Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

LIFe-GoM: Generalizable Human Rendering with Learned Iterative Feedback Over Multi-Resolution Gaussians-on-Mesh

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims a feed-forward system can turn as few as three images into an animatable avatar in under a second, rendering it at 95 FPS with quality that beats prior generalizable human rendering methods.

desk verdict LIFe-GoM is a solid, well-executed generalizable human rendering paper with strong speed/quality numbers; the iterative feedback claim is undercut by an ablation that never tests a true no-feedback baseline, but the empirical result stands and the paper deserves serious reviewing. read the letter →

arxiv 2502.09617 v1 pith:O2GNJQYR submitted 2025-02-13 cs.CV

Jing Wen , Alexander G. Schwing , Shenlong Wang This is my paper · ORCID
classification cs.CV
keywords generalizablehumanrenderinganimatableavatarsGaussians-on-Meshiterativefeedbackmulti-resolutionrepresentation3DGaussiansplattingsparse-viewreconstructionlinearblendskinning
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 sets out to prove that a single feed-forward pass need not be the limit for generalizable human avatars: it claims a system that rebuilds an animatable avatar from as few as three sparse images in 0.9 seconds, renders it from new viewpoints and poses at 95.1 FPS, and beats the prior best method on THuman2.0 (24.65 vs 21.90 PSNR). The central move is to let the network look at its own work — the reconstructor renders its current avatar back into the source views, compares with the real images, and refines the canonical 3D representation over three learned iterations, all trained end-to-end. A second move, coupled-multi-resolution Gaussians-on-Mesh, keeps that refinement cheap by updating only a low-resolution SMPL/SMPL-X mesh while rendering with many Gaussians attached to a subdivided high-resolution mesh. If the claims hold, personalized animatable avatars from a few snapshots would no longer require minutes-to-hours of per-subject optimization.

What carries the argument

The load-bearing object is the coupled-multi-resolution Gaussians-on-Mesh representation (GoMc): a low-resolution canonical mesh initialized from SMPL or SMPL-X and deformed by linear blend skinning, plus a subdivided high-resolution mesh carrying one Gaussian per triangle face, with each Gaussian's rotation, scale, color, and offset defined in the triangle's local frame. The second mechanism is the iterative feedback module: it renders the current avatar into the source views, extracts pixel-aligned features from both the real and the rendered images, fuses them across views, and encodes them with a Point Transformer to produce per-vertex updates that refine the mesh and Gaussian parameters at each step. Coupling the two keeps the expensive computation on the low-resolution mesh — updating the high-resolution vertices directly would inflate reconstruction from 0.9 s to 12.45 s — while rendering still benefits from 330K Gaussians at 10.52 ms per frame.

What would settle it

Reproduce the T=1 ablation on a held-out THuman2.0 split with all other components fixed: the paper reports PSNR dropping from 24.65 to 23.74 when the feedback loop is disabled, so a faithful implementation should recover roughly that gap and should still beat GHG's 21.90; if the single-pass version matches the full method, the iterative-feedback claim is falsified. A complementary check of the representational boundary is to render subjects in dresses or coats, which the paper's appendix predicts will fail.

Watch

Extended reading notes

Core claim

Positioned against single-pass feed-forward methods, the paper's central claim is that reconstruction quality improves when the canonical Gaussians-on-Mesh avatar is refined through an iterative feedback loop: at each of T steps the current avatar is articulated and rendered into the source views, a feedback network (multi-source fusion with transformer layers followed by a Point Transformer) compares pixel-aligned features from real and rendered images, and updates the low-resolution mesh vertices and the Gaussian parameters until, after three iterations, reconstruction finishes in 907.92 ms. The representation that makes this affordable is coupled-multi-resolution: the mesh that is deformed and updated stays low-resolution, while a twice-subdivided version of the same mesh carries 330K Gaussians whose positions and shapes are defined in each triangle's local coordinates, so rendering quality does not pay for reconstruction speed. On THuman2.0 with three source views the method reports PSNR/LPIPS*/FID of 24.65/110.82/51.27 against GHG's 21.90/133.41/61.67, improves on XHuman in a zero-fine-tuning cross-domain test, and on AIST++ matches per-scene-optimized baselines at 25.25/0.9812/21.61 while reconstructing in 589 ms.

Load-bearing premise

The canonical mesh is a deformed SMPL/SMPL-X template with fixed vertex connectivity, so the representation cannot fit clothing topology such as dresses or coats; a subject whose outfit violates the template topology cannot be represented, and rendering quality collapses on such cases.

Editorial extensions

If this is right

  • Sparse-input avatar creation becomes a one-shot operation: 3 to 5 images suffice, reconstruction takes less than one second, and the resulting avatar is cached and re-animated at 95.1 FPS at 1024×1024 resolution.
  • The method generalizes across datasets: a model trained only on THuman2.0 outperforms the prior state of the art on XHuman without fine-tuning, with PSNR/LPIPS*/FID of 25.32/99.32/42.90.
  • Because reconstruction happens in canonical T-pose space rather than in the input poses, novel poses can be synthesized directly with no post-processing such as skeleton binding.
  • Multi-frame monocular video inputs work as well as multiview image inputs: on AIST++ the method matches or exceeds per-scene-optimized baselines (25.25 PSNR, 0.9812 SSIM) while needing 589 ms instead of minutes to hours.
  • The iterative feedback loop carries a large share of the gain: going from one pass to three iterations raises PSNR from 23.74 to 24.65, at roughly 290 ms per additional iteration.

Reading between the lines

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

  • My reading: the render-and-compare loop is a portable recipe — any single-pass 3D predictor that has a differentiable renderer could be retrofitted with this self-correction mechanism at the cost of a few hundred milliseconds.
  • The saturation of gains by T=3 to T=5 suggests the loop fixes localized errors (silhouette, texture detail) rather than global topology, which fits the paper's own failure cases for dresses and coats.
  • The robustness experiments imply that pose accuracy is the tighter bottleneck than body shape, since adding pose noise of std 0.5 costs about 4.2 dB while substituting an average template shape costs only 0.5 dB.
  • A testable extension the authors leave implicit: feeding the feedback module rendered novel views in addition to source views, enforcing multi-view self-consistency, could further close the gap to per-scene optimization.
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

2 major / 5 minor

Summary. The paper presents LIFe-GoM, a generalizable human avatar rendering method that augments the Gaussians-on-Mesh representation with (i) an end-to-end iterative feedback reconstruction network and (ii) a coupled multi-resolution representation in which a low-resolution mesh is deformed while Gaussians are bound to a subdivided high-resolution mesh. The method reconstructs a canonical T-pose representation from three to five sparse source views in under one second and renders at 95 FPS at 1024x1024. It reports improved PSNR/LPIPS*/FID over GHG on THuman2.0 and XHuman and comparable quality to ActorsNeRF on AIST++, with ablation studies for the number of feedback steps and the mesh subdivision level.

Significance. Assuming the reported numbers are reproducible, the practical significance is clear: a feed-forward generalizable method that reconstructs in under one second and renders at interactive rates while improving on the previous state of the art is a useful step for avatar creation. The paper is transparent about implementation details, baseline setups, and failure cases, and it explicitly acknowledges the topology limitation of the SMPL/SMPL-X template in Section F. The main methodological weakness is that the ablation of the iterative feedback component (Table 4) does not include a true no-feedback control, so the causal attribution of the quality gains to feedback is not established. The cross-dataset evaluation on XHuman and the pose-noise and shape-sensitivity studies are valuable additions.

major comments (2)
  1. [Section 3.3, Eq. (11); Section 4.5, Table 4] The paper states that T=1 means iterative updates are disabled, but this is not a no-feedback baseline. At every step t, including t=1, the feedback feature is computed by rendering source views from GoM_{t-1} via Eq. (11) and comparing those renderings with the observed source images. Table 4 therefore varies the number of feedback refinement steps, not the presence or absence of feedback. To support the claim that the iterative feedback mechanism causes the improvement from 23.74 to 24.65 PSNR, the authors should add a single-pass control that predicts the canonical GoM directly from source image features without the rendered-image comparison branch, or otherwise ablate the render-and-compare branch at a fixed T.
  2. [Tables 1-3; Appendix D] All headline numbers are single-run results without error bars or significance tests. This is particularly consequential on AIST++, where the claimed advantage over ActorsNeRF is 25.25 vs. 25.23 PSNR and 0.9812 vs. 0.9809 SSIM. In addition, several baselines are retrained with substantially modified schedules (GoMAvatar for 100K rather than 200K iterations; iHuman for 150 rather than 15 epochs), so the 'state-of-the-art' claim needs either multiple seeds with variance reporting or a clear justification that these modified schedules are the correct comparison for the sparse-input setting.
minor comments (5)
  1. [Section 4.3; Table 3; Appendix E.2, Table 6] ActorsNeRF's LPIPS* is reported as 22.45 in Table 3, 22.11 in the Section 4.3 text, and 22.29 in the Appendix E.2 average table; these values should be reconciled.
  2. [Section 4.1] There is a typo: 'initilize' should be 'initialize'.
  3. [Section 5; Section F] The topology limitation described in Section F is a meaningful scope restriction for 'generalizable human rendering'; consider stating this limitation earlier in the paper so that the generality claim is not overbroad.
  4. [Section 4.5, Table 5] The claim that PSNR decreases at higher subdivision because 'PSNR sometimes prefers blurry results' is plausible but is not supported by any quantitative evidence; a sharpness or perceptual sharpness metric would make the explanation more convincing.
  5. [Appendix E.2] The spelling 'ActorNeRF' appears in the appendix text while the method and references use 'ActorsNeRF'; please make the spelling consistent.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the main benchmark results are measured on held-out subjects against external baselines; the notable T=1 ablation issue is a control gap, not a circular derivation.

full rationale

The central claims are not circular. The model is trained on a training split (e.g., 426 THuman2.0 subjects) and evaluated on held-out subjects (100 THuman2.0 subjects) and on XHuman/AIST++ without fine-tuning, against external baselines such as GHG, NHP, NIA, GPS-Gaussian, and ActorsNeRF. The base Gaussians-on-Mesh representation is adopted from the authors' own GoMAvatar, but GoMAvatar is also included as a baseline and is outperformed, so the self-citation is background rather than load-bearing. The iterative feedback and coupled-multi-resolution components are learned end-to-end with photometric losses, and no target metric is fitted to the test set. The most serious issue is in the ablation: the paper states 'T = 1 means a single feed-forward pass, i.e., iterative updates are disabled,' but by the paper's own Eq. 11 every step t, including t=1, renders source views from GoMc_{t-1} and uses that rendered-vs-source comparison as the feedback feature. Thus T=1 already includes one feedback-conditioned update, so the Table 4 comparison varies the number of feedback steps and total compute, not the presence or absence of feedback. This is a real experimental control gap for the causal attribution of the +0.91 PSNR gain, but it is a correctness/ablation concern rather than circularity: the final comparisons against external baselines remain independent and are not equivalent to the method's inputs by construction. The Appendix F topology/clothing limitation is explicitly acknowledged and does not bear on circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. It relies on standard learned network weights and a small set of hand-tuned hyperparameters, plus domain assumptions about input pose accuracy and SMPL topology.

free parameters (5)
  • Learned network weights (image encoder, fusion transformers, Point Transformer, Gaussian and mesh decoders) = Trained on THuman2.0 (200K iterations) and AIST++ (100K iterations)
    All prediction parameters are fitted to training data; they are the main fitted quantities of the method.
  • Iterative feedback steps T = 3
    Selected by ablation (Table 4); T=3 gives most of the PSNR gain (24.65 vs 23.74 at T=1) while keeping reconstruction under 1s.
  • Mesh subdivision level = 2
    Selected by ablation (Table 5); two subdivisions improve LPIPS/FID from 140.60/93.44 to 110.82/51.27, though PSNR slightly decreases.
  • Loss weights lambda_per, lambda_M, lambda_lap = 1.0, 5.0, 100 on THuman2.0; 1.0, 0, 100 on AIST++
    Hand-set hyperparameters in Eq. (12); they control the balance between RGB, mask, and Laplacian losses.
  • Number of Gaussians after subdivision = 330K
    Determined by the SMPL/SMPL-X template and two subdivision levels; used in the speed comparison with GHG's 2.8M Gaussians.
assumptions (4)
  • domain assumption Input human poses and camera parameters are provided and sufficiently accurate.
    Section 3.1 lists poses and cameras as given inputs; Section E.3 shows PSNR drops from 24.65 to 20.43 when pose noise std is 0.5, so the central result depends on this assumption.
  • domain assumption SMPL/SMPL-X template topology is a sufficient prior for the subjects being reconstructed.
    The canonical mesh is initialized from SMPL/SMPL-X (Section 4.1) and articulated via LBS; Section F states the representation cannot change vertex connectivity to fit dresses or coats.
  • ad hoc to paper Rendering the current reconstruction from source views and comparing it with observed images provides a useful refinement signal.
    This is the core premise of the iterative feedback module (Section 3.3). The ablation shows it helps, but no theoretical guarantee is given.
  • standard math Gaussian splatting with face-attached Gaussians is an effective differentiable renderer and animation model.
    Adopted from 3D Gaussian Splatting and GoMAvatar; used without modification in Sections 3.2.2 and 3.2.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LIFe-GoM: Generalizable Human Rendering with Learned Iterative Feedback Over Multi-Resolution Gaussians-on-Mesh." pith.science (2026). https://pith.science/paper/O2GNJQYR

@misc{pith2026250209617,
  author       = {Pith},
  title        = {Pith review of: LIFe-GoM: Generalizable Human Rendering with Learned Iterative Feedback Over Multi-Resolution Gaussians-on-Mesh},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O2GNJQYR}},
  note         = {Machine review of arXiv:2502.09617}
}
abstract

Generalizable rendering of an animatable human avatar from sparse inputs relies on data priors and inductive biases extracted from training on large data to avoid scene-specific optimization and to enable fast reconstruction. This raises two main challenges: First, unlike iterative gradient-based adjustment in scene-specific optimization, generalizable methods must reconstruct the human shape representation in a single pass at inference time. Second, rendering is preferably computationally efficient yet of high resolution. To address both challenges we augment the recently proposed dual shape representation, which combines the benefits of a mesh and Gaussian points, in two ways. To improve reconstruction, we propose an iterative feedback update framework, which successively improves the canonical human shape representation during reconstruction. To achieve computationally efficient yet high-resolution rendering, we study a coupled-multi-resolution Gaussians-on-Mesh representation. We evaluate the proposed approach on the challenging THuman2.0, XHuman and AIST++ data. Our approach reconstructs an animatable representation from sparse inputs in less than 1s, renders views with 95.1FPS at $1024 \times 1024$, and achieves PSNR/LPIPS*/FID of 24.65/110.82/51.27 on THuman2.0, outperforming the state-of-the-art in rendering quality.

Figures

Figures reproduced from arXiv: 2502.09617 by the authors.

Figure 1
Figure 1. Overview. We tackle the problem of generalizable human rendering. Given sparse source images (multiview images or multi-frame images), we reconstruct the 3D human representation in canonical T-pose space. The canonical representation can be animated and rendered in novel views. reasonably low-dimensional mesh representation is desirable for efficient reconstruction, while a high-dimensional Gaussian splat representa… view at source ↗
Figure 2
Figure 2. Multi-resolution Gaussians-on-Mesh representation. We use a low-res mesh for faster animation and simpler geometry and attach Gaus￾sians on a high-res mesh for better rendering. To achieve high-quality rendering, GoMAvatar subdivides the mesh to increase the number of Gaussians. However, in the generaliz￾able human rendering setting, naively subdi￾viding the mesh significantly increases the re￾construction time from… view at source ↗
Figure 3
Figure 3. Iterative feedback. We iteratively update in a feed-forward way the vertices of the low￾resolution mesh and the Gaussian parameters attached to the high-resolution mesh. We repeat the update for T steps. Each step t operates on the source images, camera parameters and human poses, as well as the last iteration’s results including the canonical representation GoMc t−1 and the predicted source images rendered by GoMc … view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Novel pose synthesis. Poses are from BEDLAM dataset. We compare our approach to GPS-Gaussian using five images. As GPS-Gaussian relies on depth prediction between adjacent views, five images are the minimum it needs. Despite that, it still fails in non-overlapping regi…
Figure 6
Figure 6. Figure 6: Ablation studies. We study the effect of iterative feedback (left). The geometry improves as the number of iterations increases. We show the importance of linking Gaussians to the high￾resolution mesh (right). The high-resolution mesh is subdivided from the low-resolut…
Figure 7
Figure 7. Figure 7: Iterative feedback module. The iterative feedback module takes as input the represen￾tation GoMc t−1 obtained from the previous iteration, the source images and images rendered with GoMc t−1 . The module is designed to compare the rendered images and source images, and…
Figure 8
Figure 8. Figure 8: Multi-source fusion. Multi-source fusion first samples the vertex-aligned image features from the encoded images. Then we use two Transformer encoder layers to fuse the information from each of the N source images. In the Transformer encoder layers, the three input arr…
Figure 9
Figure 9. Figure 9: Comparison to baselines in the setting of 3 source images on THuman2.0. Our method produces less noise than iHuman and GoMAvatar, and more accurate geometry and sharper details than GHG. and iHuman are scene-specific methods while GHG is a generalizable approach, the s…
Figure 10
Figure 10. Figure 10: Comparison to GPS-Gaussian in the setting of 5 source images on THuman2.0. Our method produces more complete shape and sharper details [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Comparisons on baselines on AIST++. Our method has fewer floaters compared to ActorsNeRF and produces more complete shape than GoMAvatar and iHuman. Meanwhile, ours is 11× faster than iHuman in reconstruction [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Robustness to SMPL-X shape accuracy. We use the ground-truth SMPL-X shape and the average shape as initialization of the canonical mesh. Although the average shape is smaller than the ground-truth shape, our method still captures the correct shape. (a) Reference image…
Figure 13
Figure 13. Figure 13: Examples of failure cases. F LIMITATIONS We present three types of failure cases in our method and discuss the possible next steps to resolve the issues. Failure in hallucination large regions. Without an explicit hallucination module, our method is unable to inpaint …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SAT: Supervisor Regularization and Animation Augmentation for Two-process Monocular Texture 3D Human Reconstruction

    cs.CV 2025-08 conditional novelty 4.0 of 10

    A two-stage Gaussian-splatting framework with supervisor feature regularization and online animation augmentation improves monocular textured 3D human reconstruction on CustomHuman and THuman3.0.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [2]

    Due to the explicit nature of the Gaussians-on-Mesh representation, our method produces fewer floaters than NeRF-based ActorNeRF

    The poses provided by the dataset are less accurate and the masks are predicted from off-the-shelf tools which are also less accurate. Due to the explicit nature of the Gaussians-on-Mesh representation, our method produces fewer floaters than NeRF-based ActorNeRF. Meanwhile, we capture better silhouettes and produce less noise compared to iHuman and GoMAv...

    work page 2024

  2. [5]

    Our method produces more complete shape and sharper details

    17 Published as a conference paper at ICLR 2025 (a) Ground truth(b) GPS-Gaussian(c) Ours(a) Ground truth(b) GPS-Gaussian(c) Ours Figure 10: Comparison to GPS-Gaussian in the setting of 5 source images on THuman2.0. Our method produces more complete shape and sharper details. Table 6: Per-scene breakdown on AIST++. We use lighter gray for scene-specific me...

    work page 2025

  3. [6]

    The AIST++ dataset consists of multiview dancing videos, camera calibration parameters, and human motions represented in SMPL poses

    The AIST++ dataset is used to evaluate the setting of multi-frame source images. The AIST++ dataset consists of multiview dancing videos, camera calibration parameters, and human motions represented in SMPL poses. We adopt the training and evaluation protocol of ActorsNeRF (Mu et al., 2023). Specifically, we use subjects 1-15 and 21-30 for training and le...

    work page 2023

  4. [7]

    We use 3 source images when com- paring with GoMAvatar, 3DGS-Avatar, iHuman, NHP, NIA and GHG

    on THuman2.0. We use 3 source images when com- paring with GoMAvatar, 3DGS-Avatar, iHuman, NHP, NIA and GHG. For the comparison with GPS-Gaussian, we adopt 5 source images following the setting of GHG (Kwon et al., 2024), since GPS-Gaussian requires the source views to overlap with each other and thus does not work well with very sparse views. We compared...

    work page 2024

  5. [8]

    On AIST++, we compare with HumanNeRF (Weng et al., 2022), GoMAvatar (Wen et al., 2024), 3DGS-Avatar (Qian et al., 2024b), iHuman (Paudel et al.,

    on XHuman. On AIST++, we compare with HumanNeRF (Weng et al., 2022), GoMAvatar (Wen et al., 2024), 3DGS-Avatar (Qian et al., 2024b), iHuman (Paudel et al.,

    work page 2022

  6. [9]

    HumanNeRF, GoMAvatar, 3DGS-Avatar and iHuman need to be trained per scene

    and ActorsNeRF (Mu et al., 2023). HumanNeRF, GoMAvatar, 3DGS-Avatar and iHuman need to be trained per scene. ActorsNeRF adopts a two-stage training: In the first stage, it learns a categorical prior from large-scale datasets. In the second stage, it adopts per-scene optimization given the source images. Now we detail the training setup of each baseline. N...

    work page 2023

  7. [10]

    We follow the same setting as reported in GHG (Kwon et al.,

    and GHG (Kwon et al., 2024). We follow the same setting as reported in GHG (Kwon et al.,

    work page 2024

  8. [11]

    GPS-Gaussian (Zheng et al., 2024)

    for training and evaluation. GPS-Gaussian (Zheng et al., 2024). As described in GHG (Kwon et al., 2024), GPS-Gaussian can work on as few as 5 input views. We render the THuman2.0 dataset to accommodate this setting. We use the default parameters provided in GPS-Gaussian to train the model. GoMA vatar (Wen et al., 2024). GoMAvatar originally takes SMPL par...

    work page 2024

Show all 19 references
  1. [12]

    Although the average shape is smaller than the ground-truth shape, our method still captures the correct shape. 19 Published as a conference paper at ICLR 2025 (a) Ground-truth SMPL-X shape on input images (c) Average SMPL-X shape on input images (b) Rendered image w/ (a) (d) ...

    2025

  2. [13]

    E A DDITIONAL ANALYSIS For additional qualitative results we refer the reader to the project webpage

    for training and evaluation. E A DDITIONAL ANALYSIS For additional qualitative results we refer the reader to the project webpage. It contains videos for freeview rendering, cross-domain generalization and novel pose synthesis. E.1 A DDITIONAL COMPARISONS ON THUMAN 2.0 We comp...

    2024

  3. [14]

    in the setting of 3 source images on THuman2.0. Note that GoMAvatar 16 Published as a conference paper at ICLR 2025 (a) Ground truth(b) iHuman(c) GoMAvatar(d) GHG(e) Ours Figure 9: Comparison to baselines in the setting of 3 source images on THuman2.0. Our method produces less...

    2025

  4. [15]

    During inference time, it takes as inputs two adjacent views and interpolates the novel views in between

    relies on stereo depth estimation to locate the Gaussians. During inference time, it takes as inputs two adjacent views and interpolates the novel views in between. Therefore, it requires the adjacent views to overlap with each other. As mentioned in GHG (Kwon et al., 2024), 5...

    2024

  5. [18]

    However, our approach improves upon GHG in all noise levels

    Both methods are affected by the accuracy of the input poses. However, our approach improves upon GHG in all noise levels. To make our approach less sensitive to the accuracy of input poses, we can explore a pose refinement network that is jointly trained with the iterative fe...

    2025

  6. [1192]

    #$Sample Transformer Encoder Layer𝐵×𝑁×𝑉↓×𝐶Rearrange(𝐵×𝑉↓)×𝑁×𝐶 Transformer Encoder Layer Vertex embedding 𝑒!

    We use 6 heads in the attention layers. Note that in the second Transformer encoder layer, we use a learnable vertex embedding {e↓ t−1,i}V ↓ i=1 as the query. The learnable vertex embedding is updated in iterative updates together with the low-resolution vertices. 1https://wen...

    2025

  7. [2021]

    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. CVPR, 2024a. Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-avatar...

    2025

  8. [2022]

    We follow the same setting as reported in ActorsNeRF (Mu et al.,

    and ActorsNeRF (Mu et al., 2023). We follow the same setting as reported in ActorsNeRF (Mu et al.,

    2023

  9. [2023]

    The dataset provides 20 subjects with high-quality scans and SMPL-X parameters

    We use XHuman to validate our approach to cross-domain gen- eralization quantitatively. The dataset provides 20 subjects with high-quality scans and SMPL-X parameters. We sample three scans (f00001, f00051, f00101) for each subject. We prepare the dataset in the same way as TH...

    2025

  10. [2024]

    Gart: Gaussian articulated template models

    11 Published as a conference paper at ICLR 2025 Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In CVPR,

    2025

  11. [2025]

    A summarizes mesh representations in human modeling; • Sec

    13 Published as a conference paper at ICLR 2025 APPENDIX — LIF E-GOM: G ENERALIZABLE HUMAN RENDERING WITH LEARNED ITERATIVE FEEDBACK OVER MULTI-RESOLUTION GAUSSIANS -ON-MESH This appendix is structured as follows: • Sec. A summarizes mesh representations in human modeling; • S...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.