Pith. sign in

REVIEW 4 major objections 5 minor 59 references

Snap-Snap: Taking Two Images to Reconstruct 3D Human Gaussians in Milliseconds

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Two phone photos can reconstruct a full 3D human in 190 ms

desk verdict Solid two-view human reconstruction system, but the side-view supervision is underspecified and needs clarification before the geometry claim is believable. read the letter →

arxiv 2508.14892 v1 pith:FLS7FBEV submitted 2025-08-20 cs.GR cs.CV

Jia Lu , Taoran Yi , Jiemin Fang , Chen Yang , Chuiyun Wu , Wei Shen , Wenyu Liu , Qi Tian
show 1 more author
Xinggang Wang
This is my paper · ORCID
classification cs.GRcs.CV
keywords 3Dhumanreconstructionsparse-viewGaussiansplattingfront-backimagepairsfeed-forwardpointcloudpredictiondigitalnovelviewsynthesis
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

This paper tries to establish that two casually captured photographs, one from the front and one from the back, are enough to produce a complete renderable 3D human body in about 190 milliseconds. The authors propose a feed-forward pipeline that predicts human point clouds from four synthesized viewpoints, fills in missing side color with a simple nearest-neighbor warp, and regresses 3D Gaussian attributes that can be rasterized directly. If the claim holds, it removes the main barriers to consumer 3D avatars: no multi-camera rig, no body-model fitting, and no slow per-person optimization. The payoff would be an avatar pipeline that runs at interactive speed on a single GPU from two smartphone images, including for loose clothing and unseen capture settings.

What carries the argument

The load-bearing object is a four-head point-cloud prediction network built on a DUSt3R-style geometry-reconstruction architecture. Two heads decode front and back point clouds from image tokens; two additional heads consume averaged front/back tokens to hallucinate left and right side geometry in the front view's coordinate frame, so the four clouds can be concatenated without camera poses. A learnable scale factor aligns the reconstructed body to real-world size, a nearest-neighbor search transfers front/back colors to the side clouds, and a UNet-like regression network predicts Gaussian attributes (position offset, color, opacity, scale, rotation) from each viewpoint's point cloud and col

What would settle it

Take the released model and feed it front/back photos whose person masks are produced by an automatic segmentation network instead of ground-truth masks; if end-to-end wall time or rendered PSNR degrades sharply, the 'two images in milliseconds' claim depends on that unmeasured component. Independently, reconstruct a subject with one arm raised in the front view but lowered in the back view and measure the hole fraction near the armpits: DUSt3R-derived depth supervision predicts those holes, testing the geometry prior's limit.

Watch

Extended reading notes

Core claim

The central discovery is that a complete, renderable 3D human can be predicted directly from front and back photographs. Snap-Snap does this by adapting a generalizable geometry-reconstruction network to the human domain, using four prediction heads to synthesize front, back, left, and right point clouds from the two input views; a nearest-neighbor color-warping step fills in side textures; and a UNet-like Gaussian regression module converts the colored point clouds into 3D Gaussians. The authors argue that this removes the need for SMPL-X priors, camera parameters, and dense multi-view capture, and report the best rendered-image metrics among the compared methods at 1024x1024 resolution fro

Load-bearing premise

The pipeline assumes clean person masks are already available; the reported 190 ms runs from masked images to Gaussians, and the paper does not describe how masks are obtained at test time. Stage-1 geometry also inherits the accuracy of DUSt3R depth maps, leaving holes in occluded regions.

Editorial extensions

If this is right

  • A user with two synchronized phone photos gets a complete 3D avatar; the system runs in 190 ms on a single RTX 4090 at 1024x1024 input, with no camera calibration or body-model fitting.
  • The method outperforms prior sparse-view Gaussian baselines that need more views or a body template, and handles loose clothing because it does not rely on a fitted body model.
  • Because the pipeline is feed-forward and training scales with more scan data, adding training subjects improves reconstruction quality, pointing toward larger-scale avatar training.
  • The same architecture extends to five-view input, where it removes the side-view heads and still beats a depth-plus-Gaussian baseline, suggesting the point-cloud-plus-regression design generalizes beyond two-view capture.
  • Reconstruction works on low-cost mobile phone captures without camera parameters, lowering the barrier for consumer 3D human creation.

Reading between the lines

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

  • Editorial inference: the reported 190 ms runs from masked images to Gaussians, so adding an automatic segmentation stage would raise end-to-end latency; the practical 'two photos to avatar' time depends on how masks are obtained at test time.
  • Editorial inference: because stage-1 geometry is supervised by depth maps from the DUSt3R-style network, armpit and occlusion holes trace back to that prior; a geometric generative prior or temporal multi-frame input could plausibly close them.
  • Editorial inference: the front/back-plus-hallucinated-sides design could transfer to other nearly-opposite-view capture settings, such as object or clothing digitization, where color completion by nearest-neighbor point-cloud search is a camera-free baseline.
  • Editorial inference: replacing the nearest-neighbor side-color warp with a learned inpainting or diffusion step could improve side texture consistency beyond the gains the paper's own ablations show for the current 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

4 major / 5 minor

Summary. The paper proposes Snap-Snap, a feed-forward framework that reconstructs a complete 3D human Gaussian representation from only two RGB images (front and back) in about 190 ms on an RTX 4090, without using SMPL-X or other body priors. The method adapts the DUSt3R architecture by adding two side-view prediction heads, transfers color from the front/back views to the predicted side point clouds via nearest-neighbor search, and then regresses Gaussian attributes with a U-Net-like network. Experiments on THuman2.0, 2K2K, and 4D-Dress, plus a loose-clothing subset and a five-view scenario, report improved or competitive PSNR/SSIM/LPIPS over GPS-Gaussian, GHG, and SiTH; ablations quantify the contributions of the side heads, the NNS color transfer, and DUSt3R pretraining. Qualitative results on mobile-phone captures are also shown.

Significance. If the claims hold, the work is significant: it lowers the data-capture barrier for personalized 3D human reconstruction to two smartphone images, removes the dependence on a parametric body model, and achieves millisecond-scale inference. The paper has real strengths: the final rendering quality is evaluated against independent ground-truth images and meshes, so the headline comparison is not circular; the experiments are fairly extensive, including cross-domain evaluation, a loose-clothing subset, a five-view comparison, and ablations of both main modules; and the authors promise code and demos. The main risk is not the headline comparison but the training-data pipeline for the geometry module and the unstated role of input masks, both of which are load-bearing for the central claim.

major comments (4)
  1. [§3.3, §3.6, and A.14] The supervision for the side-view heads H_l and H_r is not defined. In Eq. (6), L_stage1 compares the concatenated four-view point cloud P_h (Eq. (3)) to P_gt. However, A.14 states that the point cloud supervision is derived from DUSt3R depth maps [44]. DUSt3R, when given only the front and back images, produces depth only for those two views; it does not observe the left or right sides. The paper never specifies how side-view ground-truth points for H_l and H_r are generated (e.g., by rendering THuman2.0 meshes from left/right cameras, by a symmetry assumption, or by some other pseudo-labeling procedure). Without such a target, Eq. (6) cannot supervise the side heads as written, and the claim that the model 'builds human point clouds even with highly sparse input' is not independently verifiable. If P_gt is in fact derived from ground-truth meshes for all four viewpoints, then A.14 is m
  2. [Table 1 footnote and §4.4] The reported inference time is 'from receiving images and masks to the final Gaussians,' but the paper never explains how masks are obtained at test time. If masks come from an external segmentation network, that network is an unstated component whose cost and failure modes affect every reported timing and quality number. If masks are assumed to be user-provided, then the system is not fully automatic from two raw images, which weakens the 'just take two images' claim. The mask acquisition procedure (model choice, resolution, CPU/GPU, whether its time is included in the 190 ms budget) must be specified, and the inference-time breakdown in Fig. 16 should include it if it is part of the pipeline.
  3. [§3.4 and §3.6] The training-time construction of the side pseudo-images I_l and I_r is underspecified. Stage 2 (Eq. (7)) is trained with novel-view rendering loss against ground-truth images, but the inputs to F_g for the side views are I_l and I_r produced by the NNS color transfer from P_lf and P_rf. During training, are P_lf and P_rf the outputs of the stage-1 network (trained separately), or are they replaced by ground-truth geometry? If they are network outputs, the error from stage 1 is baked into stage 2 with no end-to-end gradient; if they are ground-truth side projections, then the inference-time behavior may differ substantially from the training-time inputs. The paper should state the exact data flow used to form the stage-2 side-view training pairs.
  4. [§4.2, Table 1] The comparison with GHG under two views uses EasyMocap-estimated SMPL-X, but the GHG 2-view row with ground-truth SMPL-X is also reported. The paper is fair in separating these, but the main text should discuss whether the EasyMocap estimates are computed from the same two input images without any additional calibration or temporal information, and what failure modes this introduces. This matters because the 'no human prior' advantage of Snap-Snap is partly quantified by the large drop from GHG 2 GT to GHG 2 Estimated; the drop could be inflated if EasyMocap is handicapped by the front/back camera setup rather than by the fundamental two-view limitation.
minor comments (5)
  1. [§3.6 / Inference] The inference text lists 'P t,f' in the point cloud set; this appears to be a typo for P r,f or a fourth head that is never defined.
  2. [§4.2] The visual comparison caption says 'the resolution of the rendered images for all methods is 2024 × 2048'; this should presumably be 2048 × 2048.
  3. [§3.4 and §4.5] The text alternates between 'wrapping' and 'warping' for the NNS color transfer. Please use the correct term consistently, since the operation described is a nearest-neighbor assignment, not an image warp.
  4. [§2 / References] The reference to the 2K2K dataset [6] is given as 'High-fidelity 3d human digitization from single 2k resolution images'; please verify that this is the correct source for the 2K2K dataset used in the evaluation.
  5. [A.13 / Fig. 17] The Gaussian regression network description says 'we only use ResNet blocks' but the figure also shows GroupNorm and SiLU; please align the text and figure.

Circularity Check

1 steps flagged · score 3.0 of 10

Geometry module is fine-tuned from DUSt3R on DUSt3R-generated pseudo point clouds, making the point-cloud 'prediction' partially self-referential; final rendering comparison against ground-truth images keeps the headline result externally grounded.

  1. fitted input called prediction [Sec. 3.3, Sec. 3.6 Eq. 6, A.14]
    "To fully leverage the priors from the foundation geometric reconstruction model, we adopt a similar architecture to DUSt3R [44]. ... point cloud supervision, which is derived from depth maps [44], wherein certain points are filtered out because of occlusions. ... Lstage1 = Lreg (P h, Pgt) + Lconf (M conf , Mgt)."

    The stage-1 point cloud predictor Rp is initialized from DUSt3R (Sec. 3.3) and its supervision P_gt in Eq. 6 is, per A.14, 'derived from depth maps [44]' (DUSt3R). The model is therefore trained to regress DUSt3R's own depth outputs while being initialized with DUSt3R weights; the 'predicted' point clouds are a fine-tuned copy of the supervision source rather than an independent geometric estimate. Eq. 3 concatenates side-view point clouds P_l,f and P_r,f into the complete P_h, but DUSt3R front/back depth does not observe left/right sides and no independent side-view target is described, so the claimed 'complete' geometry is not derived from external ground truth. The severity is limited because the final rendered images are evaluated against ground-truth images (Sec. 4.2), which is an ext

full rationale

The central speed and rendering-quality claims are tested against external THuman2.0/2K2K/4D-Dress data and ground-truth images, so the headline result is not circular. The main circular tendency is internal to stage 1: the geometry module adopts DUSt3R's architecture/pretraining and is supervised by point clouds derived from DUSt3R depth maps (A.14), so the geometry 'prediction' largely reproduces DUSt3R on humans. This is a training-label provenance problem rather than a strict by-construction identity: there are additional side heads, a learned scale δ, and a separately trained Gaussian regression network, and the final rendering is independently evaluated. No load-bearing self-citation of the authors' own prior work is used; DUSt3R and other baselines are external. The side-view supervision gap is a correctness/verifiability concern and contributes to the score, but it does not force the headline reconstruction quality to be true by definition. Overall circularity: 3/10.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on one external foundation model (DUSt3R) used both as architecture initialization and as the source of pseudo ground-truth point clouds. The pipeline also assumes the availability of person masks and roughly aligned front/back full-body photos. No new physical entities or new representations are introduced beyond existing 3D Gaussian Splatting.

free parameters (2)
  • learnable scale factor δ = learned, value not reported
    Introduced in Eq. 3 to align predicted point clouds with real-world scale; learned during training, no closed-form derivation.
  • DUSt3R pretrained weights = pretrained weights from DUSt3R
    Used as initialization for the point cloud predictor; a prior pulled from prior literature, not fit here.
assumptions (4)
  • ad hoc to paper DUSt3R provides a reliable generalizable geometric prior and its point clouds/depth maps are suitable as pseudo ground truth for human training.
    Sec. 3.3 adapts DUSt3R architecture and A.14 states supervision is derived from DUSt3R depth maps. This is a load-bearing choice.
  • domain assumption Person masks are available at inference.
    Table 1 footnote says inference time counts 'receiving images and masks'; mask source is not described in the paper.
  • domain assumption The front and back input images are roughly aligned and show the full body.
    The method assumes the subject is approximately centered and full-body; the in-the-wild setup manually adjusts phone poses (A.1).
  • ad hoc to paper Averaging front and back token representations adequately encodes side-view geometry.
    Eq. 2 uses Gv = (Gf + Gb)/2 to generate side-view tokens; this is an architectural choice with no strong theoretical justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Snap-Snap: Taking Two Images to Reconstruct 3D Human Gaussians in Milliseconds." pith.science (2026). https://pith.science/paper/FLS7FBEV

@misc{pith2026250814892,
  author       = {Pith},
  title        = {Pith review of: Snap-Snap: Taking Two Images to Reconstruct 3D Human Gaussians in Milliseconds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLS7FBEV}},
  note         = {Machine review of arXiv:2508.14892}
}
read the original abstract

Reconstructing 3D human bodies from sparse views has been an appealing topic, which is crucial to broader the related applications. In this paper, we propose a quite challenging but valuable task to reconstruct the human body from only two images, i.e., the front and back view, which can largely lower the barrier for users to create their own 3D digital humans. The main challenges lie in the difficulty of building 3D consistency and recovering missing information from the highly sparse input. We redesign a geometry reconstruction model based on foundation reconstruction models to predict consistent point clouds even input images have scarce overlaps with extensive human data training. Furthermore, an enhancement algorithm is applied to supplement the missing color information, and then the complete human point clouds with colors can be obtained, which are directly transformed into 3D Gaussians for better rendering quality. Experiments show that our method can reconstruct the entire human in 190 ms on a single NVIDIA RTX 4090, with two images at a resolution of 1024x1024, demonstrating state-of-the-art performance on the THuman2.0 and cross-domain datasets. Additionally, our method can complete human reconstruction even with images captured by low-cost mobile devices, reducing the requirements for data collection. Demos and code are available at https://hustvl.github.io/Snap-Snap/.

Figures

Figures reproduced from arXiv: 2508.14892 by the authors.

Figure 1
Figure 1. Under the setting of two input images, we propose [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of Snap-Snap. With the input front and back view images [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The framework of point cloud prediction network. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Visual comparisons with GPS-Gaussian [59] and GHG [20]. model weight. GPS-Gaussian under 5 views is completely trained following the instructions. The reconstruction quality of GPS-Gaussian is lower than ours even with five viewpoints. In [PITH_FULL_IMAGE:figures/full…
Figure 5
Figure 5. Figure 5: Qualitative comparisons with SiTH [8]. front SIFU LGM GTA Human3Diffusion ECON Ours GT [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visual comparisons of Snap-Snap with single-view re [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Reconstruction results from in-the-wild data. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: We present the visual results of ablation studies on the [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Data collection setup visualization. Input Images TRELLIS (2 view) Snap-Snap (Ours) GT [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Visual comparisons of Snap-Snap with TRELLIS [ [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 12
Figure 12. Figure 12: Visualization of different texture introduced by gener [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Reconstruction results based on arbitrary image inputs. [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: We present the visual results of the impact of the addi [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 15
Figure 15. Figure 15: Reconstruction results of the same human in different [PITH_FULL_IMAGE:figures/full_fig_p013_15.png]
Figure 16
Figure 16. Figure 16: The visual results of the time consumption for different [PITH_FULL_IMAGE:figures/full_fig_p014_16.png]
Figure 17
Figure 17. Figure 17: The network architecture of the Gaussian attribute regression network. [PITH_FULL_IMAGE:figures/full_fig_p015_17.png]
Figure 18
Figure 18. Figure 18: We visualize the comparison of reconstruction results between our method and GPS-Gaussian [ [PITH_FULL_IMAGE:figures/full_fig_p015_18.png]
Figure 19
Figure 19. Figure 19: We visualize the reconstruction results of GHG [ [PITH_FULL_IMAGE:figures/full_fig_p016_19.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 42 canonical work pages

  1. [44]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. In CVPR, 2024. 2, 3, 12, 14

    work page 2024

  2. [1]

    Easymocap - make human motion capture easier. Github,

  3. [2]

    Auto-rectify network for unsupervised indoor depth estimation

    Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Tat-Jun Chin, Chunhua Shen, and Ian Reid. Auto-rectify network for unsupervised indoor depth estimation. IEEE transactions on pattern analysis and machine intelligence , 44(12):9802– 9813, 2021. 2

    work page 2021

  4. [3]

    Geometry-guided progressive nerf for generalizable and efficient neural human rendering

    Mingfei Chen, Jianfeng Zhang, Xiangyu Xu, Lijuan Liu, Yu- jun Cai, Jiashi Feng, and Shuicheng Yan. Geometry-guided progressive nerf for generalizable and efficient neural human rendering. In European Conference on Computer Vision , pages 222–239. Springer, 2022. 1

    work page 2022

  5. [4]

    Generalizable neural performer: Learning robust radiance fields for human novel view synthesis

    Wei Cheng, Su Xu, Jingtan Piao, Chen Qian, Wayne Wu, Kwan-Yee Lin, and Hongsheng Li. Generalizable neural performer: Learning robust radiance fields for human novel view synthesis. arXiv preprint arXiv:2204.11798, 2022. 1

    arXiv 2022

  6. [5]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2

    arXiv 2010

  7. [6]

    High-fidelity 3d human digitization from single 2k resolution images

    Sang-Hun Han, Min-Gyu Park, Ju Hong Yoon, Ju-Mi Kang, Young-Jae Park, and Hae-Gon Jeon. High-fidelity 3d human digitization from single 2k resolution images. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12869–12879, 2023. 2, 5, 6

    work page 2023

  8. [7]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 14

    work page 2016

Show all 59 references
  1. [8]

    Sith: Single-view tex- tured human reconstruction with image-conditioned diffu- sion

    I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view tex- tured human reconstruction with image-conditioned diffu- sion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 538–549, 2024. 1, 7, 12

    2024

  2. [9]

    Learn- ing locally editable virtual humans

    Jie Song Hsuan-I Ho, Lixin Xue and Otmar Hilliges. Learn- ing locally editable virtual humans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2023. 8

    2023

  3. [10]

    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 Pat...

    2024

  4. [11]

    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, pages 20418–20431, 2024. 2

    2024

  5. [12]

    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 2024 Interna- tional Conference on 3D Vision (3DV) , pages 1531–1542. IEEE, 2024. 2

    2024

  6. [13]

    Splatarmor: Articulated gaussian splatting for animat- able humans from monocular rgb videos

    Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choud- hary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animat- able humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2

    2023 arXiv

  7. [14]

    Hifi4g: High-fidelity human performance rendering via compact gaussian splatting

    Yuheng Jiang, Zhehao Shen, Penghao Wang, Zhuo Su, Yu Hong, Yingliang Zhang, Jingyi Yu, and Lan Xu. Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 19...

    2024

  8. [15]

    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 Trans. Graph., 42(4):139–1,

  9. [16]

    Hugs: Human gaussian splats

    Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 505–515, 2024. 2

    2024

  10. [17]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. Advances in neural information processing systems , 25, 2012. 5

    2012

  11. [18]

    Neural human performer: Learning generalizable ra- diance fields for human performance rendering

    Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable ra- diance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741–24752,

  12. [19]

    Neural image-based avatars: Generalizable radi- ance fields for human avatar modeling

    Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural image-based avatars: Generalizable radi- ance fields for human avatar modeling. arXiv preprint arXiv:2304.04897, 2023

    2023 arXiv

  13. [20]

    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. arXiv preprint arXiv:2407.12777, 2024. 1, 2, 5, 6, 7, 1...

    2024 arXiv

  14. [21]

    Gart: Gaussian articulated template mod- els

    Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template mod- els. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 19876–19887,

  15. [22]

    Ground- ing image matching in 3d with mast3r, 2024

    Vincent Leroy, Yohann Cabon, and Jerome Revaud. Ground- ing image matching in 3d with mast3r, 2024. 3

    2024

  16. [23]

    Gaussianbody: Clothed human re- construction via 3d gaussian splatting

    Mengtian Li, Shengxiang Yao, Zhifeng Xie, Keyu Chen, and Yu-Gang Jiang. Gaussianbody: Clothed human re- construction via 3d gaussian splatting. arXiv preprint arXiv:2401.09720, 2024. 2

    2024 arXiv

  17. [24]

    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 , pages 419–436. Springer, 2022. 1

    2022

  18. [25]

    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, pages 19711–19722, 2024. 2

    2024

  19. [26]

    Raft-stereo: Multilevel recurrent field transforms for stereo matching

    Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV) , pages 218–227. IEEE, 2021. 2

    2021

  20. [27]

    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. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851–866. 2023. 2

    2023

  21. [28]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5

    2017 arXiv

  22. [29]

    Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encod- ing of keypoints

    Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. Keypointnerf: Generalizing image-based volumetric avatars using relative spatial encod- ing of keypoints. In European conference on computer vi- sion, pages 179–197. Springer, 2022. 2

    2022

  23. [30]

    Human gaussian splatting: Real-time rendering of animatable avatars

    Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo P ´erez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788–798, 2024. 2

    2024

  24. [31]

    Humansplat: Generalizable single-image human gaussian splatting with structure priors

    Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splatting with structure priors. arXiv preprint arXiv:2406.12459, 2024. 2

    2024 arXiv

  25. [32]

    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 IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 20...

    2019

  26. [33]

    Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans

    Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceed- ings of the IEEE/CVF Conference on Computer Vision a...

    2021

  27. [34]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. 2

    2021

  28. [35]

    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, pages 10684–10695, 2022. 1, 12

    2022

  29. [36]

    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–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, pa...

    2015

  30. [37]

    Efficient variants of the icp algorithm

    Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international confer- ence on 3-D digital imaging and modeling , pages 145–152. IEEE, 2001. 8

    2001

  31. [38]

    Pifu: Pixel-aligned implicit function for high-resolution clothed human digitiza- tion

    Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Mor- ishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitiza- tion. In Proceedings of the IEEE/CVF international confer- ence on computer vision, pages 2304–2314, 2019. 2

    2019

  32. [39]

    Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization

    Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84–93, 2020. 2

    2020

  33. [40]

    Floren: Real-time high-quality human performance render- ing via appearance flow using sparse rgb cameras

    Ruizhi Shao, Liliang Chen, Zerong Zheng, Hongwen Zhang, Yuxiang Zhang, Han Huang, Yandong Guo, and Yebin Liu. Floren: Real-time high-quality human performance render- ing via appearance flow using sparse rgb cameras. In SIG- GRAPH Asia 2022 Conference Papers, pages 1–10, 2022. 1

    2022

  34. [41]

    Mvdream: Multi-view diffusion for 3d gen- eration

    Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d gen- eration. arXiv:2308.16512, 2023. 1

    2023 arXiv

  35. [42]

    Sc-depthv3: Robust self- supervised monocular depth estimation for dynamic scenes

    Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self- supervised monocular depth estimation for dynamic scenes. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 2023. 2

    2023

  36. [43]

    Lgm: Large multi-view gaussian model for high-resolution 3d content creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1–18. Springer, 2025. 8

    2025

  37. [45]

    4d-dress: A 4d dataset of real-world human clothing with se- mantic annotations

    Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Ar- tur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with se- mantic annotations. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recogni...

    2024

  38. [46]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 5

    2004

  39. [47]

    Croco v2: Improved cross-view completion pre- training for stereo matching and optical flow

    Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Br ´egier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and J ´erˆome Revaud. Croco v2: Improved cross-view completion pre- training for stereo matching and optical flow. InProceedin...

    2023

  40. [48]

    Structured 3d latents for scalable and versatile 3d gen- eration

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d gen- eration. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21469–21480, 2025. 11

    2025

  41. [49]

    Icon: Implicit clothed humans obtained from nor- mals

    Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from nor- mals. In 2022 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 13286–13296. IEEE, 2022. 2

    2022

  42. [50]

    Econ: Explicit clothed humans optimized via normal integration

    Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 512–523, 2023. 2, 8

    2023

  43. [51]

    Human-3diffusion: Realistic avatar creation via explicit 3d consistent diffusion models

    Yuxuan Xue, Xianghui Xie, Riccardo Marin, and Gerard Pons-Moll. Human-3diffusion: Realistic avatar creation via explicit 3d consistent diffusion models. Advances in Neural Information Processing Systems, 37:99601–99645, 2024. 1, 8

    2024

  44. [52]

    Gaussianob- ject: Just taking four images to get a high-quality 3d object with gaussian splatting

    Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Gaussianob- ject: Just taking four images to get a high-quality 3d object with gaussian splatting. arXiv preprint arXiv:2402.10259 ,

    arXiv

  45. [53]

    Generalizable neural voxels for fast human radiance fields

    Taoran Yi, Jiemin Fang, Xinggang Wang, and Wenyu Liu. Generalizable neural voxels for fast human radiance fields. arXiv preprint arXiv:2303.15387, 2023. 1, 2

    2023 arXiv

  46. [54]

    Learning to recover 3d scene shape from a single image

    Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204–213, 2021. 2

    2021

  47. [55]

    Function4d: Real-time human vol- umetric capture from very sparse consumer rgbd sensors

    Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qiong- hai Dai, and Yebin Liu. Function4d: Real-time human vol- umetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 5746–5756, 2021...

    2021

  48. [56]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 5

    2018

  49. [57]

    Global-correlated 3d-decoupling transformer for clothed avatar reconstruction

    Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. Advances in Neural Informa- tion Processing Systems, 36:7818–7830, 2023. 8

    2023

  50. [58]

    Sifu: Side-view conditioned implicit function for real-world us- able clothed human reconstruction

    Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world us- able clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9936–9947, 2024. 1, 8

    2024

  51. [59]

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

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

    2024

Pith tools

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