Pith. sign in

REVIEW 4 major objections 4 minor 49 references

AvatarBack: Back-Head Generation for Complete 3D Avatars from Front-View Images

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read AvatarBack claims that frontal-only footage can reconstruct the missing back of the head in 3D Gaussian avatars by inverting a full-head GAN to match the subject and feeding its hallucinated back views back as aligned training supervision.

desk verdict A practically useful plug-and-play back-head completion module for Gaussian avatars, but the identity-consistency claim rests on unvalidated generative pseudo-supervision and an unvalidated GPT-4o judge. read the letter →

arxiv 2508.20623 v1 pith:M456NH76 submitted 2025-08-28 cs.CV

Shiqi Xin , Xiaolin Zhang , Yanbin Liu , Peng Zhang , Caifeng Shan This is my paper
classification cs.CV
keywords 3DheadreconstructionGaussiansplattingback-headcompletionGANinversionPanopriorpseudo-supervisionanimatableavatarsGPT-4operceptualevaluation
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

AvatarBack takes 3D head avatars built from frontal-only video — the standard setup for Gaussian-splatting head reconstruction — and tries to finish the part no camera ever sees: the back of the head. Its central claim is that the missing region can be recovered by running a pretrained full-head generator through a subject-specific inversion: the avatar's own renders and the real frontal frames together tune the generator to this person, the tuned generator then produces plausible back-view images, and a learned 3D transform maps those images into the avatar's coordinate space so they become usable training data. The paper reports that this plug-in raises back-head quality scores substantially on two benchmarks while keeping — and in one base model improving — frontal fidelity, and that the finished avatars stay fully animatable. The reason this matters: it turns sparse, cheap frontal captures into complete 360-degree virtual heads without any new capture hardware.

What carries the argument

Two coupled mechanisms carry the argument. Subject-specific Generator (SSG): a hybrid GAN inversion that optimizes a latent code and the weights of the pretrained full-head PanoHead generator against real frontal frames mixed with 3DGS-rendered novel views, so the synthesized back views inherit the subject's appearance; it then samples back views over azimuths 90°–270° and sharpens them with a super-resolution network. Adaptive Spatial Alignment (ASA): a learnable 4x4 transform, decomposed into a scale vector, a rotation vector (Rodrigues' formula) and a translation vector, applied to the FLAME mesh vertices so the hallucinated back views reach pixel-level agreement with the Gaussian renderi

What would settle it

Use the full 360-degree captures in a dataset like K-hairstyle as ground truth: train only on frontal crops, then render the completed avatar at 180 degrees and compare pixel-wise (PSNR/SSIM and hair-silhouette IoU) against the subject's true back image. If the completed back-head is not systematically closer to the same person's real back view than to another person's back view — or if it is no better than a generic average head — the identity-consistency claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim: the missing back-head is a supervision problem, not a rendering problem, and the supervision can be manufactured in a closed loop. A pretrained full-head generator (PanoHead) is inverted against a hybrid set — the subject's real frontal frames plus novel views rendered by the still-incomplete avatar — and then produces back-view images (azimuths 90°–270°) carrying the subject's identity and hairstyle. Super-resolved, these become pseudo-training views. Because the generator and the avatar live in different coordinate systems, a learnable scale–rotation–translation transform on the FLAME mesh is optimized jointly, with a FLAME-parameter regularizer guarding the fron

Load-bearing premise

The load-bearing premise is that the tuned generative prior hallucinates a back of the head that actually matches the real subject's unseen hairline, hair texture, and skull shape — because every back-region training signal in the pipeline comes from those imagined images, a wrong guess would be reconstructed with full confidence.

Editorial extensions

If this is right

  • Back-head views stop collapsing: on the GPT-4o perceptual protocol the overall score rises from 6.40 to 8.20 when AvatarBack is added to GaussianAvatars and from 6.73 to 8.43 when added to SurfHead.
  • Rear-view renders move closer to the real-image distribution on K-hairstyle: FID drops from 218.34 to 146.73 (GaussianAvatars) and from 232.46 to 165.06 (SurfHead).
  • Frontal quality is not sacrificed; in the SurfHead configuration the frontal PSNR rises by 2.68 dB on novel-view synthesis and about 3 dB on self-reenactment.
  • Completed avatars remain animatable: the added back-head Gaussians are bound to the same FLAME mesh and deform correctly under expression and pose driving.
  • The framework is protocol-agnostic at the base level: the same plug-in works on both a mesh-embedded Gaussian model and a 2D-surfel head model without architecture changes.

Reading between the lines

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

  • The closed-loop recipe generalizes: any explicit 3D representation with an under-observed region (neck, ears, chin underside, occluded limbs of a body avatar) could be supervised the same way — render from the model, invert a generative prior on the hybrid views, synthesize the missing views, and align them learnably.
  • The correctness ceiling is the prior's imagination: hairstyles rare in the generator's training distribution (tight braids, afros, shaved patterns) will be hallucinated generically, and ASA can align a wrong hairline but cannot replace it. A testable extension is measuring back-head fidelity as a function of hairstyle frequency in the prior's training set.
  • The frontal gains imply the pseudo-supervision behaves as a global regularizer on the Gaussian cloud, not a local patch; varying the pseudo-image loss weight (fixed at 0.01 in the paper) and watching frontal metrics shift would quantify that effect.
  • The GPT-4o scoring protocol could be cross-validated cheaply against human pairwise preference on the same 135°/180°/225° views, since the protocol's weights (50% rear, 25% each side) are an authors' choice rather than a measured property.
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

4 major / 4 minor

Summary. The paper proposes AvatarBack, a plug-and-play framework for completing the missing back-head region in 3D Gaussian Splatting head avatars trained from frontal views. It has two main components: a Subject-specific Generator (SSG) that adapts a pretrained PanoHead GAN to the target subject through hybrid inversion using both real frontal images and 3DGS-rendered views, and then synthesizes back-view pseudo-images; and an Adaptive Spatial Alignment (ASA) module that learns a global similarity transform to align the pseudo-images with the 3DGS coordinate frame during joint optimization. Experiments on NeRSemble and K-hairstyle report frontal-view PSNR/SSIM/LPIPS, back-view FID/KID, and GPT-4o-based perceptual scores, showing improvements over GaussianAvatars and SurfHead baselines while preserving frontal fidelity.

Significance. The paper addresses a real and under-studied problem: the collapse or absence of back-head geometry in 3DGS head avatars trained with frontal-only supervision. The plug-and-play design is practical and the qualitative results in Figs. 3, 5, and 6 show clear improvements over the base pipelines. The FID/KID reductions in Table II are large and consistent, and the ablations support the usefulness of both the SSG and the FLAME regularization. If the identity-consistency claim were properly validated, this would be a useful contribution to the head-avatar community. However, the current evidence does not fully support the central claim of identity-consistent back-head reconstruction, because the back-head metrics are either distribution-level or an unvalidated LLM-based score, and no same-subject comparison is reported.

major comments (4)
  1. [§IV-B, Table II] The paper claims 'identity-consistent' back-head reconstruction, but the only back-head quantitative metrics are FID/KID computed against K-hairstyle reference images from different subjects than the reconstructed NeRSemble avatars, plus GPT-4o scores on rendered views. Neither metric measures identity fidelity: FID/KID are distribution-level, and the GPT-4o criteria in Table S2 reward general realism and consistency with the frontal view, not correctness of the unseen back-head. Because K-hairstyle contains full 360° captures, a per-subject evaluation is feasible: reconstruct avatars from front views of K-hairstyle subjects and compare back-view renderings against the same subject's held-out back views. Without such an experiment, the central claim is unsupported.
  2. [§IV-B, Eq. (10), Table I] The GPT-4o-based perceptual scoring is load-bearing for the main back-head claim, but no validation against human raters is provided. The manuscript does not report the prompt, temperature, number of runs, or variance of the scores, and the five criteria in Table S2 are qualitative. It is therefore unclear whether the reported increases (e.g., 6.40 to 8.20) reflect human-perceived quality or merely the model's preference for coherent synthetic textures. Please add a human evaluation or a GPT-4o-vs-human correlation study, and report per-subject and per-run statistics.
  3. [§III-B, Eq. (2), Fig. 2] The SSG hybrid inversion uses 3DGS-rendered views as supervision targets, and the resulting pseudo-images are then used to supervise the same 3DGS model. This closed feedback loop risks reinforcing the 3DGS model's own errors, especially in the back-head where the initial renders are collapsed. The paper describes this as mutual reinforcement but provides no ablation that isolates the contribution of the PanoHead prior from self-reinforcement. Please report an ablation where SSG is inverted using only real frontal views, and/or quantify how the back-head region changes over successive feedback iterations.
  4. [§III-C, Eqs. (4)-(9)] ASA optimizes only a global similarity transform (scale s, rotation r, translation t) applied to FLAME vertices. It cannot correct local shape or hair-volume discrepancies between the hallucinated pseudo-image and the true subject. If the prior produces a wrong hairline, hair volume, or skull shape, the global alignment will confidently fit that wrong geometry. The paper should either justify the global-alignment assumption with evidence, for example per-vertex error maps on a dataset with ground-truth back-head scans, or extend ASA to local deformations. This issue is directly relevant to the identity-consistency claim.
minor comments (4)
  1. [Abstract/Introduction] Grammatical and typographical issues: 'an 3D avatar' in the abstract; 'the later' should be 'the latter'; the author block contains 'e-mai1'; Fig. 4 caption spells 'AvtarBack'.
  2. [§IV-B, Eq. (10)] The weights in the GPT-4o score (50% rear, 25% each side) appear arbitrary. Please provide a justification or a sensitivity analysis.
  3. [§IV-A] Implementation details omit several hyperparameters needed for reproducibility: SSG optimization steps and learning rates, the number and azimuth spacing of sampled back views, and the prompt used for GPT-4o scoring.
  4. [Tables I and III] No error bars or per-subject variances are reported. Given the small number of subjects (nine), reporting per-subject results or standard deviations would substantially strengthen the quantitative claims.

Circularity Check

1 steps flagged · score 4.0 of 10

Back-head supervision is a closed loop: 3DGS renders condition the pseudo-image generator whose outputs supervise the same 3DGS; reported back-head gains partly measure self-consistency rather than independent reconstruction.

  1. fitted input called prediction [Sec. III-A (closed feedback loop); Sec. III-B Eq. (2)-(3); Sec. III-C Eq. (4)]
    "The framework introduces a 3DGS-aware closed feedback loop... an initial 3DGS head model provides geometry- and pose-consistent cues to guide plausible back-view synthesis, and the synthesized views are fed back into the 3DGS pipeline as pseudo-supervision... I_hybrid = {I_ori} ∪ {I_render} ... I_back = G_Θ∗(w∗, c_back_j) ... T∗ = arg min Σ L(R(G(T·V(φ), c_back_j), I_back_j)."

    The pseudo-back images that supervise the 3DGS back-head are produced by a generator optimized in Eq. (2) against I_hybrid, which explicitly includes rendered images from the same 3DGS model. Eq. (4) then fits the 3DGS to those pseudo-images. So the supervision for the unobserved region is a function of the very model being supervised: any 3DGS back-head artifact can be copied into the pseudo-images and reinforced. The reported back-head gains therefore certify convergence to a self-consistent fixed point between the 3DGS and the generator, not agreement with an independent observation of the subject's true back-head. The circularity is partial because the PanoHead prior and frontal real images contribute independent content; however, the identity-consistency claim is not independently ver

full rationale

The derivation chain contains one genuine dependency: the pseudo-back images used as supervision are generated by a PanoHead inversion (Eq. 2) whose hybrid training set includes rendered images from the very 3DGS model that those pseudo-images later supervise (Eq. 4). This is an explicitly acknowledged 'closed feedback loop' (Sec. III-A). Because the back-head regions are unobserved, the loop has no independent anchor for identity/geometry; the reported GPT-4o and FID/KID gains mainly show that the 3DGS converges to the generator's self-consistent output. This is a partial circularity, not a total one: the PanoHead prior is an external, pre-trained model, and the frontal real images are independent. The paper's own K-hairstyle evaluation admits the references are 'pseudo-ground-truth' (different subjects, distribution-level), and the GPT-4o criteria reward consistency with the frontal view—exactly what SSG optimizes—so no experiment validates the back-head against the same subject's real capture. These limitations are in-scope: they mean the central 'identity-consistent' claim is unverified, but they do not by themselves make the derivation equivalent to its inputs. Score 4 reflects a real but partial feedback-loop circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The load-bearing assumptions are the PanoHead generative prior's generalizability, the usefulness of 3DGS renderings for inversion, the anatomical validity of FLAME, and the validity of GPT-4o as a perceptual judge. The ASA parameters and loss weights are fitted/optimized and should be treated as free parameters.

free parameters (3)
  • ASA transformation (s, r, t) = learned during training
    A 9-dimensional similarity transform (scale, rotation, translation) is optimized to align pseudo-images with the 3DGS space. Its values are fitted and directly affect back-head supervision.
  • Loss weights lambda, lambda_FLAME = 0.01, 0.5
    Chosen by hand. The pseudo-image loss weight and FLAME regularization strength shape the balance between hallucinated supervision and geometric prior.
  • GPT-4o scoring weights = 0.5 (rear), 0.25 (side-back each), 0.2 per criterion
    Hand-set weights in the perceptual score S; they are ad hoc and not justified by human ratings.
assumptions (4)
  • domain assumption PanoHead's pretrained generative prior can produce identity-consistent back-head images for an arbitrary subject after hybrid inversion.
    The entire SSG relies on this prior to hallucinate missing geometry. Section III-B uses PanoHead as the generator without evidence that its latent space covers the diversity of real users' back-heads.
  • domain assumption Rendered 3DGS novel views provide useful geometry-aware cues for GAN inversion.
    The hybrid inversion in Eq. (2) mixes real and rendered images. This assumes that the current 3DGS reconstruction, which is initially poor at the back, still helps the inversion rather than hurting it.
  • standard math The FLAME parametric model is an accurate anatomical prior for head geometry.
    FLAME is used as the underlying mesh and for L_FLAME regularization in Eq. (5). This is a standard, accepted assumption in the field.
  • ad hoc to paper GPT-4o perceptual scores correlate with human visual judgment for back-head quality.
    The paper introduces a GPT-4o scoring protocol (Section IV-B) and uses it as the main back-head quality metric, but provides no human study or calibration to support this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AvatarBack: Back-Head Generation for Complete 3D Avatars from Front-View Images." pith.science (2026). https://pith.science/paper/M456NH76

@misc{pith2026250820623,
  author       = {Pith},
  title        = {Pith review of: AvatarBack: Back-Head Generation for Complete 3D Avatars from Front-View Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M456NH76}},
  note         = {Machine review of arXiv:2508.20623}
}
read the original abstract

Recent advances in Gaussian Splatting have significantly boosted the reconstruction of head avatars, enabling high-quality facial modeling by representing an 3D avatar as a collection of 3D Gaussians. However, existing methods predominantly rely on frontal-view images, leaving the back-head poorly constructed. This leads to geometric inconsistencies, structural blurring, and reduced realism in the rear regions, ultimately limiting the fidelity of reconstructed avatars. To address this challenge, we propose AvatarBack, a novel plug-and-play framework specifically designed to reconstruct complete and consistent 3D Gaussian avatars by explicitly modeling the missing back-head regions. AvatarBack integrates two core technical innovations,i.e., the Subject-specific Generator (SSG) and the Adaptive Spatial Alignment Strategy (ASA). The former leverages a generative prior to synthesize identity-consistent, plausible back-view pseudo-images from sparse frontal inputs, providing robust multi-view supervision. To achieve precise geometric alignment between these synthetic views and the 3D Gaussian representation, the later employs learnable transformation matrices optimized during training, effectively resolving inherent pose and coordinate discrepancies. Extensive experiments on NeRSemble and K-hairstyle datasets, evaluated using geometric, photometric, and GPT-4o-based perceptual metrics, demonstrate that AvatarBack significantly enhances back-head reconstruction quality while preserving frontal fidelity. Moreover, the reconstructed avatars maintain consistent visual realism under diverse motions and remain fully animatable.

Figures

Figures reproduced from arXiv: 2508.20623 by the authors.

Figure 1
Figure 1. Motivation for our AvatarBack. (a) Current datasets are limited to [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the AvatarBack framework. The framework leverages a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of back-head reconstructions. The first column displays ground-truth images, followed by five rendered views under azimuth [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of reconstructed frontal-views. The proposed [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: SSG improves back-head geometry, while Lflame preserves frontal facial structure and symmetry. The full model integrates both components to produce anatomically plausible and view-consistent 3D head reconstructions. reconstruction methods, i.e., GaussianAvatars [5] and…
Figure 6
Figure 6. Figure 6: Ablation comparison of mesh quality and rendered images under four [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 43 canonical work pages

  1. [1]

    Blanz and T

    V. Blanz and T. Vetter, ``A morphable model for the synthesis of 3d faces,'' in Seminal Graphics Papers: Pushing the Boundaries, 2023, vol. 2, pp. 157--164

    work page 2023

  2. [2]

    T. Li, T. Bolkart, M. J. Black, H. Li, and J. Romero, ``Learning a model of facial shape and expression from 4d scans,'' ACM Transactions on Graphics, vol. 36, no. 6, 2017

    work page 2017

  3. [3]

    Y. Feng, H. Feng, M. J. Black, and T. Bolkart, ``Learning an animatable detailed 3d face model from in-the-wild images,'' ACM Transactions on Graphics, vol. 40, no. 4, 2021

    work page 2021

  4. [4]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk \"u hler, and G. Drettakis, ``3d gaussian splatting for real-time radiance field rendering,'' ACM Transactions on Graphics, vol. 42, no. 4, 2023

    work page 2023

  5. [5]

    S. Qian, T. Kirschstein, L. Schoneveld, D. Davoli, S. Giebenhain, and M. Niessner, ``Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20\,299--20\,309

    work page 2024

  6. [6]

    Z. Shao, Z. Wang, Z. Li, D. Wang, X. Lin, Y. Zhang, M. Fan, and Z. Wang, ``Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 1606--1616

    work page 2024

  7. [7]

    Xiang, X

    J. Xiang, X. Gao, Y. Guo, and J. Zhang, ``Flashavatar: High-fidelity head avatar with efficient gaussian embedding,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 1802--1812

    work page 2024

  8. [8]

    J. Lee, T. Kang, M. Buehler, M.-J. Kim, S. Hwang, J. Hyung, H. Jang, and J. Choo, ``Surfhead: Affine rig blending for geometrically accurate 2d gaussian surfel head avatars,'' in The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

Show all 49 references
  1. [9]

    Z. Zhao, Z. Bao, Q. Li, G. Qiu, and K. Liu, ``Psavatar: A point-based shape model for real-time head avatar animation with 3d gaussian splatting,'' 2024. [Online]. Available: https://arxiv.org/abs/2401.12900

    2024 arXiv

  2. [10]

    Zhong, X

    Y. Zhong, X. Zhang, L. Liu, Y. Zhao, and Y. Wei, ``Avatarmakeup: Realistic makeup transfer for 3d animatable head avatars,'' 2025. [Online]. Available: https://arxiv.org/abs/2507.02419

    2025 arXiv

  3. [11]

    S. An, H. Xu, Y. Shi, G. Song, U. Y. Ogras, and L. Luo, ``Panohead: Geometry-aware 3d full-head synthesis in 360°,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 20\,950--20\,959

    2023

  4. [12]

    Kirschstein, S

    T. Kirschstein, S. Qian, S. Giebenhain, T. Walter, and M. Nie ner, ``Nersemble: Multi-view radiance field reconstruction of human heads,'' ACM Transactions on Graphics, vol. 42, no. 4, 2023

    2023

  5. [13]

    T. Kim, C. Chung, S. Park, G. Gu, K. Nam, W. Choe, J. Lee, and J. Choo, ``K-hairstyle: A large-scale korean hairstyle dataset for virtual hair editing and hairstyle classification,'' in IEEE International Conference on Image Processing, 2021, pp. 1299--1303

    2021

  6. [14]

    H. Cai, Y. Xiao, X. Wang, J. Li, Y. Guo, Y. Fan, S. Gao, and J. Zhang, ``Hera: Hybrid explicit representation for ultra-realistic head avatars,'' in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2025, pp. 260--270

    2025

  7. [15]

    H. Sun, C. Wang, T.-X. Xu, J. Huang, D. Kang, C. Guo, and S.-H. Zhang, ``Svg-head: Hybrid surface-volumetric gaussians for high-fidelity head reconstruction and real-time editing,'' 2025. [Online]. Available: https://arxiv.org/abs/2508.09597

    2025 arXiv

  8. [16]

    C. Wang, D. Kang, H. Sun, S. Qian, Z. Wang, L. Bao, and S.-H. Zhang, ``Mega: Hybrid mesh-gaussian head avatar for high-fidelity rendering and head editing,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 26\,274--26\,284

    2025

  9. [17]

    Y. Xu, Z. Su, Q. Wu, and Y. Liu, ``Gphm: Gaussian parametric head model for monocular head avatar reconstruction,'' 2024. [Online]. Available: https://arxiv.org/abs/2407.15070

    2024 arXiv

  10. [18]

    Giebenhain, T

    S. Giebenhain, T. Kirschstein, M. R\" u nz, L. Agapito, and M. Nie ner, ``Npga: Neural parametric gaussian avatars,'' in SIGGRAPH Asia 2024 Conference, 2024

    2024

  11. [19]

    L. Li, Y. Li, Y. Weng, Y. Zheng, and K. Zhou, ``Rgbavatar: Reduced gaussian blendshapes for online modeling of head avatars,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 10\,747--10\,757

    2025

  12. [20]

    X. Wei, P. Chen, M. Lu, H. Chen, and F. Tian, ``Graphavatar: compact head avatars with gnn-generated 3d gaussians,'' in Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence, 2025

    2025

  13. [21]

    G. Li, P. Gotardo, T. Bolkart, S. Garbin, K. Sarkar, A. Meka, A. Lattas, and T. Beeler, ``Tega: Texture space gaussian avatars for high-resolution dynamic head modeling,'' in Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference C...

    2025

  14. [22]

    Serifi and M

    G. Serifi and M. C. Bühler, ``Hypergaussians: High-dimensional gaussian splatting for high-fidelity animatable face avatars,'' 2025. [Online]. Available: https://arxiv.org/abs/2507.02803

    2025

  15. [23]

    Schmidt, S

    J. Schmidt, S. Giebenhain, and M. Niessner, ``Becominglit: Relightable gaussian avatars with hybrid neural shading,'' 2025. [Online]. Available: https://arxiv.org/abs/2506.06271

    2025

  16. [24]

    Zhang, Y

    D. Zhang, Y. Liu, L. Lin, Y. Zhu, K. Chen, M. Qin, Y. Li, and H. Wang, ``Hravatar: High-quality and relightable gaussian head avatar,'' in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 26\,285--26\,296

    2025

  17. [25]

    P. Manu, A. Srivastava, A. Raj, V. Jampani, A. Sharma, and P. J. Narayanan, ``Lightheaded: Relightable & editable head avatars from a smartphone,'' 2025. [Online]. Available: https://arxiv.org/abs/2504.09671

    2025 arXiv

  18. [26]

    Z. Liao, Y. Xu, Z. Li, Q. Li, B. Zhou, R. Bai, D. Xu, H. Zhang, and Y. Liu, ``Hhavatar: Gaussian head avatar with dynamic hairs,'' 2024. [Online]. Available: https://arxiv.org/abs/2312.03029

    2024 arXiv

  19. [27]

    Zheng, M

    Y. Zheng, M. Chai, D. Vicini, Y. Zhou, Y. Xu, L. Guibas, G. Wetzstein, and T. Beeler, ``Groomlight: Hybrid inverse rendering for relightable human hair appearance modeling,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 16\,04...

    2025

  20. [28]

    Zhang, Z

    J. Zhang, Z. Wu, Z. Liang, Y. Gong, D. Hu, Y. Yao, X. Cao, and H. Zhu, ``Fate: Full-head gaussian avatar with textural editing from monocular video,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 5535--5545

    2025

  21. [29]

    Z. Yu, T. Li, J. Sun, O. Shapira, S. Park, M. Stengel, M. Chan, X. Li, W. Wang, K. Nagano, and S. De Mello, ``Gaia: Generative animatable interactive avatars with expression-conditioned gaussians,'' in Proceedings of the Special Interest Group on Computer Graphics and Interact...

    2025

  22. [30]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, ``Generative adversarial networks,'' 2014. [Online]. Available: https://arxiv.org/abs/1406.2661

    2014 arXiv

  23. [31]

    Karras, S

    T. Karras, S. Laine, and T. Aila, ``A style-based generator architecture for generative adversarial networks,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 4401--4410

    2019

  24. [32]

    Patsouras, A

    O. Patsouras, A. Tefas, N. Nikolaidis, and I. Pitas, ``Styleclip: Text-driven manipulation of stylegan imagery,'' in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 2085--2094

    2021

  25. [33]

    M. W. Khan, M. Jia, X. Zhang, E. Yu, C. Shan, and K. Musial-Gabrys, ``Instaface: Identity-preserving facial editing with single image inference,'' 2025. [Online]. Available: https://arxiv.org/abs/2502.20577

    2025 arXiv

  26. [34]

    Zhong, X

    Y. Zhong, X. Zhang, Y. Zhao, and Y. Wei, ``Dreamlcm: Towards high quality text-to-3d generation via latent consistency model,'' in Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 1731--1740

    2024

  27. [35]

    J. Wu, C. Zhang, T. Xue, W. T. Freeman, and J. B. Tenenbaum, ``Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling,'' in Proceedings of the International Conference on Neural Information Processing Systems, 2016, p. 82–90

    2016

  28. [36]

    R. Li, X. Li, K.-H. Hui, and C.-W. Fu, ``Sp-gan: sphere-guided 3d shape generation and manipulation,'' ACM Transactions on Graphics, vol. 40, no. 4, 2021

    2021

  29. [37]

    Z. Yang, Y. Chen, X. Zheng, Y. Chang, and X. Li, ``Conditional gan for point cloud generation,'' in 16th Asian Conference on Computer Vision, 2022, p. 117–133

    2022

  30. [38]

    Pemasiri, K

    A. Pemasiri, K. Nguyen, S. Sridharan, and C. Fookes, ``Accurate 3d hand mesh recovery from a single rgb image,'' Scientific Reports, vol. 12, 2022

    2022

  31. [39]

    J. Gao, T. Shen, Z. Wang, W. Chen, K. Yin, D. Li, O. Litany, Z. Gojcic, and S. Fidler, ``Get3d: A generative model of high quality 3d textured shapes learned from images,'' in Proceedings of the International Conference on Neural Information Processing Systems, 2022, pp. 31\,8...

    2022

  32. [40]

    E. R. Chan, C. Z. Lin, M. A. Chan, K. Nagano, B. Pan, S. de Mello, O. Gallo, L. Guibas, J. Tremblay, S. Khamis, T. Karras, and G. Wetzstein, ``Efficient geometry-aware 3d generative adversarial networks,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Patter...

    2022

  33. [41]

    C. Sun, Y. Liu, J. Han, and S. Gould, ``Nerfeditor: Differentiable style decomposition for 3d scene editing,'' in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 7306--7315

    2024

  34. [42]

    Roich, R

    D. Roich, R. Mokady, A. H. Bermano, and D. Cohen-Or, ``Pivotal tuning for latent-based editing of real images,'' ACM Transactions on Graphics, vol. 42, no. 1, 2022

    2022

  35. [43]

    J. He, W. Shi, K. Chen, L. Fu, and C. Dong, ``Gcfsr: a generative and controllable face super resolution method without facial and gan priors,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 1879--1888

    2022

  36. [44]

    Szegedy, W

    C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, ``Going deeper with convolutions,'' in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 1--9

    2015

  37. [45]

    [Online]

    OpenAI, ``Gpt-4o system card,'' 2024. [Online]. Available: https://arxiv.org/abs/2410.21276

    2024 arXiv

  38. [46]

    Y. Xu, L. Wang, X. Zhao, H. Zhang, and Y. Liu, ``Avatarmav: Fast 3d head avatar reconstruction using motion-aware neural voxels,'' in ACM SIGGRAPH 2023 Conference Proceedings, 2023

    2023

  39. [47]

    Zheng, W

    Y. Zheng, W. Yifan, G. Wetzstein, M. J. Black, and O. Hilliges, ``Pointavatar: Deformable point-based head avatars from videos,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 21\,057--21\,067

    2023

  40. [48]

    Zielonka, T

    W. Zielonka, T. Bolkart, and J. Thies, ``Instant volumetric head avatars,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4574--4584

    2023

  41. [49]

    Bharadwaj, Y

    S. Bharadwaj, Y. Zheng, O. Hilliges, M. J. Black, and V. F. Abrevaya, ``Flare: Fast learning of animatable and relightable mesh avatars,'' ACM Transactions on Graphics, vol. 42, no. 6, 2023

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.