]

    In: CVPR (2025)

    Zhang, D., Liu, Y., Lin, L., Zhu, Y., Chen, K., Qin, M., Li, Y., Wang, H.: Hravatar: High-quality and relightable gaussian head avatar. In: CVPR (2025)

    2025

  50. [58]

    In: ICCV (2025)

    Zhang, D., Liu, Y., Lin, L., Zhu, Y., Li, Y., Qin, M., Li, Y., Wang, H.: Guava: General- izable upper body 3d gaussian avatar. In: ICCV (2025)

    2025

  51. [59]

    In: CVPR (2025)

    Zhang, J., Wu, Z., Liang, Z., Gong, Y., Hu, D., Yao, Y., Cao, X., Zhu, H.: Fate: Full-head gaussian avatar with textural editing from monocular video. In: CVPR (2025)

    2025

  52. [60]

    In: CVPR (2018)

    Zhang,R.,Isola,P.,Efros,A.A.,Shechtman,E.,Wang,O.:Theunreasonableeffectiveness of deep features as a perceptual metric. In: CVPR (2018)

    2018

  53. [61]

    In: ICML (2025)

    Zhang, Y., Gu, J., Wang, L.W., Wang, H., Cheng, J., Zhu, Y., Zou, F.: Mimicmotion: High-quality human motion video generation with confidence-aware pose guidance. In: ICML (2025)

    2025

  54. [62]

    In: Computer Graphics International Conference (2023)

    Zheng, W., Zhao, J., Liu, X., Pan, Y., Gan, Z., Han, H., Liu, N.: Flame-based multi-view 3d face reconstruction. In: Computer Graphics International Conference (2023)

    2023

  55. [63]

    In: ECCV (2024)

    Zhu, S., Chen, J.L., Dai, Z., Dong, Z., Xu, Y., Cao, X., Yao, Y., Zhu, H., Zhu, S.: Champ: Controllable and consistent human image animation with 3d parametric guidance. In: ECCV (2024)

    2024

  56. [64]

    In: CVPR (2025)

    Zielonka, W., Bolkart, T., Beeler, T., Thies, J.: Gaussian eigen models for human heads. In: CVPR (2025)

    2025

  57. [65]

    Zielonka, W., Bolkart, T., Thies, J.: Instant volumetric head avatars. In: CVPR (2023) FlexiA vatar: Unified 3D Gaussian Human A vatars Under Arbitrary Body Visibility Supplementary Materials Yihalem Yimolal Tiruneh1, Muhammad Salman Ali1, Uyoung Jeong1, Muneeb A. Khan1, MD Kh...

    2023

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.