Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

AdaHuman: Animatable Detailed 3D Human Generation with Compositional Multiview Diffusion

T0 review · 4 major / 4 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 AdaHuman claims that a single in-the-wild photo is enough to build a detailed, animation-ready 3D avatar, reconstructing the input pose and a standard A-pose with reduced self-occlusion via pose-conditioned joint diffusion and…

desk verdict Solid engineering advance with a likely-contaminated CustomHumans test split; the headline numbers need a re-run or a clear disjointness statement before I'd trust them. read the letter →

arxiv 2505.24877 v1 pith:HNADPATV submitted 2025-05-30 cs.CV

Yangyi Huang , Ye Yuan , Xueting Li , Jan Kautz , Umar Iqbal This is my paper · ORCID
classification cs.CV
keywords single-image3Davatargenerationpose-conditioneddiffusionGaussiansplattinganimatablehumanmulti-viewconsistencycompositionalrefinementnovelposesynthesis
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

AdaHuman claims that one in-the-wild photograph of a person is enough to create a high-fidelity, animation-ready 3D avatar, without multi-view capture or per-subject optimization. The paper's central bet is that doing 3D Gaussian reconstruction inside every step of a pose-conditioned multi-view diffusion model, and then refining body parts locally before composing them into one avatar, yields both pixel-aligned reconstruction in the input pose and a reposed A-pose avatar that can be rigged and animated. The work reports that this approach beats existing methods on novel view synthesis and novel pose synthesis metrics on CustomHumans and Sizer, and is preferred by users on in-the-wild images. If the claim holds, avatar creation for games, animation, and virtual reality becomes practical from ordinary single images.

What carries the argument

The load-bearing object is the pose-conditioned 3D joint diffusion model: a multi-view latent diffusion U-Net with 3D cross-view attention that, at every denoising step, feeds its predicted clean images to a 3D Gaussian generator and renders 3D-consistent predictions back into the next diffusion step. Its conditioning inputs are a 2D semantic pose map rendered from SMPL and a camera ray map, which together let the same model reconstruct or repose. The second mechanism is compositional 3DGS refinement: the crop-aware camera ray map converts local-view pixel coordinates into global-view ray coordinates through the crop box, so locally refined parts and global full-body views share one 3D space, and a visibility-aware composition rule keeps only trustworthy, visible Gaussians when merging parts.

What would settle it

Look up the identities of the 50 CustomHumans evaluation subjects and check whether any of them appear among the 589 meshes rendered for training augmentation; if there is overlap, re-run the reconstruction and reposing comparisons on a provably disjoint split and see whether the reported PSNR, SSIM, LPIPS, and FID advantages persist.

Watch

Extended reading notes

Core claim

The core discovery is that multi-view consistency and pose transfer do not need to be separated into independent stages. By conditioning a multi-view latent diffusion model on a target semantic pose map and generating a 3D Gaussian splat avatar at every reverse-diffusion step, AdaHuman obtains images and geometry that agree across views while also changing pose on demand. A second stage renders close-up views of the head, upper body, and lower body, refines them with image-to-image editing, and stitches them back using a crop-aware camera ray map plus a visibility-aware composition rule. The result is a detailed 3DGS avatar in the input pose, and a separate A-pose version with reduced self-occlusion that can be skinned and animated with arbitrary motions. The paper also states that this is achieved without training on standard-pose images, since the required poses are synthesized by the pose-conditioned diffusion model.

Load-bearing premise

The evaluation on CustomHumans is only valid if the 50 test subjects are not among the 589 human meshes used to add rendered training images, and the paper does not say that these sets are disjoint.

Editorial extensions

If this is right

  • A single image yields both a pixel-aligned 3DGS avatar and a reposed A-pose avatar, so occlusion-heavy input poses no longer block rigging and animation.
  • The avatar can be driven by arbitrary input motion through SMPL skinning weights, despite never being trained on standard-pose images.
  • Joint diffusion with per-step 3D reconstruction is what keeps generated multi-view images consistent; the 'No Joint Diffusion' ablation drops in quality.
  • Local body-part refinement improves fine details such as the face and reduces floating artifacts compared with direct or learned composition.
  • Training on multi-view video plus rendered meshes lets the method generalize to loose and unusual clothing better than SMPL-deformation baselines.

Reading between the lines

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

  • If the central claim holds, single-image avatar generation reaches a production-usable point, and the practical bottleneck shifts from capture to downstream animation quality, which the paper itself flags for hands, facial expressions, and garment deformation.
  • The pose-conditioned joint diffusion design should transfer beyond humans to any articulated class with a parametric pose model; a natural test is quadrupeds or hands.
  • The crop-aware camera ray map is a general mechanism for fusing local and global 3D reconstructions, and could be tested on object part refinement rather than only body parts.
  • Because the reported metrics are on still images, the biggest open test is temporal coherence of the animated 4D video, which the current evaluation does not measure.
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 / 4 minor

Summary. The paper proposes AdaHuman, a framework for single-image 3D human avatar generation. Its two main contributions are a pose-conditioned 3D joint diffusion model that interleaves multiview image synthesis with 3D Gaussian Splatting reconstruction at each denoising step, and a compositional 3DGS refinement module that refines local body parts via SDEdit-style image-to-image editing and merges them into a holistic avatar using a crop-aware camera ray map and visibility-aware composition. The method is evaluated on CustomHumans and Sizer for novel view synthesis, on MVHumanNet for novel pose synthesis, and on SHHQ in-the-wild images for a user study. The authors report consistent improvements over LGM, SiTH, SIFU, and Human3Diffusion on rendering metrics, with a runtime of about 70 seconds per avatar.

Significance. If the reported results are valid, the paper makes a strong practical contribution: a single in-the-wild image can produce both a pixel-aligned 3DGS avatar and a standardized A-pose avatar that supports rigging and animation, without requiring multiview capture or standard-pose training data. The architecture is coherent, the pose-conditioned diffusion plus joint 3D reconstruction is a sensible extension of Human3Diffusion, and the compositional refinement with crop-aware ray maps is a novel and plausible way to add local detail. The paper also provides qualitative results, a user study, and a clear statement that code and models will be released. However, the main quantitative support for the central claim depends on a potentially contaminated CustomHumans evaluation, as discussed in the major comments, and no error bars or significance tests are provided for the metric tables.

major comments (4)
  1. [§4 (Datasets)] The CustomHumans evaluation appears to overlap with the CustomHumans training renders, and the paper never states that the 50 testing subjects are disjoint from the 589 meshes used for training augmentation. The text says "we mixed the training data with multiview images rendered from 589 human meshes in the CustomHumans dataset" and later "50 testing subjects from the CustomHumans dataset" are used for evaluation. If 589 is the total number of meshes in CustomHumans, which is the commonly cited figure, the 50 test subjects are arithmetically contained in the training-render set. This is load-bearing because Table 1 (all CustomHumans columns) and the entire ablation study in Table 4 are computed on CustomHumans; an overlap would inflate AdaHuman's metrics relative to baselines that did not augment with renders of all meshes, and would invalidate the headline claim of substantial improvement. Please state explicitly whether the 50 test subjects are disjoint from the 589 training meshes, and if they are not, re-run the evaluation on a disjoint held-out set or provide per-subject results that demonstrate the advantage is not due to memorization.
  2. [Table 1 and Table 4] No error bars, confidence intervals, or significance tests are reported for any quantitative comparison. This matters especially for the Sizer F-score, where AdaHuman (0.6075) is lower than Human3Diffusion (0.6336), and for the small margins in some SSIM and CD values. Without per-subject variance or a paired test, the reader cannot tell whether the reported differences are meaningful or noise. Please report standard deviations or confidence intervals across the 97 Sizer subjects and 50 CustomHumans subjects, and ideally a paired significance test for the headline comparisons.
  3. [§4.1 (Baselines)] The comparison with SIFU is conducted with its text-guided texture refinement disabled because prompts are unavailable. This may disadvantage SIFU, since texture refinement is part of its intended pipeline. The paper should either provide prompts (e.g., derived from the subject or generic clothing descriptions) or explicitly discuss how disabling this component affects the comparison. This is not a fatal flaw, but it should be acknowledged so the reader can appropriately weigh the SIFU rows in Table 1.
  4. [§3.2 (Visibility-aware composition)] The visibility-aware composition rule uses several hand-set thresholds: a splat is reliable if covered by more than 2 views, by 3 views for the head part, and is dropped if it has higher visibility in another part of similar detail. These thresholds are free parameters, but their sensitivity is not analyzed in the ablation study. Since the composition strategy is one of the two main contributions, a small sensitivity analysis (e.g., varying the coverage threshold from 2 to 4) would strengthen the claim that the proposed rule, rather than the specific threshold values, drives the improvement in Table 4.
minor comments (4)
  1. [Throughout] There are several typographical and formatting errors that should be corrected: "appedix" in §3.1, "seemlessly" in §5, "reconstructs pixel-aligned a 3DGS avatar" in the Figure 1 caption, "generatd" in §3.2, and "1,0000 iterations" in Appendix A (appears to be 10,000).
  2. [Table 3] The baseline labels "SiTH" and "SIFU" are not marked with a symbol indicating that they are reposed by linear blend skinning, while the two AdaHuman rows are labeled "GR + deform" and "GPt (Ours)". Adding a footnote that all baselines use the same input pose and target pose conditions would make the comparison easier to follow.
  3. [§4.4 (Ablation study)] The ablation table reports "Ours + GT Pose Condition" shows large gains (PSNR 21.46 to 23.00). The text says this indicates potential for improvement, but it is worth discussing whether the pose estimation errors are a practical concern for the claimed "minimal self-occlusion" A-pose generation, since the final animation pipeline relies on estimated poses.
  4. [Appendix B] The claim that some baselines "rely on private data or synthesized meshes for training" is not substantiated with references to the specific training data of each baseline. Please be more precise about which baseline uses which data, since this affects the fairness of the comparison.

Circularity Check

1 steps flagged · score 4.0 of 10

No equation-level circularity; the one load-bearing circular-adjacent step is that the 50 CustomHumans test subjects are never shown to be disjoint from the 589 CustomHumans meshes rendered into training data, so the CH columns of Table 1 and the entire Table 4 ablation are partly recall of training identities rather than held-out prediction.

  1. fitted input called prediction [Section 4, Datasets paragraph; Tables 1 and 4]
    "we mixed the training data with multiview images rendered from 589 human meshes in the CustomHumans [12] dataset to improve generalizability. 50 testing subjects from the CustomHumans [12] dataset and 97 subjects from Sizer [46] dataset are used to quantitatively compare our method against SOTA approaches."

    The model is trained with multiview images rendered from 589 CustomHumans meshes, and the paper then reports 'novel view synthesis' numbers on 50 subjects 'from the CustomHumans [12] dataset' (Table 1, CH columns), with Table 4's ablation also run on CustomHumans. The paper never states that the 50 test subjects are disjoint from the 589 rendered meshes; since 589 is the natural reading of the dataset's mesh count, the 50 test identities are arithmetically contained in the training-render set. The CH 'prediction' columns are then in part re-renderings of training identities — a fitted input renamed as prediction, statistically forced by exposure rather than measured generalization.

full rationale

This is an empirical ML paper, not a derivation chain; there are no equation-level reductions (no Eq. X = Eq. Y by construction) in the method itself. The components — pose-conditioned 3D joint diffusion (Sec. 3.1) and compositional 3DGS refinement with the crop-aware ray map (Sec. 3.2) — are trained end-to-end on MVHumanNet videos plus CustomHumans renders, with losses (latent MSE; rendering MSE/LPIPS/surface regularization) measured against ground-truth multiview video frames; no predicted quantity is defined by the fit. Self-citations (TeCH [18], GAvatar [60]) appear only as background items in Related Work lists and carry none of the argument; the load-bearing priors are external and reproducible (stable-diffusion-v1-5 for the LDM, LGM for the 3DGS generator, and the joint-diffusion scheme following Human3Diffusion [55] and CAT3D [10]). No uniqueness theorem or ansatz is smuggled in via self-citation. The one flagged step is a benchmark-integrity issue bordering on pattern 2: the 50 CustomHumans test subjects are never stated to be disjoint from the 589 CustomHumans meshes rendered into training data, so the CH columns of Table 1 and all of Table 4 are partly recall rather than held-out prediction. This is weighed as partial, not total, circularity because the central claim retains independent content: the Sizer evaluation on 97 subjects (training includes no Sizer renders) shows consistent gains over all baselines; the reposing claim is validated on 50 explicitly 'unseen' MVHumanNet subjects with a clear disjointness statement; and the SHHQ user study on 53 in-the-wild images is external. A single sentence in Section 4 stating that the 50 CH test subjects are excluded from the 589-mesh render set would resolve the issue; absent that, the CH superiority margins and the Table 4 ablation conclusions are unverified as generalization evidence.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method is empirical and depends on several domain assumptions about pose conditioning, Gaussian reconstruction, and data distribution, plus a handful of hand-set parameters in the composition and refinement stages. No new physical entities are introduced.

free parameters (3)
  • Visibility coverage threshold for splat retention = 2 views, 3 views for head
    Section 3.2: a splat is considered reliable only if covered by more than 2 input views, or 3 for the head. This hand-set threshold directly controls which Gaussians survive composition.
  • SDEdit strength for local refinement = s = 0.5
    Appendix A: local views are denoised starting at t=500 with 3D joint diffusion between t=350 and 500. This hand-set value controls the detail versus consistency trade-off in local refinement.
  • Local body-part crop scale ratios = 1.0, 0.5, 0.5, 0.25
    Table 5 assigns crop scales for full body, upper, lower, and head. These ratios determine which image regions are refined and at what zoom, so they directly affect the detail and composition balance.
assumptions (5)
  • domain assumption MVHumanNet multi-view video provides consistent appearance and pose ground truth for the same subject across viewpoints and frames.
    Used to supervise both multi-view image synthesis and 3DGS reconstruction in Section 3.1; inconsistent training views would degrade 3D consistency.
  • domain assumption The semantic pose map rendered from SMPL is an adequate conditioning signal to transfer identity and appearance to an arbitrary target pose.
    Reposing relies entirely on pose conditioning in Eq. 1; if the pose map under-specifies appearance or geometry, the reposed avatar quality claim fails.
  • ad hoc to paper The linear crop-to-global coordinate mapping of Eq. 3 correctly places local-view pixels in the global camera ray embedding space.
    The compositional refinement module uses this mapping to feed local crops to a 3DGS generator trained on global views; any misalignment breaks the fusion of local and global reconstructions.
  • domain assumption The 50 CustomHumans test subjects are disjoint from the 589 CustomHumans meshes used for training augmentation.
    Section 4 reports both numbers but does not state disjointness; the quantitative comparison on CustomHumans is only interpretable under this assumption.
  • domain assumption The pose distribution in MVHumanNet is dense enough that the fine-tuned model can synthesize a canonical A-pose without standard-pose training examples.
    The animation-ready A-pose claim in Section 4.3 depends on this generalization; if A-pose is out of distribution, the generated canonical avatar will be unreliable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AdaHuman: Animatable Detailed 3D Human Generation with Compositional Multiview Diffusion." pith.science (2026). https://pith.science/paper/HNADPATV

@misc{pith2026250524877,
  author       = {Pith},
  title        = {Pith review of: AdaHuman: Animatable Detailed 3D Human Generation with Compositional Multiview Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HNADPATV}},
  note         = {Machine review of arXiv:2505.24877}
}
read the original abstract

Existing methods for image-to-3D avatar generation struggle to produce highly detailed, animation-ready avatars suitable for real-world applications. We introduce AdaHuman, a novel framework that generates high-fidelity animatable 3D avatars from a single in-the-wild image. AdaHuman incorporates two key innovations: (1) A pose-conditioned 3D joint diffusion model that synthesizes consistent multi-view images in arbitrary poses alongside corresponding 3D Gaussian Splats (3DGS) reconstruction at each diffusion step; (2) A compositional 3DGS refinement module that enhances the details of local body parts through image-to-image refinement and seamlessly integrates them using a novel crop-aware camera ray map, producing a cohesive detailed 3D avatar. These components allow AdaHuman to generate highly realistic standardized A-pose avatars with minimal self-occlusion, enabling rigging and animation with any input motion. Extensive evaluation on public benchmarks and in-the-wild images demonstrates that AdaHuman significantly outperforms state-of-the-art methods in both avatar reconstruction and reposing. Code and models will be publicly available for research purposes.

Figures

Figures reproduced from arXiv: 2505.24877 by the authors.

Figure 1
Figure 1. Given a single input image, AdaHuman reconstructs pixel-aligned a 3DGS avatar with detailed appearance. It can also generate [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Method Overview. Left: Given an RGB image of an unseen person as input, AdaHuman could (1) reconstruct a high-fidelity pixel-aligned 3D Gaussian Splat (3DGS) avatar, as well as (2) generate an reposed 3DGS avatar with a target pose condition, enable building animatable avatar in a standard A-pose. Right: A pose-conditioned joint 3D diffusion process is utilized to generate global or local 3DGS reconstruction or repo… view at source ↗
Figure 3
Figure 3. Compositional 3DGS Refinement. Given the coarse 3DGS reconstruction Gcoarse as input, we render initial coarse views, and refine them with image-to-image editing for enhancing local 3DGS Gupper, Glower, Ghead. Finally, a refined holistic 3DGS avatar Grefined is generated from these results by our proposed visibility-aware 3DGS Composition. into a holistic 3DGS avatar. The 3DGS generator in [45] uses four fixed canon… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on vatar reconstruction task on CustomHumans[12] dataset. Input AdaHuman (Ours) Coarse 3DGS Human3Diffusion SiTH SiFU [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison on in-the-wild images. AdaHuman generalizes well to images with diverse appearances, body shapes, and clothing styles, while SIFU[63] and SiTH[13] fail on loose and complex clothing, and Human3Diffusion [55] fail to preserve appearance details. Coarse 3DGS i…
Figure 6
Figure 6. Figure 6: Qualitative comparison on novel pose synthesis task. Baseline methods SiTH[13] SIFU[63] H3D[55] Coarse 3DGS Preference of AdaHuman (%) 88.3 99.2 79.7 93.8 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: AdaHuman generates animation-ready avatar in a standard pose, which can be animated with unseen input motion. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 9
Figure 9. Figure 9: compares the performance of these two modes. Here we analysis by comparing their pros and cons. Mode 1: Direct Avatar Reposing - This mode directly 3DGS avatars directly generated on the target pose sequence Animated standard pose avatars by skinning weights Input [PI…
Figure 8
Figure 8. Figure 8: Reposing avatars with challenging garments. Avatar Animation [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: Comparison of our method and ablation variants. ter FID while maintaining comparable PSNR, SSIM, and LPIPS scores, demonstrating that compositional refinement improves details without sacrificing accuracy. Composition Strategy. We compare our visibility-aware approach…
Figure 11
Figure 11. Figure 11: Network Architectures of (1) Pose-Conditioned Multi-View LDM Model and (2) Compositional 3DGS Generator. 2D diffusion to generate more detailed appearances. Dur￾ing image-to-image local refinement, we utilize SDEdit [31] with a strength of s = 0.5, meaning that denois…

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. CubifyGS: Object-Centric 3D Gaussian Splatting for Lifelong Dynamic Scene Maintenance

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    CubifyGS maintains 3DGS maps under rigid object rearrangements via reusable Gaussian assets, explicit prune/retrieve/align operations, and event-triggered local optimization, outperforming baselines on a new dynamic b...

Reference graph

Works this paper leans on

67 extracted references · 55 canonical work pages · cited by 1 Pith paper

  1. [1]

    Easymocap - make human motion capture easier. Github,

  2. [2]

    Single-image 3d human digitization with shape-guided diffusion

    Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia, 2023. 2

    work page 2023

  3. [3]

    Photorealistic monocular 3d reconstruction of humans wear- ing clothing

    Thiemo Alldieck, Mihai Zanfir, and Cristian Sminchisescu. Photorealistic monocular 3d reconstruction of humans wear- ing clothing. In CVPR, 2022. 2

    work page 2022

  4. [4]

    DreamAvatar: Text-and-Shape Guided 3D Human 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 Human Avatar Generation via Diffusion Models. In CVPR,

    work page

  5. [5]

    Effi- cient 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. Effi- cient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 2

    2022

  6. [6]

    Fan- tasia3d: Disentangling geometry and appearance for high- quality text-to-3d content creation

    Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fan- tasia3d: Disentangling geometry and appearance for high- quality text-to-3d content creation. In ICCV, 2023. 2

    work page 2023

  7. [7]

    Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dim- itrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In ECCV, pages 20–40, 2020. 2

    work page 2020

  8. [8]

    Objaverse: 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. Objaverse: A universe of annotated 3d objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13142–13153, 2023. 13

    work page 2023

Show all 67 references
  1. [9]

    9 StyleGAN-Human: A data-centric odyssey of human gener- ation

    Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. 9 StyleGAN-Human: A data-centric odyssey of human gener- ation. In ECCV, 2022. 6, 8

    2022

  2. [10]

    Srinivasan, Jonathan T

    Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. CAT3D: Create any- thing in 3d with multi-view diffusion models. In NeurIPS,

  3. [11]

    Arch++: Animation-ready clothed human re- construction revisited

    Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human re- construction revisited. In CVPR, 2021. 2

    2021

  4. [12]

    Learn- ing locally editable virtual humans

    Hsuan-I Ho, Lixin Xue, Jie Song, and Otmar Hilliges. Learn- ing locally editable virtual humans. In CVPR, 2023. 4, 5, 6, 7, 13

    2023

  5. [13]

    Sith: Single- view textured human reconstruction with image-conditioned diffusion

    Hsuan-I Ho, Jie Song, and Otmar Hilliges. Sith: Single- view textured human reconstruction with image-conditioned diffusion. In CVPR, 2024. 2, 6, 7, 13

    2024

  6. [14]

    LRM: Large reconstruction model for single image to 3d

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3d. In ICLR, 2024. 2, 4

    2024

  7. [15]

    2d gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields. In SIGGRAPH. Association for Com- puting Machinery, 2024. 12

    2024

  8. [16]

    Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation

    Xin Huang, Ruizhi Shao, Qi Zhang, Hongwen Zhang, and Ying Feng. Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation. InCVPR,

  9. [17]

    Dreamwaltz: Make a scene with complex 3d animatable avatars

    Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. In NeurIPS, 2023. 2

    2023

  10. [18]

    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 3DV, 2024. 1, 2

    2024

  11. [19]

    ARCH: Animatable reconstruction of clothed humans

    Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: Animatable reconstruction of clothed humans. In CVPR, 2020. 2

    2020

  12. [20]

    Avatarcraft: Transforming text into neural human avatars with parameter- ized shape and pose control

    Ruixiang Jiang, Can Wang, Jingbo Zhang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Avatarcraft: Transforming text into neural human avatars with parameter- ized shape and pose control. In ICCV, 2023. 2

    2023

  13. [21]

    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, 42 (4), 2023. 2, 3

    2023

  14. [22]

    Dreamhuman: Animatable 3d avatars from text

    Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Ed- uard Gabriel Bazavan, Mihai Fieraru, and Cristian Sminchis- escu. Dreamhuman: Animatable 3d avatars from text. In NeurIPS, 2023. 2

    2023

  15. [23]

    Instant 3d human avatar generation using image diffusion models

    Nikos Kolotouros, Thiemo Alldieck, Enric Corona, Ed- uard Gabriel Bazavan, and Cristian Sminchisescu. Instant 3d human avatar generation using image diffusion models. In ECCV, 2024. 2

    2024

  16. [24]

    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 3DV, 2024. 2

    2024

  17. [25]

    Magic3D: High-Resolution Text-to-3D Content Creation

    Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. In CVPR, 2023. 2

    2023

  18. [26]

    One-2-3-45: Any single image to 3d mesh in 45 seconds without per- shape optimization

    Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per- shape optimization. In NeurIPS, 2023. 2

    2023

  19. [27]

    One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion

    Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Ji- ayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In CVPR, 2024. 2

    2024

  20. [28]

    Zero-1-to-3: Zero-shot one image to 3d object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tok- makov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3d object. In CVPR, 2023. 2

    2023

  21. [29]

    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 CVPR,

  22. [30]

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi- person linear model. SIGGRAPH Asia, 2015. 2, 3

    2015

  23. [31]

    SDEdit: Guided image synthesis and editing with stochastic differential equa- tions

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia- jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equa- tions. In ICLR, 2022. 4, 13

    2022

  24. [32]

    Charactergen: Efficient 3d character generation from single images with multi-view pose canoni- calization

    Hao-Yang Peng, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, and Shi-Min Hu. Charactergen: Efficient 3d character generation from single images with multi-view pose canoni- calization. ACRM Trans. Graph., 43(4):1–13, 2024. 2

    2024

  25. [33]

    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 ICLR,

  26. [34]

    Lhm: Large animatable human reconstruction model from a single image in seconds

    Lingteng Qiu, Xiaodong Gu, Peihao Li, Qi Zuo, Weichao Shen, Junfei Zhang, Kejie Qiu, Weihao Yuan, Guanying Chen, Zilong Dong, and Liefeng Bo. Lhm: Large animatable human reconstruction model from a single image in seconds. In arXiv preprint arXiv:2503.10625, 2025. 2

    2025 arXiv

  27. [35]

    Hierarchical text-conditional image gen- eration with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen- eration with clip latents. arXiv preprint arXiv:2204.06125,

    arXiv

  28. [36]

    Texture: Text-guided texturing of 3d shapes

    Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. SIGGRAPH, 2023. 2

    2023

  29. [37]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 2, 4, 12

    2022

  30. [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 ICCV, 2019. 2

    2019

  31. [39]

    Saragih, and Han- byul Joo

    Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Han- byul Joo. PIFuHD: Multi-level pixel-aligned implicit func- tion for high-resolution 3d human digitization. In CVPR,

  32. [40]

    DiffHu- man: Probabilistic Photorealistic 3D Reconstruction of Hu- mans

    Akash Sengupta, Thiemo Alldieck, Nikos Kolotouros, Enric Corona, Andrei Zanfir, and Cristian Sminchisescu. DiffHu- man: Probabilistic Photorealistic 3D Reconstruction of Hu- mans. In CVPR, 2024. 2

    2024

  33. [41]

    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. 2

    arXiv

  34. [42]

    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. In ICLR, 2023. 2

    2023

  35. [43]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In ICLR, 2021. 3

    2021

  36. [44]

    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 ICLR, 2024. 2

    2024

  37. [45]

    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 ECCV,

  38. [46]

    Sizer: A dataset and model for parsing 3d clothing and learning size sensitive 3d clothing

    Garvita Tiwari, Bharat Lal Bhatnagar, Tony Tung, and Ger- ard Pons-Moll. Sizer: A dataset and model for parsing 3d clothing and learning size sensitive 3d clothing. In ECCV, pages 1–18. Springer, 2020. 6, 7

    2020

  39. [47]

    Triposr: Fast 3d object reconstruction from a single image

    Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 2

    2024 arXiv

  40. [48]

    Attention is all you need

    A Vaswani. Attention is all you need. Adv. Neural Inform. Process. Syst., 2017. 12

    2017

  41. [49]

    ProlificDreamer: High-fidelity and diverse text-to-3d generation with variational score dis- tillation

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and diverse text-to-3d generation with variational score dis- tillation. In NeurIPS, 2023. 2

    2023

  42. [50]

    MVHumanNet: A large- scale dataset of multi-view daily dressing human captures

    Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. MVHumanNet: A large- scale dataset of multi-view daily dressing human captures. In CVPR, 2024. 4, 5, 8, 12

    2024

  43. [51]

    Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit clothed humans obtained from normals. In CVPR, 2022. 2

    2022

  44. [52]

    Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit clothed humans optimized via normal integration. In CVPR, 2023. 2

    2023

  45. [53]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191,

    arXiv

  46. [54]

    DMV3D: Denoising multi-view diffusion using 3d large reconstruction model

    Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Ji- ahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. DMV3D: Denoising multi-view diffusion using 3d large reconstruction model. In ICLR, 2024. 2

    2024

  47. [55]

    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. InNeurIPS, 2024. 1, 2, 3, 4, 6, 7, 12, 13

    2024

  48. [56]

    S3: Neural shape, skeleton, and skinning fields for 3d human modeling

    Ze Yang, Shenlong Wang, Sivabalan Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinning fields for 3d human modeling. In CVPR, 2021. 2

    2021

  49. [57]

    GRM: Large gaussian reconstruction model for efficient 3d reconstruction and generation

    Xu Yinghao, Shi Zifan, Yifan Wang, Chen Hansheng, Yang Ceyuan, Peng Sida, Shen Yujun, and Wetzstein Gordon. GRM: Large gaussian reconstruction model for efficient 3d reconstruction and generation. In ECCV, 2024. 2

    2024

  50. [58]

    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 IEEE Conf. Comput. Vis. Pattern Recog., 2021. 13

    2021

  51. [59]

    Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACRM Trans. Graph., 2024. 12

    2024

  52. [60]

    GAvatar: Animatable 3d gaussian avatars with implicit mesh learning

    Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. GAvatar: Animatable 3d gaussian avatars with implicit mesh learning. In CVPR,

  53. [61]

    Avatarverse: High-quality & stable 3d avatar creation from text and pose

    Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Kang Du, and Min Zheng. Avatarverse: High-quality & stable 3d avatar creation from text and pose. In AAAI, 2024

    2024

  54. [62]

    Getavatar: Generative textured meshes for animatable human avatars

    Xuanmeng Zhang, Jianfeng Zhang, Chacko Rohan, Hongyi Xu, Guoxian Song, Yi Yang, and Jiashi Feng. Getavatar: Generative textured meshes for animatable human avatars. In ICCV, 2023. 2

    2023

  55. [63]

    SIFU: Side- view conditioned implicit function for real-world usable clothed human reconstruction

    Zechuan Zhang, Zongxin Yang, and Yi Yang. SIFU: Side- view conditioned implicit function for real-world usable clothed human reconstruction. In CVPR, 2024. 2, 6, 7, 13

    2024

  56. [64]

    Texmesh: Recon- structing detailed human texture and geometry from rgb-d video

    Tiancheng Zhi, Christoph Lassner, Tony Tung, Carsten Stoll, Srinivasa G Narasimhan, and Minh V o. Texmesh: Recon- structing detailed human texture and geometry from rgb-d video. In ECCV, 2020. 2

    2020

  57. [65]

    Idol: Instant photorealistic 3d human creation from a single image, 2024

    Yiyu Zhuang, Jiaxi Lv, Hao Wen, Qing Shuai, Ailing Zeng, Hao Zhu, Shifeng Chen, Yujiu Yang, Xun Cao, and Wei Liu. Idol: Instant photorealistic 3d human creation from a single image, 2024. 2

    2024

  58. [66]

    Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers

    Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, 2024. 2 11 A. Implementation Details Network Structure.In Fig. ...

    2024

  59. [2025]

    2, 3, 4, 5, 6, 7, 12, 13

Pith tools

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