Pith. sign in

REVIEW 3 major objections 6 minor 54 references

MaintaAvatar: A Maintainable Avatar Based on Neural Radiance Fields by Continual Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read A neural-radiance-field human avatar can learn new outfits and poses from a few images per task while still rendering every earlier appearance at near-joint quality, if two targeted fixes are added to replay-based continual learning.

desk verdict A useful continual-learning method for human NeRF avatars, but the main experiment may be testing identity changes rather than same-person appearance changes. read the letter →

arxiv 2502.02372 v1 pith:BWKC4JFD submitted 2025-02-04 cs.CV cs.AI

classification cs.CVcs.AI
keywords neuralradiancefieldshumanavatarcontinuallearningcatastrophicforgettingfew-shotadaptationposedistillationgenerativereplayappearanceediting
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

The paper tries to establish that a neural-radiance-field human avatar can be maintained incrementally: each new task brings only a handful of images of a changed appearance or pose, and after the last task the same weights still render all earlier appearances at quality close to the joint-training upper bound. The authors argue that naive replay of past renderings causes two specific failures — colors from different outfits bleeding into each other, and drift in the rendered body pose — and that both are fixable with a Global-Local Joint Storage Module and a Pose Distillation Module. Experiments on ZJU-MoCap and THuman2.0 report MaintaAvatar approaching the joint-training bound and outperforming continual-NeRF baselines. The practical bet is that a real-world avatar can be updated with limited data collection instead of being retrained from scratch.

What carries the argument

The load-bearing mechanism is a replay-based continual-learning loop built on a deformable SMPL-conditioned NeRF, plus two modules. The Global-Local Joint Storage Module keeps a shared network $\mathrm{MLP}_o$ conditioned by a global geometry embedding $\ell_g$, a global color embedding $\ell_c$, and a tri-plane local embedding $\ell_t = q(\mathcal{G}(\ell_a), x)$ generated per appearance; the local term is what stops color bleeding between appearances. The Pose Distillation Module supervises the pose-residual network $MLP_p$ with the frozen past model's output $\bar{\Delta\Omega}(p)$ through the L2 loss $L_{POSE}=\|\hat{\Delta\Omega}(p)-\bar{\Delta\Omega}(p)\|_2^2$, activated only in the final phase of each task. Together they let the same weights render both old and new appearances without retraining.

What would settle it

Look up the identity metadata for ZJU-MoCap subjects 377, 392, 393, and 394. If the dataset records them as distinct people, the reported near-Joint numbers do not establish the maintainable same-person avatar claim; a decisive experiment would train the pipeline on four outfit changes of one person and check that old outfits remain renderable.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that catastrophic forgetting in a personalized human NeRF can be separated into two defects and corrected by targeted modules. Appearances are separated by storing a global geometry embedding and a global color embedding per identity, plus a per-appearance tri-plane local embedding generated from a condition embedding, so that sampling points query appearance-specific local features and colors do not bleed between outfits. Pose memory is preserved by distilling the pose-correction residual $\Delta\Omega(p)$ of the frozen previous model into the current model's $MLP_p$ during a late training phase while all other weights are frozen. With one frozen copy of the previous network generating replay supervision, the final task's weights render all earlier tasks' appearances at near-Joint quality.

Load-bearing premise

The claim that the ZJU-MoCap experiments test appearance changes of a single person rests on the paper's unsupported statement that subjects 377, 392, 393, and 394 are the same individual in different clothing; if those IDs are distinct people, the main evaluation tests identity changes instead.

Editorial extensions

If this is right

  • After learning the final task, the same network parameters render all previous tasks' appearances at quality close to joint training, so an avatar can be updated incrementally instead of retrained from scratch.
  • A new outfit or pose can be added from roughly five images with viewpoint spread, without storing old training images — only a frozen copy of the previous network.
  • The pose-distillation phase prevents the pose-correction network from overfitting to new poses, avoiding visibly incorrect limbs in older appearances.
  • The Global-Local Joint Storage Module extends replay-based continual NeRF from static scenes to articulated human bodies by separating global and local appearance factors.

Reading between the lines

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

  • If the ZJU-MoCap subject IDs 377, 392, 393, and 394 turn out to be distinct people rather than one person in different outfits, the reported results would still demonstrate continual adaptation across identities, but they would not validate the claimed same-person appearance-maintenance setting; a true test needs a single person photographed in multiple outfits over time.
  • The design suggests a cheap stress test: measure whether color bleeding reappears when the number of sequential tasks grows well beyond four, since per-appearance tri-planes add storage linearly with each identity.
  • The same recipe of a frozen replay network plus late-phase pose-residual distillation could plausibly transfer to 3D Gaussian avatar representations, where forgetting may show up as geometry drift rather than color bleeding.
  • Pose distillation as a late-phase fine-tuning step could serve as a general mechanism for any animatable NeRF that has an auxiliary pose-correction MLP, independent of the appearance-storage module.
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 / 6 minor

Summary. The paper introduces MaintaAvatar, a continual learning framework for animatable NeRF-based human avatars that can be updated with a few images of new appearances and poses while retaining the ability to render previously seen appearances. The method builds on a PersonNeRF-style deformable NeRF and combines three components: a global-local joint storage module that represents each appearance with global embeddings and a per-appearance triplane, a replay-based continual learning loss that distills the frozen previous model's outputs, and a pose distillation module that regularizes the pose-correction MLP across tasks. The authors evaluate on ZJU-MoCap and THuman2.0, comparing against CLNeRF, MEIL-NeRF, PersonNeRF, and a continual variant of PersonNeRF, and report that the method approaches the Joint upper bound and outperforms all baselines. They also provide ablations for each module and a hyperparameter study for the replay weight.

Significance. If the results hold, MaintaAvatar addresses a practical and understudied problem: maintaining a single person's avatar as their appearance and poses change over time, using small amounts of new data and without retraining from scratch. The problem formulation is clearly motivated, and the modular design is sensible: per-appearance triplanes and global embeddings are a reasonable way to separate appearance information, and pose distillation directly targets a failure mode that the paper identifies and visualizes. The paper includes ablations for each proposed module, and the comparisons to several NeRF continual learning baselines are useful. However, the central experimental evidence depends on an unverified assumption about the ZJU-MoCap dataset, and the quantitative claims rest on a single seed with no error bars; these issues are load-bearing for the paper's main claim and prevent acceptance in the current form.

major comments (3)
  1. [Section 4.1] The ZJU-MoCap paragraph states that subjects (377, 392, 393, 394) "all feature the same individual in different sets of clothing," but no citation or metadata reference is provided. The widely used ZJU-MoCap metadata lists these IDs as distinct subjects. If that is correct, then the sequence of tasks in Table 1 is a sequence of identity changes, not appearance changes of one person, and the experiment does not validate the claimed maintainable-avatar setting. The THuman2.0 experiment explicitly uses different subjects (262, 220, 207, 125), so it cannot substitute for a same-person multi-outfit test. The authors must either provide verifiable evidence for the same-individual claim or re-run the evaluation on a dataset or data split in which the same person genuinely appears in multiple outfits across tasks.
  2. [Tables 1-3 and Section 4.2] All quantitative results are reported for a single random seed (set to 42 in Section 4.2), with no error bars or significance tests. Several of the claimed improvements over the strongest baseline are small; for example, in the novel-view column of Table 1, the proposed method achieves PSNR 29.495 versus 28.605 for PersonNeRFCL. Without multiple seeds or per-task variance, the state-of-the-art claim is not statistically established. The ablations in Table 3 show similarly small gaps (full model 29.50 versus w/o G-L 28.89 in novel-view PSNR), so the module-level contributions also need variance estimates. Please report results over multiple seeds and, ideally, per-task breakdowns for each previously learned appearance.
  3. [Section 4.3, Table 1] The comparison table for ZJU-MoCap is incomplete: the CLNeRF row reports only PSNR and SSIM, with LPIPS listed as "--". Since LPIPS is one of the three headline metrics and the paper argues that the proposed method avoids color bleeding (which LPIPS is sensitive to), omitting this value for a primary baseline makes the comparison harder to assess. Please provide the missing metric or explain why it is unavailable.
minor comments (6)
  1. [Section 3.4, Eq. (8) and Eq. (12)] The notation b∆Ω(p) and e∆Ω(p) is used in Eq. (8) before the text defines these as outputs of the novel and frozen pose-correction networks; please define the symbols at first use.
  2. [Section 3.4] The two-phase training schedule is described in prose but is not stated precisely in the equations; the condition that freezes all networks except MLPp and ℓc appears only as a sentence after Eq. (12). A pseudo-code algorithm or a clear statement of which parameters are trainable in each phase would improve reproducibility.
  3. [Section 4.2] The implementation details state that one camera is used for training and 22 cameras for evaluation in ZJU-MoCap; please clarify whether the same single camera is used for all tasks and whether the five training images per task come from that camera only.
  4. [Section 5, Limitations] The Limitations paragraph concedes that the method "shows performance drops with significant clothing shape changes," which is directly relevant to the paper's core claim of maintaining old appearances. This caveat should be stated in the abstract or conclusion, and the evaluation should quantify how much degradation occurs for large clothing-shape changes.
  5. [Throughout] The dataset name is spelled inconsistently as "Thuman2.0" and "THuman2.0"; please use a single spelling throughout, and also check that all references to the THuman2.0 paper are formatted consistently.
  6. [Section 3.1] The deformable NeRF backbone follows PersonNeRF, but the reference to PersonNeRF appears only in the preliminaries text and is not cited at the point where Eq. (1) is introduced; please cite Weng et al. (2023) explicitly at that equation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported continual-learning pipeline is a standard replay/distillation scheme with per-appearance memory, and the central numbers do not reduce to their inputs by construction.

full rationale

The paper's derivation chain is not circular in the sense defined by the seven patterns. The deformable-NeRF backbone is imported from the external PersonNeRF work; the continual-learning objective LCL compares the current model's render to the frozen previous model's render (Eq. 11), which is a standard self-distillation/replay loss rather than a tautology. The Global-Local Joint Storage Module stores per-appearance condition embeddings and triplanes that are optimized during each task; these are explicit stored parameters used as conditioning for the shared MLP, not a renamed prediction of the evaluation metrics. The Pose Distillation Module similarly distills the previous model's pose-correction output (Eq. 8), a normal teacher-student objective. No equation in the paper is equal to its own input by construction, and no fitted parameter is relabeled as a prediction. The only self-citation of note (Yang et al. 2023, with an overlapping author) merely supports the general phenomenon of catastrophic forgetting and is not load-bearing. The uncited assertion in Section 4.1 that ZJU-MoCap subjects 377, 392, 393, and 394 are the same individual in different clothing is a dataset-validity and external-validity concern, not a circularity: it does not make any derived result equivalent to its input by definition. The THuman2.0 experiments use different subjects and still show the method outperforming baselines, so the central continual-learning mechanism has independent empirical content. Accordingly, no circular step is identified and the score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the replay setup, per-appearance storage, and the dataset identity assumption. The hyperparameters are hand-set. The most fragile input is the ZJU-MoCap identity claim, which the paper asserts without documentation. The frozen past model is also assumed to be a trustworthy teacher. No new physical entities are introduced.

free parameters (6)
  • lambda_1 (current task weight) = 0.2
    Hand-set loss weight for current-task color and LPIPS losses (Eq. 10); no sensitivity analysis reported.
  • lambda_2 (LPIPS weight) = 0.2
    Hand-set weight for LPIPS in both current and replay losses (Eq. 10-11).
  • lambda_beta (pose distillation weight) = 800
    Set to 0 before t0 and 800 after; hand-tuned to balance pose retention and rendering sharpness (Section 3.3).
  • t0 (pose distillation activation iteration) = 10,000 (ZJU), 70,000 (THuman)
    Hand-set switch point for the second training phase.
  • lambda_p (replay weight schedule) = sine ramp from 0 to 1 over [0, tmax-t0]
    Defined in Eq. 12; shape and endpoints are hand-chosen, with tinit and tmax per dataset.
  • Tri-plane resolution = 3*512*512*8 per appearance
    Architectural capacity choice for local appearance storage; chosen by hand.
assumptions (4)
  • domain assumption ZJU-MoCap subjects 377, 392, 393, 394 are the same individual in different sets of clothing
    Section 4.1 states this without citation; standard ZJU-MoCap metadata treats these IDs as distinct subjects. The experiment's validity for the claimed single-person appearance-change setting depends on this assumption.
  • domain assumption SMPL pose parameters and skinning weights are available for all training images
    Used throughout Section 3.1 for the deformation field; accuracy of SMPL fits is assumed.
  • domain assumption The frozen previous model's rendered images are faithful enough to serve as replay supervision
    Replay loss in Eq. 11 uses outputs of the frozen past model; if the teacher is wrong, errors propagate to the current model.
  • ad hoc to paper A pretrained multi-human initialization generalizes to new identities
    Section 4.2 introduces a pretrained model but does not specify its training data or whether it overlaps with test identities; few-shot gains may depend on this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MaintaAvatar: A Maintainable Avatar Based on Neural Radiance Fields by Continual Learning." pith.science (2026). https://pith.science/paper/BWKC4JFD

@misc{pith2026250202372,
  author       = {Pith},
  title        = {Pith review of: MaintaAvatar: A Maintainable Avatar Based on Neural Radiance Fields by Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BWKC4JFD}},
  note         = {Machine review of arXiv:2502.02372}
}
read the original abstract

The generation of a virtual digital avatar is a crucial research topic in the field of computer vision. Many existing works utilize Neural Radiance Fields (NeRF) to address this issue and have achieved impressive results. However, previous works assume the images of the training person are available and fixed while the appearances and poses of a subject could constantly change and increase in real-world scenarios. How to update the human avatar but also maintain the ability to render the old appearance of the person is a practical challenge. One trivial solution is to combine the existing virtual avatar models based on NeRF with continual learning methods. However, there are some critical issues in this approach: learning new appearances and poses can cause the model to forget past information, which in turn leads to a degradation in the rendering quality of past appearances, especially color bleeding issues, and incorrect human body poses. In this work, we propose a maintainable avatar (MaintaAvatar) based on neural radiance fields by continual learning, which resolves the issues by utilizing a Global-Local Joint Storage Module and a Pose Distillation Module. Overall, our model requires only limited data collection to quickly fine-tune the model while avoiding catastrophic forgetting, thus achieving a maintainable virtual avatar. The experimental results validate the effectiveness of our MaintaAvatar model.

Figures

Figures reproduced from arXiv: 2502.02372 by the authors.

Figure 1
Figure 1. In reality, a person’s pose and appearance con [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. MaintaAvatar Pipeline. In this paper, we propose a continual learning strategy pipeline primarily based on the replay method. During the training of Task T, we replicate and freeze the network ΘT −1 from the past Task T-1. Given the camera parameters from Task T-1, the network ΘT −1 can generate corresponding patches and the residual human body pose of one randomly selected past appearance, which are utilized to sup… view at source ↗
Figure 3
Figure 3. Pipeline for MaintaAvatar Network Θ Structure. For any given human body pose, we utilize skeletal motion based on the SMPL model to transform the body from the observation space to the canonical space. Meanwhile, we employ a network MLPp to predict the residual ∆Ω(p) between the current pose parameters and the true pose parameters. Subsequently, our Global-Local Joint Storage Module generates Tri-plane-based local e… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The visualization comparison results for the free [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: The rendering result for the past task (yellow ap [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 5
Figure 5. Figure 5: The Global-Local Joint Storage Module can help [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 8
Figure 8. Figure 8: The hyperparameter [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 35 canonical work pages

  1. [1]

    Bao, C.; Zhang, Y.; Yang, B.; Fan, T.; Yang, Z.; Bao, H.; Zhang, G.; and Cui, Z. 2023. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20919--20929

  2. [2]

    T.; Mildenhall, B.; Tancik, M.; Hedman, P.; Martin-Brualla, R.; and Srinivasan, P

    Barron, J. T.; Mildenhall, B.; Tancik, M.; Hedman, P.; Martin-Brualla, R.; and Srinivasan, P. P. 2021. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 5855--5864

  3. [3]

    Cai, Z.; and M \"u ller, M. 2023. CLNeRF: Continual Learning Meets NeRF. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 23185--23194

  4. [4]

    Chen, Y.; Zhang, J.; Xie, Z.; Li, W.; Zhang, F.; Lu, J.; and Zhang, L. 2024. S-NeRF++: Autonomous Driving Simulation via Neural Reconstruction and Generation. arXiv preprint arXiv:2402.02112

  5. [5]

    Chen, Z.; Wang, F.; and Liu, H. 2023. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585

  6. [6]

    Cheng, K.; Long, X.; Yin, W.; Wang, J.; Wu, Z.; Ma, Y.; Wang, K.; Chen, X.; and Chen, X. 2023. UC-NeRF: Neural Radiance Field for Under-Calibrated multi-view cameras in autonomous driving. arXiv preprint arXiv:2311.16945

  7. [7]

    Chung, J.; Lee, K.; Baik, S.; and Lee, K. M. 2022. MEIL-NeRF: Memory-Efficient Incremental Learning of Neural Radiance Fields. arXiv preprint arXiv:2212.08328

  8. [8]

    Feldmann, C.; Siegenheim, N.; Hars, N.; Rabuzin, L.; Ertugrul, M.; Wolfart, L.; Pollefeys, M.; Bauer, Z.; and Oswald, M. R. 2024. NeRFmentation: NeRF-based Augmentation for Monocular Depth Estimation. arXiv preprint arXiv:2401.03771

Show all 54 references
  1. [9]

    French, R. M. 1999. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4): 128--135

  2. [10]

    Gao, X.; Yang, J.; Kim, J.; Peng, S.; Liu, Z.; and Tong, X. 2022. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence

  3. [11]

    Hore, A.; and Ziou, D. 2010. Image quality metrics: PSNR vs. SSIM. In 2010 20th international conference on pattern recognition, 2366--2369. IEEE

  4. [12]

    Hu, L.; Zhang, H.; Zhang, Y.; Zhou, B.; Liu, B.; Zhang, S.; and Nie, L. 2024 a . Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 634--644

  5. [13]

    Hu, S.; Hong, F.; Pan, L.; Mei, H.; Yang, L.; and Liu, Z. 2023. SHERF: Generalizable Human NeRF from a Single Image. arXiv preprint arXiv:2303.12791

  6. [14]

    Hu, S.; Hu, T.; and Liu, Z. 2024. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20418--20431

  7. [15]

    Hu, X.; Xiong, G.; Zang, Z.; Jia, P.; Han, Y.; and Ma, J. 2024 b . PC-NeRF: Parent-Child Neural Radiance Fields Using Sparse LiDAR Frames in Autonomous Driving Environments. arXiv preprint arXiv:2402.09325

  8. [16]

    M.; Samei, G.; Tuzel, O.; and Ranjan, A

    Jiang, W.; Yi, K. M.; Samei, G.; Tuzel, O.; and Ranjan, A. 2022. Neuman: Neural human radiance field from a single video. In European Conference on Computer Vision, 402--418. Springer

  9. [17]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4)

  10. [18]

    Kolotouros, N.; Alldieck, T.; Zanfir, A.; Bazavan, E.; Fieraru, M.; and Sminchisescu, C. 2024. Dreamhuman: Animatable 3d avatars from text. Advances in Neural Information Processing Systems, 36

  11. [19]

    Kwon, Y.; Kim, D.; Ceylan, D.; and Fuchs, H. 2023. Neural Image-based Avatars: Generalizable Radiance Fields for Human Avatar Modeling. arXiv preprint arXiv:2304.04897

  12. [20]

    Li, Z.; Zheng, Z.; Wang, L.; and Liu, Y. 2024. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19711--19722

  13. [21]

    Liu, L.; Xu, W.; Habermann, M.; Zollh \"o fer, M.; Bernard, F.; Kim, H.; Wang, W.; and Theobalt, C. 2020. Neural human video rendering by learning dynamic textures and rendering-to-video translation. arXiv preprint arXiv:2001.04947

  14. [22]

    Liu, X.; Zhan, X.; Tang, J.; Shan, Y.; Zeng, G.; Lin, D.; Liu, X.; and Liu, Z. 2024. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6646--6657

  15. [23]

    Martin-Brualla, R.; Pandey, R.; Yang, S.; Pidlypenskyi, P.; Taylor, J.; Valentin, J.; Khamis, S.; Davidson, P.; Tkach, A.; Lincoln, P.; et al. 2018. Lookingood: Enhancing performance capture with real-time neural re-rendering. arXiv preprint arXiv:1811.05029

  16. [24]

    S.; Barron, J

    Martin-Brualla, R.; Radwan, N.; Sajjadi, M. S.; Barron, J. T.; Dosovitskiy, A.; and Duckworth, D. 2021. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7210--7219

  17. [25]

    Mendiratta, M.; Pan, X.; Elgharib, M.; Teotia, K.; Tewari, A.; Golyanik, V.; Kortylewski, A.; and Theobalt, C. 2023. Avatarstudio: Text-driven editing of 3d dynamic human head avatars. ACM Transactions on Graphics (ToG), 42(6): 1--18

  18. [26]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

  19. [27]

    Mu, J.; Sang, S.; Vasconcelos, N.; and Wang, X. 2023. ActorsNeRF: Animatable Few-shot Human Rendering with Generalizable NeRFs. arXiv preprint arXiv:2304.14401

  20. [28]

    Pan, X.; Yang, Z.; Ma, J.; Zhou, C.; and Yang, Y. 2023. Transhuman: A transformer-based human representation for generalizable neural human rendering. In Proceedings of the IEEE/CVF International conference on computer vision, 3544--3555

  21. [29]

    Peng, S.; Zhang, Y.; Xu, Y.; Wang, Q.; Shuai, Q.; Bao, H.; and Zhou, X. 2021. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

  22. [30]

    W.; and Wetzstein, G

    Po, R.; Dong, Z.; Bergman, A. W.; and Wetzstein, G. 2023. Instant continual learning of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 3334--3344

  23. [31]

    Qian, Z.; Wang, S.; Mihajlovic, M.; Geiger, A.; and Tang, S. 2024. 3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5020--5030

  24. [32]

    K.; Kim, J.; and Kim, J

    Shin, H.; Lee, J. K.; Kim, J.; and Kim, J. 2017. Continual learning with deep generative replay. Advances in neural information processing systems, 30

  25. [33]

    Sun, J.-M.; Wu, T.; and Gao, L. 2024. Recent advances in implicit representation-based 3d shape generation. Visual Intelligence, 2(1): 9

  26. [34]

    Sun, M.; Yang, D.; Kou, D.; Jiang, Y.; Shan, W.; Yan, Z.; and Zhang, L. 2022. Human 3d avatar modeling with implicit neural representation: A brief survey. In 2022 14th International Conference on Signal Processing Systems (ICSPS), 818--827. IEEE

  27. [35]

    Sun, S.; Cheng, Y.; Gan, Z.; and Liu, J. 2019. Patient knowledge distillation for bert model compression. arXiv preprint arXiv:1908.09355

  28. [36]

    Tang, J.; Ren, J.; Zhou, H.; Liu, Z.; and Zeng, G. 2023. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653

  29. [37]

    Taud, H.; and Mas, J. 2018. Multilayer perceptron (MLP). Geomatic approaches for modeling land change scenarios, 451--455

  30. [38]

    Y.; Ferroni, F.; and Ramanan, D

    Turki, H.; Zhang, J. Y.; Ferroni, F.; and Ramanan, D. 2023. SUDS: Scalable Urban Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12375--12385

  31. [39]

    P.; Zhou, H.; Barron, J

    Wang, Q.; Wang, Z.; Genova, K.; Srinivasan, P. P.; Zhou, H.; Barron, J. T.; Martin-Brualla, R.; Snavely, N.; and Funkhouser, T. 2021. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4690--4699

  32. [40]

    Wang, Z.; Shen, T.; Gao, J.; Huang, S.; Munkberg, J.; Hasselgren, J.; Gojcic, Z.; Chen, W.; and Fidler, S. 2023. Neural Fields meet Explicit Geometric Representations for Inverse Rendering of Urban Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patter...

  33. [41]

    P.; Barron, J

    Weng, C.-Y.; Curless, B.; Srinivasan, P. P.; Barron, J. T.; and Kemelmacher-Shlizerman, I. 2022. Humannerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, 16210--16220

  34. [42]

    P.; Curless, B.; and Kemelmacher-Shlizerman, I

    Weng, C.-Y.; Srinivasan, P. P.; Curless, B.; and Kemelmacher-Shlizerman, I. 2023. PersonNeRF: Personalized Reconstruction from Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 524--533

  35. [43]

    Wu, G.; Yi, T.; Fang, J.; Xie, L.; Zhang, X.; Wei, W.; Liu, W.; Tian, Q.; and Wang, X. 2023. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528

  36. [44]

    Wu, X.; Dai, P.; Deng, W.; Chen, H.; Wu, Y.; Cao, Y.-P.; Shan, Y.; and Qi, X. 2024. CL-NeRF: Continual Learning of Neural Radiance Fields for Evolving Scene Representation. Advances in Neural Information Processing Systems, 36

  37. [45]

    Xu, L.; Xiangli, Y.; Peng, S.; Pan, X.; Zhao, N.; Theobalt, C.; Dai, B.; and Lin, D. 2023. Grid-guided Neural Radiance Fields for Large Urban Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8296--8306

  38. [46]

    Yang, Y.; Cui, Z.; Xu, J.; Zhong, C.; Zheng, W.-S.; and Wang, R. 2023. Continual learning with Bayesian model based on a fixed pre-trained feature extractor. Visual Intelligence, 1(1): 5

  39. [47]

    Yu, T.; Zheng, Z.; Guo, K.; Liu, P.; Dai, Q.; and Liu, Y. 2021. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5746--5756

  40. [48]

    Zhang, H.; Feng, Y.; Kulits, P.; Wen, Y.; Thies, J.; and Black, M. J. 2023 a . Text-guided generation and editing of compositional 3d avatars. arXiv preprint arXiv:2309.07125

  41. [49]

    Zhang, K.; Riegler, G.; Snavely, N.; and Koltun, V. 2020. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492

  42. [50]

    Zhang, L.; Li, M.; Chen, C.; and Xu, J. 2023 b . IL-NeRF: Incremental Learning for Neural Radiance Fields with Camera Pose Alignment. arXiv preprint arXiv:2312.05748

  43. [51]

    Zhao, F.; Yang, W.; Zhang, J.; Lin, P.; Zhang, Y.; Yu, J.; and Xu, L. 2022. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7743--7753

  44. [52]

    Zhou, S.; Chang, H.; Jiang, S.; Fan, Z.; Zhu, Z.; Xu, D.; Chari, P.; You, S.; Wang, Z.; and Kadambi, A. 2024. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

  45. [53]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  46. [54]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.