Pith. sign in

REVIEW 3 major objections 4 minor 62 references

FA-LAM claims that one-shot animatable head avatars fail at fine details and wide views because attention is unconstrained and reconstruction/animation training conflicts, and fixes both with a symmetry-aware attention loss and a two-stage

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 08:58 UTC pith:Q7V7KV3T

load-bearing objection A solid, well-evaluated incremental advance in one-shot animatable head avatars, but the wide-angle claim rests on unvalidated Qwen-Image-Edit pseudo-views, and the paper overstates its checklist compliance. the 3 major comments →

arxiv 2607.20922 v1 pith:Q7V7KV3T submitted 2026-07-23 cs.CV

FA-LAM: Focus-Aware Large Avatar Model for One-Shot 4D Animatable Gaussian Head

Yingdong Hu , Yisheng He , Yiming Jiang , Zehong Lin , Steven Hoi , Jun Zhang This is my paper
classification cs.CV
keywords one-shot avatar reconstruction3D Gaussian headattention regularizationbilateral symmetry priordual-phase trainingautoregressive reconstructionvisibility-aware fusionnovel view synthesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

FA-LAM argues that the two biggest quality problems in one-shot animatable head avatars — blurred fine facial details and distorted wide-angle views — come from identifiable, fixable causes rather than model capacity: cross-attention is spatially unconstrained, and joint training for reconstruction and animation pushes gradients in conflicting directions. The paper's fix is a regularization loss that forces each UV-space token to attend to its semantic and mirror-symmetric image location, plus a two-stage recipe that first learns static multi-view geometry and then adds expression residuals. It also shows how to extend the same model to streaming video with a visibility-gated autoregressive fusion that keeps memory constant. If these mechanisms work as claimed, a single portrait produces an animatable Gaussian full head whose side views and fine regions render far more faithfully than previous feed-forward avatars, and long head-video streams can be processed without quadratic memory growth. A careful reader should note that the wide-angle supervision comes from an image-editing model, which is the load-bearing assumption for the large-view improvements.

Core claim

The paper's central claim is that one-shot animatable 3D Gaussian head reconstruction fails in fine facial regions and large viewing angles because (1) transformer cross-attention between UV-space tokens and image features is trained without explicit spatial supervision, letting invisible or occlusion-prone tokens attend to spurious patches, and (2) optimizing novel-view synthesis and expression animation simultaneously produces conflicting gradients that degrade wide-angle geometry. To test and fix this, FA-LAM injects a semantic-and-symmetric attention prior into every cross-attention layer, matching predicted attention maps to a Gaussian mixture centered on the reprojected FLAME correspon

What carries the argument

The key mechanism is attention regularization: a KL loss (Eq. 4) between each predicted cross-attention map and a Gaussian mixture placed at the token's semantic correspondence (reprojected FLAME landmark) and its symmetric counterpart, which teaches invisible tokens to borrow from their visible mirror. The second mechanism is the dual-phase training schedule — reconstruction-only on multi-view and pseudo-multi-view data, then expression-residual fine-tuning — which avoids gradient conflict documented by negative cosine similarities between the two objectives. The third is visibility-gated autoregressive fusion: a running UV-token state is updated by blending previous and current tokens only

Load-bearing premise

The large-angle supervision comes from pseudo side views generated by an image-editing model; if those generated views are not identity-consistent and multi-view consistent, the reported wide-angle improvements may be an artifact of fitting the generator's distribution rather than genuine 3D generalization.

What would settle it

Take a held-out identity from a real multi-view head dataset, render or capture true images at 45° and 90° yaw, and compare FA-LAM's novel-view outputs against those ground-truth frames using PSNR/SSIM/LPIPS, alongside a baseline that was trained without any pseudo-multiview data. If the gap over baselines shrinks or disappears on real wide-angle frames while persisting on Qwen-generated frames, the claim of superior large-view synthesis is not sustained by real geometry.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • On held-out identities from NeRSemble-v2, the model reports 25.80 PSNR for novel views from a single image, a large jump over the LAM baseline's 18.89, suggesting attention regularization mainly recovers geometry that prior attention wasted.
  • Ablations show recon-then-animation training (25.80 PSNR) beats joint training (24.99), supporting the claim that disentangling the two losses is what preserves wide-view quality.
  • Autoregressive visibility-aware fusion achieves comparable or better image quality than per-frame reconstruction while adding only about 20 MB of GPU memory per additional frame, enabling long streams without out-of-memory.
  • The attention prior generalizes across self- and cross-reenactment: cross-identity CSIM improves to 0.692 vs LAM's 0.651, meaning the same model transfers better to unseen faces.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the analysis is right, the bottleneck for large-view hallucination in feed-forward head models is not expressive capacity but under-constrained attention; the same symmetry-and-semantics regularization could be dropped into other UV-space or UV-mapped generative models with little modification.
  • Because the wide-angle supervision is generated by a text-to-image editing model (Qwen-Image-Edit), the reported large-angle gains may partly reflect matching that generator's distribution; a decisive test would be evaluating on real multi-view captures with cameras at 45–90 degrees from frontal.
  • The visibility-gated fusion is a generic mechanism for any streaming reconstruction setting where observations are partial and overlapping (e.g., full-body scans or object-level multi-view), not just heads; the constant-memory property makes it attractive for on-device capture.
  • The attention-mapping loss could double as an interpretability tool: after training, attention maps are concentrated on semantically meaningful regions, which might help diagnose other failures in avatar models.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes FA-LAM, a unified feed-forward model for one-shot animatable Gaussian head avatars, static 3D reconstruction, and streaming 4D reconstruction. The contributions are: (1) a semantic and symmetric attention regularization that uses FLAME UV correspondences to shape cross-attention targets; (2) a two-stage training pipeline that first trains for reconstruction on multi-view data and then fine-tunes for animation, motivated by an empirical gradient-conflict analysis; and (3) an autoregressive extension with visibility-gated token fusion for constant-memory streaming inputs. The model is trained on VFHQ, NeRSemble-v2, Ava256, and a self-curated MV-VFHQ dataset generated with Qwen-Image-Edit. Experiments report consistent improvements over state-of-the-art baselines on self- and cross-reenactment, novel-view synthesis, few-shot input, and streaming 4D reconstruction, with ablations for the attention regularization and training strategy.

Significance. If the reported results hold, FA-LAM provides a practical, unified solution to several related tasks in head avatar reconstruction, and the proposed mechanisms are intuitive and well-motivated: symmetry-awareness is a natural prior for heads, the reconstruction/animation decoupling is a plausible remedy for a real training conflict, and the visibility-gated fusion addresses a genuine scalability bottleneck. The experimental coverage is broad (four datasets, multiple tasks, ablations), and the quantitative gains over prior work are consistent, not cherry-picked. However, the current evidence has two structural weaknesses that affect the strength of the claims: there are no error bars or significance tests despite a checklist item asserting they exist, and the large-view training signal relies on an unvalidated pseudo-multi-view dataset. These issues prevent the paper from being accepted in its present form.

major comments (3)
  1. [Sec. 4 (all tables) and NeurIPS Checklist item 7] The checklist answers [Yes] to 'Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?' with justification 'Please see Sec. 4', but no table or figure in Sec. 4 or the appendix reports error bars, confidence intervals, or significance tests. The main claims rest on absolute differences such as the +2.42 dB PSNR gain over GAGAvatar in Table 3 and +0.27 dB over LAM in Table 2; without variance across test identities or runs, the reader cannot judge whether these gaps are meaningful. Add error bars (e.g., per-identity standard deviation) or significance tests to the primary tables, and correct the checklist answer accordingly.
  2. [Sec. 4.1 and Appendix B (MV-VFHQ)] The large-view reconstruction claim is trained on MV-VFHQ, where Qwen-Image-Edit is prompted to 'turn camera right/left 45/90°' and the outputs are treated as ground-truth novel views. The paper provides no validation that these synthetic images preserve identity, geometry, or multi-view consistency with the input frame. Because the ablation tables (Tables 6 and 7) are all trained under this same pseudo-supervision, they cannot establish that the supervised target is correct; the gains could reflect matching the generator's distribution rather than true 3D generalization. The comparison with baselines (e.g., LAM) is also partially confounded because those models were trained on monocular VFHQ without the four synthetic views. Please add (a) quantitative consistency checks of the pseudo-views (e.g., identity similarity, face-reprojection error) or a user study, and (b) an ablation that tr
  3. [Sec. 3.4, Eq. (6)] The visibility-gated fusion is a core contribution for the streaming claim, but the paper never defines the visibility mask C(t). The text only says it 'measures the visibility of each token in the current view.' Without a precise definition — e.g., whether it is derived from rasterizing the FLAME mesh, from the 3D Gaussian positions, or from learned head pose — the fusion update in Eq. (6) is not reproducible. Please provide the exact computation of C(t) and the update rule for the history visibility C'(t-1), and state whether the mask is binary or soft.
minor comments (4)
  1. [Sec. 3.3 / Fig. 3] The gradient-conflict analysis is described qualitatively ('cosine similarity < -0.1 for nearly half of the pairs') but the numerical distribution is not reported. Please provide the mean/median cosine similarity, the fraction of negative pairs, and a brief description of the t-SNE protocol, so the reader can assess the strength of the conflict.
  2. [Sec. 5] The limitations discussion is a single sentence about future work on occiputs, yet the checklist states limitations are discussed in Sec. 5. The paper does not address other important limitations, such as dependence on FLAME tracking quality, the unvalidated pseudo-multi-view supervision, or sensitivity to extreme expressions/poses. Please expand the limitation statement.
  3. [Table 4] The AKD value for GAGAvatar in the one-shot block is listed as '18,585'; this appears to be a typo for 18.585 and should be corrected.
  4. [Throughout] Minor typos: 'VHFQ' in Sec. 4.2 should be 'VFHQ'; 'Self-Renactment' in Table 2 caption should be 'Self-Reenactment'; Figure 3's axes and legends appear garbled in the submitted text and should be checked in the camera-ready.

Circularity Check

0 steps flagged

No circular derivation found: the avatar is decoded from inputs, the headline wide-angle/4D claims are evaluated on held-out real identities, and the proposed training/regularization choices are tested by ablations rather than assumed by construction.

full rationale

I walked the claimed derivation chain and found no step that reduces to its own inputs. Eq. 1 defines G3D as a decoder output over UV tokens and ViT features; the final reconstruction is never defined in terms of the reported metrics. Eq. 4 supervises attention maps toward FLAME-derived semantic and symmetric correspondences; those correspondences come from external tracking, not from the reconstruction or from the metrics, so the attention regularizer is not self-definitional. The two-stage training claim is supported by an empirical gradient-correlation observation in Sec. 3.3 and then tested by ablations in Table 6 that vary supervision; the improvement of reconstruction-then-animation over joint training is a contingent experimental result, not a tautology. The autoregressive fusion in Eqs. 5-6 updates UV tokens from visibility maps and previous/current frame features; no fitted parameter is renamed as a prediction. The headline wide-angle and 4D claims are evaluated on held-out real identities (NeRSemble-v2 ids 017/018, Ava256 held-out id), not on the MV-VFHQ pseudo-views used during training, so the pseudo-multi-view supervision is a data-quality and confound risk, not a circularity. The paper is heavily self-citational (LAM, PanoLAM, Forge4D), but those are used as baselines and architectural starting points, not as an imported uniqueness theorem or as external proof that forces the result. The checklist's pointer to Sec. 5 for limitations is weak because Sec. 5 is only a conclusion, but that is a reporting defect, not a circular step.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central claim rests on two hand-chosen hyperparameters (tau, loss weights), two domain assumptions about human head symmetry and FLAME tracking reliability, and one dataset-generation assumption about Qwen-Image-Edit's synthetic views. No new physical or model entities are introduced. The pseudo-label assumption is the most load-bearing because it underpins the wide-angle reconstruction gain.

free parameters (2)
  • Gaussian mixture standard deviation tau in attention regularization loss = not reported
    Eq. 4 defines Lattn = KL(Mattn, GMM(Mreg, tau)) with a 'fixed standard deviation tau' whose value is never given. The behavior of the regularization changes with tau, making it a hand-chosen constant that the central claim depends on.
  • Composite loss weights in Eq. 7 = not reported
    Eq. 7 sums MSE, LPIPS, mask loss, and attention loss without reporting weighting coefficients or how they were chosen. The relative weighting directly affects how strongly attention regularization and animation objectives are balanced; its omission makes the training recipe incomplete.
axioms (3)
  • domain assumption Human heads are approximately bilaterally symmetric in UV space, so symmetric correspondences can be used to supervise attention for occluded regions.
    Invoked in Sec. 3.2: the method relies on facial symmetry as a reliable prior for hallucinating invisible regions. Faces are not perfectly symmetric (blemishes, hair, asymmetric expressions), and the paper does not quantify how violations affect the regularization.
  • domain assumption FLAME tracking provides sufficiently accurate semantic correspondences, UV extraction, and visibility information.
    Used in Sec. 3.1-3.2 and Eq. 6: the whole method depends on FLAME pose/expression/camera parameters and on the projected FLAME template to define UV coordinates and visibility masks. Tracking failures would corrupt both the attention targets and the autoregressive fusion.
  • ad hoc to paper Qwen-Image-Edit generated four-synthetic-views-per-frame images are reliable multi-view supervision for training.
    Appendix B describes the construction of MV-VFHQ with Qwen-Image-Edit side views. The paper provides no verification that these pseudo-views are identity-consistent or geometrically valid, and this is the core training signal for the large-view-angle capability claimed in the abstract.

pith-pipeline@v1.3.0-alltime-deepseek · 21066 in / 8376 out tokens · 88722 ms · 2026-08-01T08:58:04.992204+00:00 · methodology

0 comments
read the original abstract

We propose FA-LAM, a Focus-Aware Large Avatar Model for one-shot animatable Gaussian head creation, while simultaneously enabling static 3D and dynamic 4D full-head recovery. The core of our method lies in a thorough analysis of the attention mechanisms and the entangled reconstruction and animation training pipeline adopted by prior state-of-the-art approaches. Our analysis identifies two main factors that compromise the quality of 3D full-head generation: (1) incorrect and noisy attention activations, and (2) conflicts between the tasks of reconstruction and animation. To address the first issue, we introduce a symmetric and semantic attention regularization strategy that leverages the inherent semantics and structural symmetry of human heads. To disentangle the objectives of reconstruction and animation, we develop a novel dual-phase training pipeline that separates the model's capabilities for large-view hallucination and animation into distinct modules. Moreover, we enhance our model to support multi-view and streaming 4D reconstruction in an efficient and memory-friendly manner through a core autoregressive modification with tailored visibility-aware token fusion. Collectively, these innovations enable FA-LAM to reconstruct animatable Gaussian full heads with superior quality, particularly in fine facial regions and large viewing angles.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 14 linked inside Pith

  1. [1]

    Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians

    Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299–20309, 2024

    2024

  2. [2]

    Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data

    Yu Deng, Duomin Wang, Xiaohang Ren, Xingyu Chen, and Baoyuan Wang. Portrait4d: Learning one-shot 4d head avatar synthesis using synthetic data. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  3. [3]

    Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer.arXiv preprint arXiv:2403.13570, 2024

    Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer.arXiv preprint arXiv:2403.13570, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Fate: Full-head gaussian avatar with textural editing from monocular video

    Jiawei Zhang, Zijian Wu, Zhiyang Liang, Yicheng Gong, Dongfang Hu, Yao Yao, Xun Cao, and Hao Zhu. Fate: Full-head gaussian avatar with textural editing from monocular video. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5535–5545, 2025

    2025

  5. [5]

    Surfhead: Affine rig blending for geometrically accurate 2d gaussian surfel head avatars

    Jaeseong Lee, Taewoong Kang, Marcel Buehler, Min-Jung Kim, Sungwon Hwang, Junha Hyung, Hyojin Jang, and Jaegul Choo. Surfhead: Affine rig blending for geometrically accurate 2d gaussian surfel head avatars. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=1x1gGg49jr

    2025

  6. [6]

    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. InACM SIGGRAPH 2023 Conference Proceedings, 2023

    2023

  7. [7]

    3d gaussian parametric head model

    Yuelang Xu, Lizhen Wang, Zerong Zheng, Zhaoqi Su, and Yebin Liu. 3d gaussian parametric head model. InEuropean Conference on Computer Vision, pages 129–147. Springer, 2024

    2024

  8. [8]

    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, Guidong Wang, and Lan Xu. HeadGAP: Few-Shot 3D Head Avatar via Generalizable Gaussian Priors . In2025 International Conference on 3D Vision (3DV), pages 946–957, Los Alamitos, CA, USA, March 2025. IEEE Computer Society. doi: 10.1109/3D...

    arXiv 2025

  9. [9]

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

    Zhixuan Yu, Ziqian Bai, Abhimitra Meka, Feitong Tan, Qiangeng Xu, Rohit Pandey, Sean Fanello, Hyun Soo Park, and Yinda Zhang. One2avatar: Generative implicit head avatar for few-shot user adaptation, 2024. URLhttps://arxiv.org/abs/2402.11909

    Pith/arXiv arXiv 2024

  10. [10]

    Monogaussianavatar: Monocular gaussian point-based head avatar

    Yufan Chen, Lizhen Wang, Qijing Li, Hongjiang Xiao, Shengping Zhang, Hongxun Yao, and Yebin Liu. Monogaussianavatar: Monocular gaussian point-based head avatar. InACM SIGGRAPH 2024 conference papers, pages 1–9, 2024. 10

    2024

  11. [11]

    LRM: Large reconstruction model for single image to 3d

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3d. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=sllU8vvsFF

    2024

  12. [12]

    Lam: Large avatar model for one-shot animatable gaussian head

    Yisheng He, Xiaodong Gu, Xiaodan Ye, Chao Xu, Zhengyi Zhao, Yuan Dong, Weihao Yuan, Zilong Dong, and Liefeng Bo. Lam: Large avatar model for one-shot animatable gaussian head. InProceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, pages 1–13, 2025

    2025

  13. [13]

    Meshlam: Feed-forward one-shot animatable textured mesh avatar reconstruction, 2026

    Yisheng He and Steven Hoi. Meshlam: Feed-forward one-shot animatable textured mesh avatar reconstruction, 2026. URLhttps://arxiv.org/abs/2604.22865

    Pith/arXiv arXiv 2026

  14. [14]

    Panolam: Large avatar model for gaussian full-head synthesis from one-shot unposed image, 2025

    Peng Li, Yisheng He, Yingdong Hu, Yuan Dong, Weihao Yuan, Yuan Liu, Siyu Zhu, Gang Cheng, Zilong Dong, and Yike Guo. Panolam: Large avatar model for gaussian full-head synthesis from one-shot unposed image, 2025. URL https://arxiv.org/abs/2509.07552

    arXiv 2025

  15. [15]

    Avat3r: Large animatable gaussian reconstruction model for high-fidelity 3d head avatars

    Tobias Kirschstein, Javier Romero, Artem Sevastopolsky, Matthias Nießner, and Shunsuke Saito. Avat3r: Large animatable gaussian reconstruction model for high-fidelity 3d head avatars. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12089–12100, October 2025

    2025

  16. [16]

    Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans.ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1–194:17, 2017. URLhttps://doi.org/10.1145/3130800.3130813

    arXiv 2017

  17. [17]

    Forge4d: Feed-forward 4d human reconstruction and interpolation from uncalibrated sparse-view videos, 2025

    Yingdong Hu, Yisheng He, Jinnan Chen, Weihao Yuan, Kejie Qiu, Zehong Lin, Siyu Zhu, Zilong Dong, and Jun Zhang. Forge4d: Feed-forward 4d human reconstruction and interpolation from uncalibrated sparse-view videos, 2025. URLhttps://arxiv.org/abs/2509.24209

    arXiv 2025

  18. [18]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  19. [19]

    Streaming 4d visual geometry transformer.arXiv preprint arXiv:2507.11539, 2025

    Dong Zhuo, Wenzhao Zheng, Jiahe Guo, Yuqi Wu, Jie Zhou, and Jiwen Lu. Streaming 4d visual geometry transformer.arXiv preprint arXiv:2507.11539, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Vfhq: A high-quality dataset and benchmark for video face super-resolution

    Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. InThe IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022

    2022

  21. [21]

    Qwen-image technical report, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun...

    Pith/arXiv arXiv 2025

  22. [22]

    Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    2014

  23. [23]

    Neural head reenactment with latent pose descriptors

    Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786–13795, 2020

    2020

  24. [24]

    Progressive disentangled representation learning for fine-grained controllable talking head synthesis

    Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979–17989, 2023. 11

    2023

  25. [25]

    Few-shot ad- versarial learning of realistic neural talking head models

    Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot ad- versarial learning of realistic neural talking head models. InProceedings of the IEEE/CVF international conference on computer vision, pages 9459–9468, 2019

    2019

  26. [26]

    Depth-aware generative adversarial network for talking head video generation

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

    2022

  27. [27]

    A style-based generator architecture for generative adversarial networks

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

    2019

  28. [28]

    Megaportraits: One-shot megapixel neural head avatars

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

    2022

  29. [29]

    Liveportrait: Efficient portrait animation with stitching and retargeting control.arXiv preprint arXiv:2407.03168, 2024

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

    Pith/arXiv arXiv 2024

  30. [30]

    Metaportrait: Identity-preserving talking head generation with fast personalized adaptation

    Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, HsiangTao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096–22105, 2023

    2023

  31. [31]

    Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions

    Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions. InEuropean Conference on Computer Vision, pages 244–260. Springer, 2024

    2024

  32. [32]

    Hallo2: Long-duration and high-resolution audio-driven portrait image animation.arXiv preprint arXiv:2410.07718, 2024

    Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2: Long-duration and high-resolution audio-driven portrait image animation.arXiv preprint arXiv:2410.07718, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Hallo: Hierarchical audio-driven visual synthesis for portrait image animation.arXiv preprint arXiv:2406.08801, 2024

    Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation.arXiv preprint arXiv:2406.08801, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Felix Taubner, Ruihang Zhang, Mathieu Tuli, and David B. Lindell. CAP4D: Creating animat- able 4D portrait avatars with morphable multi-view diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5318–5330, June 2025

    2025

  35. [35]

    Mvp4d: Multi-view portrait video diffusion for animatable 4d avatars

    Felix Taubner, Ruihang Zhang, Mathieu Tuli, Sherwin Bahmani, and David B Lindell. Mvp4d: Multi-view portrait video diffusion for animatable 4d avatars. InProceedings of the SIGGRAPH Asia 2025 Conference Papers, pages 1–11, 2025

    2025

  36. [36]

    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. InSeminal Graphics Papers: Pushing the Boundaries, V olume 2, pages 851–866. 2023

    2023

  37. [37]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. InProceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  38. [38]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoor- thi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021

    2021

  39. [39]

    GPAvatar: Generalizable and precise head avatar from image(s)

    Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. GPAvatar: Generalizable and precise head avatar from image(s). InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum? id=hgehGq2bDv. 12

    2024

  40. [40]

    GGHead: Fast and Generalizable 3D Gaussian Heads

    Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner. GGHead: Fast and Generalizable 3D Gaussian Heads. InSIGGRAPH Asia 2024 Conference Papers, SA ’24, New York, NY , USA, 2024. Association for Computing Machinery. ISBN 9798400711312. doi: 10.1145/3680528.3687686. URL https://doi.org/10.1145/ 3680528.3687686

    arXiv 2024

  41. [41]

    One-Shot High-Fidelity Talking- Head Synthesis with Deformable Neural Radiance Field

    Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhigang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, Liefeng Bo, and Xuelong Li. One-Shot High-Fidelity Talking- Head Synthesis with Deformable Neural Radiance Field . In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17969–17978, Los Alamitos, CA, USA, June 2023. IEEE Co...

    arXiv 2023

  42. [42]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis, et al. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1, 2023

    2023

  43. [43]

    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 (CVPR), 2024

    2024

  44. [44]

    Soap: Style-omniscient animatable portraits

    Tingting Liao, Yujian Zheng, Adilbek Karmanov, Liwen Hu, Leyang Jin, Yuliang Xiu, and Hao Li. Soap: Style-omniscient animatable portraits. 2025. URL https://arxiv.org/abs/ 2505.05022

    Pith/arXiv arXiv 2025

  45. [45]

    Generalizable and animatable gaussian head avatar

    Xuangeng Chu and Tatsuya Harada. Generalizable and animatable gaussian head avatar. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=gVM2AZ5xA6

    2024

  46. [46]

    Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis

    Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  47. [47]

    Eva-gaussian: 3d gaussian-based real-time human novel view synthesis under diverse camera settings, 2024

    Yingdong Hu, Zhening Liu, Jiawei Shao, Zehong Lin, and Jun Zhang. Eva-gaussian: 3d gaussian-based real-time human novel view synthesis under diverse camera settings, 2024. URL https://arxiv.org/abs/2410.01425

    Pith/arXiv arXiv 2024

  48. [48]

    Flexavatar: Flexible large reconstruction model for animatable gaussian head avatars with detailed deformation, 2025

    Cheng Peng, Zhuo Su, Liao Wang, Chen Guo, Zhaohu Li, Chengjiang Long, Zheng Lv, Jingxiang Sun, Chenyangguang Zhang, and Yebin Liu. Flexavatar: Flexible large reconstruction model for animatable gaussian head avatars with detailed deformation, 2025. URL https: //arxiv.org/abs/2512.17717

    arXiv 2025

  49. [49]

    Flexavatar: Learning complete 3d head avatars with partial supervision.arXiv preprint arXiv:2512.15599, 2025

    Tobias Kirschstein, Simon Giebenhain, and Matthias Nießner. Flexavatar: Learning complete 3d head avatars with partial supervision.arXiv preprint arXiv:2512.15599, 2025

    Pith/arXiv arXiv 2025

  50. [50]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. InCVPR, 2024

    2024

  51. [51]

    Uika: Fast universal head avatar from pose-free images, 2026

    Zijian Wu, Boyao Zhou, Liangxiao Hu, Hongyu Liu, Yuan Sun, Xuan Wang, Xun Cao, Yujun Shen, and Hao Zhu. Uika: Fast universal head avatar from pose-free images, 2026. URL https://arxiv.org/abs/2601.07603

    Pith/arXiv arXiv 2026

  52. [52]

    Bringing your portrait to 3d presence, 2025

    Jiawei Zhang, Lei Chu, Jiahao Li, Zhenyu Zang, Chong Li, Xiao Li, Xun Cao, Hao Zhu, and Yan Lu. Bringing your portrait to 3d presence, 2025. URL https://arxiv.org/abs/2511. 22553

    2025

  53. [53]

    Gradient surgery for multi-task learning.arXiv preprint arXiv:2001.06782, 2020

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning.arXiv preprint arXiv:2001.06782, 2020

    Pith/arXiv arXiv 2001

  54. [54]

    Visualizing data using t-sne.Journal of Machine Learning Research, 9(86):2579–2605, 2008

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of Machine Learning Research, 9(86):2579–2605, 2008. URL http://jmlr.org/papers/v9/ vandermaaten08a.html. 13

    2008

  55. [55]

    Nersemble: Multi-view radiance field reconstruction of human heads.ACM Trans

    Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads.ACM Trans. Graph., 42 (4), jul 2023. ISSN 0730-0301. doi: 10.1145/3592455. URL https://doi.org/10.1145/ 3592455

    doi:10.1145/3592455 2023

  56. [56]

    Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A

    Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shoou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman,...

    2024

  57. [57]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685–4694, 2019. doi: 10.1109/CVPR.2019.00482

    arXiv 2019

  58. [58]

    How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks)

    Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In2017 IEEE International Conference on Computer Vision (ICCV), pages 1021–1030, 2017. doi: 10.1109/ICCV .2017.116

    doi:10.1109/iccv 2017

  59. [59]

    Real3d-portrait: One-shot realistic 3d talking portrait synthesis.arXiv preprint arXiv:2401.08503, 2024

    Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiawei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis.arXiv preprint arXiv:2401.08503, 2024

    Pith/arXiv arXiv 2024

  60. [60]

    GPAvatar: Generalizable and precise head avatar from image(s)

    Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. GPAvatar: Generalizable and precise head avatar from image(s). InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum? id=hgehGq2bDv

    2024

  61. [61]

    Invertavatar: In- cremental gan inversion for generalized head avatars

    Xiaochen Zhao, Jingxiang Sun, Lizhen Wang, Jinli Suo, and Yebin Liu. Invertavatar: In- cremental gan inversion for generalized head avatars. InACM SIGGRAPH 2024 Con- ference Papers, SIGGRAPH ’24, New York, NY , USA, 2024. Association for Comput- ing Machinery. ISBN 9798400705250. doi: 10.1145/3641519.3657478. URL https: //doi.org/10.1145/3641519.3657478

    arXiv 2024

  62. [62]

    Diffusionrig: Learning personalized priors for facial appearance editing

    Zhihao Xia Lars Jebe Zhuowen Tu Zheng Ding, Cecilia Zhang and Xiuming Zhang. Diffusionrig: Learning personalized priors for facial appearance editing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023