Pith. sign in

REVIEW 3 major objections 5 minor 99 references

This paper claims that dynamic human avatars can be generated by diffusing over the weights of per-person pose-to-Gaussian UNets, unifying person-specific photorealism with generative diversity.

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 ↗

A diffusion model over per-person UNet weights generates new dynamic human avatars that render pose-dependent 3D Gaussians in real time.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A plausible new application of weight-space diffusion with a load-bearing flaw: the per-person UNet's poor generalization to unseen poses, admitted in Sec 7, undermines the controllability claim. the 3 major comments →

arxiv 2509.04145 v2 pith:4MUB7BJN submitted 2025-09-04 cs.GR cs.CV

Hyper Diffusion Avatars: Dynamic Human Avatar Generation using Network Weight Space Diffusion

Dongliang Cao , Guoxing Sun , Marc Habermann , Florian Bernard This is my paper
classification cs.GR cs.CV
keywords network weight diffusiondynamic human avatar generation3D Gaussian Splattingpose-dependent deformationhyper diffusiontransformer diffusionMVHumanNetUV-space Gaussian maps
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.

The reading

Most photorealistic human avatars are person-specific: one model per person, trained for days from multi-view video. Generative alternatives generalize across people but produce static or cartoonish avatars moved only by skeleton skinning, so they miss pose-dependent cloth wrinkles and shadows. This paper argues that both problems can be solved by moving the generative model into network weight space: first overfit a lightweight UNet per person that turns a skeletal pose into 3D Gaussian parameters, then train a transformer-based diffusion model over the collection of weight vectors. At inference, sampling the diffusion model yields the weights of a new, never-trained avatar network that can be posed and rendered in real time. On 500 MVHumanNet sequences, the generated avatars beat the state-of-the-art static avatar generators on all reported quality and diversity metrics.

Core claim

The central claim is that a dynamic human avatar can be encoded entirely in the weights of a small pose-conditioned network, and that the space of such weight vectors is learnable by a diffusion model. The paper's pipeline first trains, for each identity, a UNet that maps pose-dependent normal and position UV textures to a UV map of 3D Gaussian parameters; this network captures pose-dependent deformations such as clothing wrinkles. All these per-person networks share the same architecture, a shared SMPL-X UV space, and a consistent weight initialization, so their weights form a dataset of 'avatar networks.' A transformer-based hyper diffusion model is then trained to denoise these weight vec

What carries the argument

The load-bearing object is the network weight space of the per-person UNet: a diffusion model operates on the weights themselves rather than on images, latent codes, or static 3D representations. Each avatar is the function U_w(N_uv(theta), P_uv(theta)) = G_uv(theta), so the roughly 0.6M weights of this UNet compactly encode the full pose-dependent Gaussian map of one person. The hyper diffusion model is a transformer that tokenizes network weights layer by layer — one token per convolutional kernel-plus-bias — and predicts denoised weights, preserving the network's hierarchical structure instead of flattening or embedding the weights in a latent space.

Load-bearing premise

The load-bearing premise is that a per-person UNet, trained for 700k iterations on one individual's multi-view frames, maps both seen and unseen skeletal poses to valid 3D Gaussian parameters; the paper itself reports limited generalization to unseen poses, so a sampled weight vector may not yield a freely poseable avatar.

What would settle it

Take several weight vectors sampled from the hyper diffusion and render them under a held-out motion sequence far outside the training pose distribution of the source identities; if most sampled avatars produce distorted or collapsed bodies, the method's controllable-dynamic-avatar claim fails. The underlying UNet's poor unseen-pose generalization, acknowledged in the paper, makes this end-to-end test decisive.

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

If this is right

  • A sampled weight vector is a complete dynamic avatar: no further optimization or per-instance fitting is needed, only feeding the desired pose into the generated UNet and rasterizing the resulting Gaussians.
  • Because the pose enters the UNet as input, rendered results inherit pose-dependent deformations such as cloth wrinkles and cast shadows, which static-representation generative methods cannot express.
  • The layer-wise tokenization result indicates that respecting the internal structure of the network is what makes weight-space diffusion tractable, giving a recipe for future per-instance neural representations.
  • The reported metric improvements — higher coverage, lower 1-NNA, FID, and KID than PrimDiffusion and E3Gen — support that the generated weight distribution covers real identities without collapsing to a few modes.

Where Pith is reading between the lines

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

  • If the learned weight distribution is genuinely smooth, conditioning the hyper diffusion on attributes such as identity, text, or a pose sequence should be a natural next step; the paper trains only unconditionally, but a conditional version would make sampled avatars steerable at generation time.
  • The paper's own limitation — UNets generalize poorly to unseen poses — implies that sampled avatars inherit the pose coverage of the training videos; a direct evaluation that re-targets sampled avatars onto held-out pose sequences would separate 'novel avatar' quality from 'novel pose' robustness.
  • The success of layer-wise tokens over flattened vectors suggests architectural symmetry matters; aligning away permutation symmetries or generating low-rank weight updates could improve sample diversity and scale.
  • A practical bottleneck is the per-person 700k-iteration overfit stage; extending the idea to amortized weight prediction or meta-learning would be the difference between a proof of concept and a deployable avatar generator.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper proposes a two-stage generative model for dynamic human avatars. In Stage 1, a lightweight UNet is trained per person to map SMPL-X pose inputs (encoded as UV-space normal and position maps) to UV-space 3D Gaussian parameters, thereby capturing pose-dependent deformations (Eq. 6). In Stage 2, a transformer-based diffusion model is trained over the flattened weights of these per-person UNets. At inference, sampling the hyper-diffusion yields new network weights, which are then used to render controllable, real-time avatars. The method is evaluated on 500 manually selected sequences from MVHumanNet and compared against PrimDiffusion and E3Gen using MMD, COV, 1-NNA, FID, and KID, reporting favorable numbers on all metrics.

Significance. If the central claim holds, this is a novel and potentially impactful combination: it unifies per-person photorealistic avatar rendering with generative diffusion by operating in network-weight space, avoiding the need to generate raw 3D Gaussian parameters while preserving pose-dependent deformations. The layer-wise tokenization of network weights is a sensible design choice, and the shared UV space across identities is a useful contribution. The paper also deserves credit for being explicit about its limitations in Sec. 7. However, the evidence is currently incomplete: the load-bearing pose-generalization assumption is acknowledged as weak, the quantitative evaluation lacks error bars and a clear data split, and the comparison set is too narrow to support the claimed superiority over 'state-of-the-art human avatar generation.' The potential is real, but the validation needs substantial strengthening.

major comments (3)
  1. [Sec. 7 and Eq. (6)] The paper's own limitation statement in Sec. 7 — 'the UNet shows limited generalization to unseen poses' — targets the load-bearing assumption of the entire method. Eq. (6) defines the per-person UNet as the map that turns any pose θ into valid Gaussian parameters, and inference uses a sampled U_w exactly this way to obtain a controllable dynamic avatar. Table 1 evaluates generation with image-based metrics, but it does not report whether the poses used for rendering generated avatars are inside or outside the Stage-1 training poses of the corresponding identities. A generated weight vector is therefore not demonstrated to produce an avatar that can be posed freely; it may only replay near-training poses. Please add a held-out-pose experiment: split each subject's sequence by pose (e.g., time intervals or clustering of θ), train Stage-1 on one portion, and report PSNR/LPIPS/mask on held-
  2. [Sec. 5.1 and Tab. 1] The quantitative evaluation lacks the information needed to interpret the superiority claims. No error bars or number of seeds are reported for any metric in Table 1 or Table 2, so it is impossible to tell whether the margins (e.g., FID 12.68 vs. 32.17) are significant. Moreover, the paper does not state the split between the identities/sequences used to train the per-person UNets and the diffusion model and the reference set Sr used for MMD/COV/1-NNA/FID/KID. If Sr contains the same identities and poses used in Stage 1, the metrics can be inflated by memorization. Please specify the exact training/reference split, report statistics over at least three seeds, and describe the 500-sequence selection criterion (currently 'manually selected ... based on the SMPL-X pose parameter estimation accuracy') so that selection bias can be assessed.
  3. [Sec. 5.2] The comparison is limited to two static avatar generators (PrimDiffusion, E3Gen), both of which use LBS for articulation. The central claim is about pose-dependent deformations and controllability, yet no baseline that models pose-dependent deformations is compared, and no quantitative comparison of deformation quality (e.g., wrinkle fidelity, pose-transfer accuracy) is provided. The qualitative example in Fig. 6 is a single sequence. Please include either a dynamic-avatar baseline or, at minimum, a quantitative pose-transfer experiment rendering generated avatars on held-out poses and comparing to ground truth or to a person-specific dynamic renderer such as ASH [50] under the same protocol.
minor comments (5)
  1. [Eq. (7)] The barycentric quantities a_a, a_b, a_c and the canonical vertex positions vbar are not defined in the text. Defining them would make the LBS step reproducible.
  2. [Sec. 4.2] The exact construction of tokens in Eq. (10)-(13) should be clarified: how the bias is concatenated to the flattened kernel, whether all layers have the same token dimension after projection, and how zero-padding handles layers of different sizes.
  3. [Sec. 4.3] The '30k-step warm-up' is mentioned but not described. Please state what is warmed up (e.g., learning rate) and the schedule.
  4. [Sec. 5.1] The definition of 1-NNA can be ambiguous when there are ties in the nearest-neighbor computation. A short clarification would help.
  5. [Fig. 4] Indicate which pose is used for the renderings at different noise levels and whether the same pose is used across columns; otherwise the visual denoising story is hard to interpret.

Circularity Check

0 steps flagged

No significant circularity: the two-stage weight-space diffusion pipeline is self-contained, and the evaluation is external; the unseen-pose limitation is an acknowledged soundness issue, not definitional circularity.

full rationale

The paper's derivation chain has three load-bearing pieces: (1) per-subject UNets are optimized against multi-view images via Eq. (9), so each weight vector w_i is fitted to image data; (2) a transformer diffusion model is trained with an MSE loss on the layer-wise tokens of those weights (Eqs. 10-13); (3) at inference, DDIM samples new weight vectors that are rendered through Eqs. (6)-(8). None of these steps defines its output in terms of the evaluation metric or fits a parameter that is then relabeled as a prediction. The hyper-diffusion is trained on the same 500-subject MVHumanNet collection that is later used as the reference distribution for MMD/COV/1-NNA/FID/KID, but that is the standard protocol for unconditional generation and does not make the sampled weights 'forced by construction': the diffusion must still learn the weight distribution, and the comparison against PrimDiffusion and E3Gen is external. The only self-citation in the paper is [12] (Bernard) in Sec. 7, where it is used to name neural permutation symmetry as a challenge for future work; it is not load-bearing for any central claim. The Sec. 7 concession that 'the UNet shows limited generalization to unseen poses' is a genuine limitation that threatens the controllability promise, but it is a soundness/correctness issue, not circularity: the paper does not redefine unseen-pose rendering in terms of the training poses, nor does it secretly fit the evaluation metric. No circular step can be exhibited, so the circularity score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method's scientific content is an empirical system; the listed free parameters and domain assumptions are the hand-set choices and unproven premises the central claim rests on. None of the entries has independent evidence beyond the paper's own qualitative results.

free parameters (5)
  • loss weights lambda_pix=1.0, lambda_str=0.1, lambda_per=0.01, lambda_m=0.1 = 1.0, 0.1, 0.01, 0.1
    Set empirically (Sec 4.3) to balance L1, SSIM, LPIPS, and mask losses in Eq. 9; changes affect per-person UNet quality and hence the weight distribution being diffused.
  • UNet hidden channels = 64
    Empirically reduced from the original UNet's 30M params to 0.6M params (Sec 4.3); this capacity choice directly shapes the weight space the diffusion model must learn.
  • transformer depth/width = 12 blocks, 16 heads, hidden 2048
    Model size for the weight-space diffusion (Sec 4.3); chosen empirically, affects the diffusion model's capacity.
  • UV map resolution = 256 x 256
    Resolution of the UV Gaussian maps (Sec 4.3); trades efficiency against the fidelity of pose-dependent Gaussian parameters.
  • weight standardization = zero mean, unit variance
    Preprocessing of network weights before diffusion (Sec 4.3); a normalization choice that determines the target distribution and the unstandardization step at sampling.
axioms (5)
  • domain assumption The mean SMPL-X template (beta=0, psi=0) is a sufficient canonical geometry for all identities, providing a shared UV space and LBS skinning across individuals.
    Used in Sec 4.1 to define input textures N_uv and P_uv and output Gaussians G_uv; if the mean template poorly matches some body shapes, the per-person UNet capacity must compensate or avatars will be distorted.
  • domain assumption Each per-person UNet, trained through Eq. 9 on that person's multi-view frames, learns a valid pose-to-Gaussian map that supports pose-dependent deformation (Eq. 6).
    Stage-1 training is the source of the 'dynamic avatar'; Sec 7 admits generalization to unseen poses is limited, so the assumption is partially false in practice.
  • domain assumption Layer-wise tokenization preserves cross-identity weight alignment, i.e., corresponding kernels across different UNets occupy the same token positions.
    Sec 4.2 partitions weights by layer; permutation symmetry (mentioned as a limitation in Sec 7) means the alignment is not guaranteed, so the diffusion model must learn a distribution over one particular permutation of each network.
  • domain assumption DDIM sampling on standardized weight vectors yields valid UNet weights after unstandardization.
    Sec 4.2-4.3 rely on standard diffusion practice; no analysis of mode collapse or invalid samples is given in the main evaluation.
  • domain assumption MVHumanNet SMPL-X pose estimates for the manually selected 500 sequences are accurate enough to serve as conditioning inputs.
    Sec 5.1 selects sequences 'based on the SMPL-X pose parameter estimation accuracy'; errors in pose would propagate into the learned pose-to-Gaussian mapping.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Hyper Diffusion Avatars: Dynamic Human Avatar Generation using Network Weight Space Diffusion." pith.science (2026). https://pith.science/paper/4MUB7BJN

@misc{pith2026250904145,
  author       = {Pith},
  title        = {Pith review of: Hyper Diffusion Avatars: Dynamic Human Avatar Generation using Network Weight Space Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4MUB7BJN}},
  note         = {Machine review of arXiv:2509.04145}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Creating human avatars is a highly desirable yet challenging task. Recent advancements in radiance field rendering have achieved unprecedented photorealism and real-time performance for personalized dynamic human avatars. However, these approaches are typically limited to person-specific rendering models trained on multi-view video data for a single individual, limiting their ability to generalize across different identities. On the other hand, generative approaches leveraging prior knowledge from pre-trained 2D diffusion models can produce cartoonish, static human avatars, which are animated through simple skeleton-based articulation. Therefore, the avatars generated by these methods suffer from lower rendering quality compared to person-specific rendering methods and fail to capture pose-dependent deformations such as cloth wrinkles. In this paper, we propose a novel approach that unites the strengths of person-specific rendering and diffusion-based generative modeling to enable dynamic human avatar generation with both high photorealism and realistic pose-dependent deformations. Our method follows a two-stage pipeline: first, we optimize a set of person-specific UNets, with each network representing a dynamic human avatar that captures intricate pose-dependent deformations. In the second stage, we train a hyper diffusion model over the optimized network weights. During inference, our method generates network weights for real-time, controllable rendering of dynamic human avatars. Using a large-scale, cross-identity, multi-view video dataset, we demonstrate that our approach outperforms state-of-the-art human avatar generation methods.

Figures

Figures reproduced from arXiv: 2509.04145 by Dongliang Cao, Florian Bernard, Guoxing Sun, Marc Habermann.

Figure 1
Figure 1. Figure 1: Our method enables dynamic human avatar generation via diffusion in network weight space. First, we optimize a set of UNets, each representing an individual dynamic human avatar (top left). Next, we train a transformer network to model a diffusion process over these optimized network weights (top right). At inference time, our approach samples new network weights for real-time, controllable dynamic human a… view at source ↗
Figure 2
Figure 2. Figure 2: Dynamic human representation learning based on UNet. Given a specific human pose, the pose-dependent position and normal maps are generated via inverse texture mapping. These maps serve as inputs to the UNet, which predicts pose-dependent 3D Gaussians for rendering. During training, the UNet is optimized using multi-view RGB image sequences along with their corresponding segmentation masks. Gaussian is par… view at source ↗
Figure 3
Figure 3. Figure 3: Diffusion process on network weight space. During the forward diffusion process, the standard Gaussian noise at time step t is added to the network weights and the transformer take the noisy weights as well as the time step t to predict the denoised weights. For each camera view k with projection matrix πk, the resulting 3D Gaussians Guv are rendered using a differ￾entiable Gaussian rasterizer R, producing… view at source ↗
Figure 4
Figure 4. Figure 4: Denoising network weights at various time steps. Here the network weights are visualized based on the rendering images. The rendering images show the reverse diffusion process based on DDIM sampling. We observe that UNet weights corrupted by noise fail to represent a valid human avatar. However, the iterative denoising process yields a high-quality human avatar. model [PITH_FULL_IMAGE:figures/full_fig_p00… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results on unconditional human avatar generation. Compared to baseline methods, our method is able to generate more photorealistic human avatars. Methods MMDPSNR ↑ MMDLPIPS ↓ COVPSNR(%) ↑ 1-NNAPSNR(%) ↓ FID ↓ KID ↓ PrimDiffusion [10] 22.23 26.45 52.3 26.8 41.97 328.46 E3Gen [93] 21.14 32.28 58.2 21.3 32.17 284.31 Ours 27.52 12.13 63.8 15.7 12.68 123.26 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: An example of unconditional human avatar generation of our method. Rendering sequence demonstrates our method’s ability to generate dynamic human avatars. Pose-dependent deformations are emphasized with green rectangles. MMDPSNR ↑ MMDLPIPS ↓ COVPSNR(%) ↑ 1-NNAPSNR(%) ↓ FID ↓ KID ↓ Latent diffusion 21.23 27.28 0.4 98.0 58.52 480.34 1D vector flatten 27.12 12.30 54.2 27.6 14.73 134.65 Ours 27.52 12.13 63.8 1… view at source ↗
Figure 7
Figure 7. Figure 7: Latent diffusion model on network weight space. The network weights are first reshaped into a 2D feature map. An en￾coder then converts this 2D feature map into a latent space repre￾sentation. The diffusion process takes place on the latent space. Afterward, a decoder transforms the latent features back into a 2D feature map. Finally, the network weights are recovered by reshap￾ing the 2D feature map. used… view at source ↗

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

99 extracted references · 69 canonical work pages

  1. [1]

    Generative neural articulated radiance fields

    Alexander W Bergman, Petr Kellnhofer, Wang Yifan, Eric R Chan, David B Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. In Advances in Neural In- formation Processing Systems, 2022. 2

    2022

  2. [2]

    Multi-scale capture of facial geometry and motion

    Bernd Bickel, Mario Botsch, Roland Angst, Wojciech Ma- tusik, Miguel Otaduy, Hanspeter Pfister, and Markus Gross. Multi-scale capture of facial geometry and motion. ACM Transactions on Graphics (ToG), 26(3):33–es, 2007. 2

    2007

  3. [3]

    Demystifying mmd gans

    Mikołaj Bi ´nkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations, 2018. 6

    2018

  4. [4]

    Sampling weights of deep neural networks

    Erik L Bolager, Iryna Burak, Chinmay Datar, Qing Sun, and Felix Dietrich. Sampling weights of deep neural networks. In Advances in Neural Information Processing Systems , 2023. 8

    2023

  5. [5]

    Hexplane: A fast representa- tion for dynamic scenes

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ,

  6. [6]

    Authen- tic volumetric avatars from a phone scan

    Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shoou-I Yu, et al. Authen- tic volumetric avatars from a phone scan. ACM Transactions on Graphics (ToG), 41(4):1–19, 2022. 5

    2022

  7. [7]

    Dreamavatar: Text-and-shape guided 3d hu- man avatar generation via diffusion models

    Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and Kwan- Yee K Wong. Dreamavatar: Text-and-shape guided 3d hu- man avatar generation via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 2024. 2, 3

    2024

  8. [8]

    Efficient geometry-aware 3d generative adversarial networks

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

    2022

  9. [9]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3

    Pith/arXiv arXiv 2015

  10. [10]

    Primdiffusion: V olumetric primitives diffusion for 3d human generation

    Zhaoxi Chen, Fangzhou Hong, Haiyi Mei, Guangcong Wang, Lei Yang, and Ziwei Liu. Primdiffusion: V olumetric primitives diffusion for 3d human generation. In Advances in Neural Information Processing Systems, 2023. 3, 6, 7

    2023

  11. [11]

    Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation

    Matt Cordner and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 811–818. Association for Com- puting Machinery, 2023. 3, 6

    2023

  12. [12]

    Cycle-consistent multi- model merging

    Donato Crisostomi, Marco Fumero, Daniele Baieri, Florian Bernard, and Emanuele Rodola. Cycle-consistent multi- model merging. In Advances in Neural Information Process- ing Systems, 2024. 8

    2024

  13. [13]

    Obja- verse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Obja- verse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 3

    Pith/arXiv arXiv 2022

  14. [14]

    Objaverse-xl: A universe of 10m+ 3d objects

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Chris- tian Laforte, Vikram V oleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl V ondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objaverse-xl: A universe of 10m+ 3d objects. arXiv preprint arXiv:2307.05663, 2023. 3

    Pith/arXiv arXiv 2023

  15. [15]

    Hyperdiffusion: Generating implicit neu- ral fields with weight-space diffusion

    Ziya Erkoc ¸, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neu- ral fields with weight-space diffusion. In Proceedings of the IEEE/CVF international conference on computer vision,

  16. [16]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 2, 3

    2024

  17. [17]

    Capturing and animation of body and clothing from monocular video

    Yao Feng, Jinlong Yang, Marc Pollefeys, Michael J Black, and Timo Bolkart. Capturing and animation of body and clothing from monocular video. In SIGGRAPH Asia, 2022. 2

    2022

  18. [18]

    Reconstructing detailed dynamic face geome- try from monocular video

    Pablo Garrido, Levi Valgaerts, Chenglei Wu, and Christian Theobalt. Reconstructing detailed dynamic face geome- try from monocular video. ACM Transactions on Graphics (ToG), 32(6):158–1, 2013. 2

    2013

  19. [19]

    Hdhu- mans: A hybrid approach for high-fidelity digital humans

    Marc Habermann, Lingjie Liu, Weipeng Xu, Gerard Pons- Moll, Michael Zollhoefer, and Christian Theobalt. Hdhu- mans: A hybrid approach for high-fidelity digital humans. Proceedings of the ACM on Computer Graphics and Inter- active Techniques, 6(3):1–23, 2023. 2

    2023

  20. [20]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. In Advances in Neural Information Processing Sys- tems, 2017. 6

    2017

  21. [21]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In Advances in Neural Informa- tion Processing Systems, 2020. 2, 4, 5

    2020

  22. [22]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In In- ternational Conference on Learning Representations , 2022. 8

    2022

  23. [23]

    Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians

    Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 2

    work page 2024

  24. [24]

    Sherf: Generalizable human nerf from a single image

    Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2023. 2 9

    work page 2023

  25. [25]

    Gauhuman: Articu- lated gaussian splatting from monocular human videos

    Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articu- lated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2024. 2

    work page 2024

  26. [26]

    Tech: Text-guided reconstruction of lifelike clothed humans

    Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji- axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In International Conference on 3D Vision (3DV), 2024. 2, 3

    work page 2024

  27. [27]

    Speeding up convolutional neural networks with low rank expansions

    Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. arXiv preprint arXiv:1405.3866, 2014. 8

    Pith/arXiv arXiv 2014

  28. [28]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):139–1, 2023. 2, 3, 5

    work page 2023

  29. [29]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    Pith/arXiv arXiv

  30. [30]

    Dreamhuman: Animatable 3d avatars from text

    Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Ed- uard Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. InAdvances in Neural Information Processing Systems, 2023. 2

    work page 2023

  31. [31]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. In Advances in Neural Information Processing Sys- tems, 2012. 5, 6

    work page 2012

  32. [32]

    Neural mechan- ics: Symmetry and broken conservation laws in deep learn- ing dynamics

    Daniel Kunin, Javier Sagastuy-Brena, Surya Ganguli, Daniel LK Yamins, and Hidenori Tanaka. Neural mechan- ics: Symmetry and broken conservation laws in deep learn- ing dynamics. In International Conference on Learning Rep- resentations, 2020. 8

    work page 2020

  33. [33]

    Deliffas: Deformable light fields for fast avatar synthesis

    Youngjoong Kwon, Lingjie Liu, Henry Fuchs, Marc Haber- mann, and Christian Theobalt. Deliffas: Deformable light fields for fast avatar synthesis. In Advances in Neural Infor- mation Processing Systems, 2023. 2

    work page 2023

  34. [34]

    Gen- eralizable human gaussians for sparse view synthesis

    Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella- Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, et al. Gen- eralizable human gaussians for sparse view synthesis. In Eu- ropean Conference on Computer Vision. Springer, 2024. 4

    work page 2024

  35. [35]

    Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation

    JP Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th an- nual conference on Computer graphics and interactive tech- niques, pages 165–172, 2000. 3, 4

    work page 2000

  36. [36]

    Tava: Template-free animatable volumetric actors

    Ruilong Li, Julian Tanke, Minh V o, Michael Zollh ¨ofer, J¨urgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In Eu- ropean Conference on Computer Vision, 2022. 2

    work page 2022

  37. [37]

    Ani- matable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling

    Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Ani- matable 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, 2024. 2, 4

    work page 2024

  38. [38]

    Tada! text to animatable digital avatars

    Tingting Liao, Hongwei Yi, Yuliang Xiu, Jiaxiang Tang, Yangyi Huang, Justus Thies, and Michael J Black. Tada! text to animatable digital avatars. In International Conference on 3D Vision (3DV), 2024. 2, 3

    work page 2024

  39. [39]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matthew Le. Flow matching for generative modeling. In International Conference on Learning Repre- sentations, 2022. 2

    work page 2022

  40. [40]

    Neural actor: Neural free-view synthesis of human actors with pose con- trol

    Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose con- trol. ACM Transactions on Graphics (ToG) , 40(6):1–16,

    work page

  41. [41]

    Humangaus- sian: Text-driven 3d human generation with gaussian splat- ting

    Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaus- sian: Text-driven 3d human generation with gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024. 2, 3

    work page 2024

  42. [42]

    Mix- ture of volumetric primitives for efficient neural rendering

    Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mix- ture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG), 40(4):1–13, 2021. 3

    work page 2021

  43. [43]

    Wonder3d: Sin- gle image to 3d using cross-domain diffusion

    Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Sin- gle image to 3d using cross-domain diffusion. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3

    work page 2024

  44. [44]

    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. ACM Transactions on Graphics (ToG), 34(6), 2015. 2

    work page 2015

  45. [45]

    Diffusion probabilistic models for 3d point cloud generation

    Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition ,

    work page

  46. [46]

    On distillation of guided diffusion models

    Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 2023. 3

    work page 2023

  47. [47]

    Latent-nerf for shape-guided generation of 3d shapes and textures

    Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition ,

    work page

  48. [48]

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

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 2

    2021

  49. [49]

    Gsd: View-guided gaussian splatting diffusion for 3d reconstruction

    Yuxuan Mu, Xinxin Zuo, Chuan Guo, Yilin Wang, Juwei Lu, Xiaofeng Wu, Songcen Xu, Peng Dai, Youliang Yan, and Li Cheng. Gsd: View-guided gaussian splatting diffusion for 3d reconstruction. In European Conference on Computer Vision, 2024. 3

    work page 2024

  50. [50]

    Ash: Animatable gaussian splats for efficient and photoreal human rendering

    Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6 10

    work page 2024

  51. [51]

    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. In Proceedings of the IEEE/CVF international conference on computer vision,

    work page

  52. [52]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF interna- tional conference on computer vision, 2023. 7

    work page 2023

  53. [53]

    Learning to learn with genera- tive models of neural network checkpoints

    William Peebles, Ilija Radosavovic, Tim Brooks, Alexei A Efros, and Jitendra Malik. Learning to learn with genera- tive models of neural network checkpoints. arXiv preprint arXiv:2209.12892, 2022. 5

    Pith/arXiv arXiv 2022

  54. [54]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3

    Pith/arXiv arXiv 2023

  55. [55]

    Dreamfusion: Text-to-3d using 2d diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. In Inter- national Conference on Learning Representations, 2022. 2, 3

    work page 2022

  56. [56]

    Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors

    Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Sko- rokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In Interna- tional Conference on Learning Representations, 2024. 3

    work page 2024

  57. [57]

    Language models are unsu- pervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsu- pervised multitask learners. OpenAI blog, 1(8):9, 2019. 4

    work page 2019

  58. [58]

    Drivable volumet- ric avatars using texel-aligned features

    Edoardo Remelli, Timur Bagautdinov, Shunsuke Saito, Chenglei Wu, Tomas Simon, Shih-En Wei, Kaiwen Guo, Zhe Cao, Fabian Prada, Jason Saragih, et al. Drivable volumet- ric avatars using texel-aligned features. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1–9, 2022. 2

    work page 2022

  59. [59]

    Dreamgaussian4d: Genera- tive 4d gaussian splatting

    Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Genera- tive 4d gaussian splatting. arXiv preprint arXiv:2312.17142,

    Pith/arXiv arXiv

  60. [60]

    L3dg: Latent 3d gaussian diffusion

    Barbara Roessle, Norman M ¨uller, Lorenzo Porzi, Samuel Rota Bul `o, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024. 3

    work page 2024

  61. [61]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2, 3, 7

    work page 2022

  62. [62]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In Medical image computing and computer-assisted intervention. Springer, 2015. 2, 4, 6, 7

    work page 2015

  63. [63]

    Relightable gaussian codec avatars

    Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2024. 2

    work page 2024

  64. [64]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, 2024. 3

    work page 2024

  65. [65]

    Holo- ported characters: Real-time free-viewpoint rendering of humans from sparse rgb cameras

    Ashwath Shetty, Marc Habermann, Guoxing Sun, Diogo Lu- vizon, Vladislav Golyanik, and Christian Theobalt. Holo- ported characters: Real-time free-viewpoint rendering of humans from sparse rgb cameras. In Proceedings of the IEEE/CVF international conference on computer vision ,

    work page

  66. [66]

    Zero123++: a single image to consistent multi-view dif- fusion base model

    Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view dif- fusion base model. arXiv preprint arXiv:2310.15110, 2023. 3

    Pith/arXiv arXiv 2023

  67. [67]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Confer- ence on Machine Learning, 2015. 4

    work page 2015

  68. [68]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In International Conference on Learning Representations, 2020. 2, 4, 5

    work page 2020

  69. [69]

    Generative modeling by es- timating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by es- timating gradients of the data distribution. In Advances in Neural Information Processing Systems, 2019. 4

    work page 2019

  70. [70]

    Diffusion-based neural network weights generation

    Bedionita Soro, Bruno Andreis, Hayeon Lee, Wonyong Jeong, Song Chong, Frank Hutter, and Sung Ju Hwang. Diffusion-based neural network weights generation. In In- ternational Conference on Learning Representations , 2025. 3, 5, 7

    work page 2025

  71. [71]

    Real-time free- view human rendering from sparse-view rgb videos us- ing double unprojected textures

    Guoxing Sun, Rishabh Dabral, Heming Zhu, Pascal Fua, Christian Theobalt, and Marc Habermann. Real-time free- view human rendering from sparse-view rgb videos us- ing double unprojected textures. In Proceedings of the IEEE/CVF international conference on computer vision ,

    work page

  72. [72]

    Rethinking the inception archi- tecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception archi- tecture for computer vision. InProceedings of the IEEE/CVF international conference on computer vision, 2016. 6

    work page 2016

  73. [73]

    Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior

    Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, 2023. 3

    work page 2023

  74. [74]

    Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation. In International Conference on Learning Representations, 2024. 3

    work page 2024

  75. [75]

    Gaus- sianheads: End-to-end learning of drivable gaussian head avatars from coarse-to-fine representations

    Kartik Teotia, Hyeongwoo Kim, Pablo Garrido, Marc Haber- mann, Mohamed Elgharib, and Christian Theobalt. Gaus- sianheads: End-to-end learning of drivable gaussian head avatars from coarse-to-fine representations. ACM Transac- tions on Graphics (ToG), 43(6):1–12, 2024. 2

    work page 2024

  76. [76]

    Disentangled3d: Learn- ing a 3d generative model with disentangled geometry and appearance from monocular images

    Ayush Tewari, Xingang Pan, Ohad Fried, Maneesh Agrawala, and Christian Theobalt. Disentangled3d: Learn- ing a 3d generative model with disentangled geometry and appearance from monocular images. In Proceedings of 11 the IEEE/CVF conference on computer vision and pattern recognition, 2022. 8

    work page 2022

  77. [77]

    Ren- derpeople, 2018

    https://renderpeople.com/3d-people/ . Ren- derpeople, 2018. 3

    work page 2018

  78. [78]

    Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation

    Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 3

    work page 2023

  79. [79]

    Neus: Learning neural im- plicit surfaces by volume rendering for multi-view recon- struction

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural im- plicit surfaces by volume rendering for multi-view recon- struction. In Advances in Neural Information Processing Systems, 2021. 2

    work page 2021

  80. [80]

    Neus2: Fast learning of neural implicit surfaces for multi-view recon- struction

    Yiming Wang, Qin Han, Marc Habermann, Kostas Dani- ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view recon- struction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2

    work page 2023

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.