Pith. sign in

REVIEW 5 major objections 8 minor 100 references

HuGeDiff: 3D Human Generation via Diffusion with Gaussian Splatting

T0 review · 5 major / 8 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 A text prompt yields an animatable 3D human avatar in about half a minute of compute, the paper claims, with rendering quality and prompt alignment that surpass current text-to-3D human methods.

desk verdict A genuinely new pipeline for text-to-3D human generation, but the SOTA claims are ahead of the evidence. read the letter →

arxiv 2506.04351 v1 pith:R2YGLLIY submitted 2025-06-04 cs.CV

Maksym Ivashechkin , Oscar Mendez , Richard Bowden This is my paper
classification cs.CV
keywords 3Dhumangenerationtext-to-3DGaussiansplattingpoint-clouddiffusiontransformercross-attentionsyntheticdatasetappearancecontrolSMPL-X
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that high-quality, text-controllable 3D human generation can be done far faster than current methods by combining a synthetic data loop with a direct image-to-point-cloud reconstruction and a point-cloud diffusion model. It proposes a weakly supervised pipeline: generate diverse human images with a stable diffusion model, fit 3D Gaussian splatting avatars to them, learn a transformer that maps image features straight to 3D Gaussian parameters, and then train a text-conditioned diffusion model on those parameters. If the paper is right, a single text prompt produces a renderable, re-articulable human avatar in roughly 0.45 minutes, orders of magnitude quicker than existing text-guided approaches, while also improving realism and faithfulness to the prompt. This would matter because 3D human generation is currently slow, expensive, and limited by low-quality human image datasets, and the paper's pipeline sidesteps the data problem by generating its own photorealistic training set.

What carries the argument

The load-bearing mechanism is a two-part transformer architecture. The first part maps image features to a 3D point cloud: the query is the canonical SMPL-X point set, the keys and values are image features from a Sapiens foundation model, and multi-head cross-attention assigns 2D information to 3D points, with an upsampling module and a self-attention layer that injects relative positional distances to refine dense point features. The second part is a point-cloud diffusion model that operates directly on 3DGS parameters (scales, displacements, colors) rather than on a compressed latent space, using an adapted U-Net with MLP and attention on k-nearest neighbors. Together they convert a single image into a Gaussian-splat human and then let text prompts generate new splat parameter sets that can be rendered and re-posed through SMPL-X.

What would settle it

Feed the trained diffusion model prompts that are absent from the template set, such as 'woman with long flowing hair wearing a floral-patterned dress and kneeling', and check whether the output actually matches the requested clothing pattern, hair length, pose, and text alignment; a consistent failure to reproduce patterns or long hair, or a large drop in LLM alignment scores compared with template prompts, would show that the synthetic dataset's coverage is insufficient for the paper's general text-to-3D claim.

Watch

Extended reading notes

Core claim

The central claim is that a closed loop of synthetic data generation, UV-free transformer-based 3D reconstruction, and text-conditioned point-cloud diffusion lets a text prompt directly produce a 3D Gaussian Splatting human with high fidelity and appearance control. The method first uses FLUX to generate about 10,000 images from template prompts that vary race, gender, hair, clothing, colors, and viewpoint. Each image is segmented and fitted to SMPL-X with 3DGS parameters, and a transformer cross-attention model learns to map Sapiens image features to canonical SMPL-X mesh points, estimating scales, displacements, and colors without any UV atlas. Finally, a diffusion model trained on these 3DGS parameters (scales, displacements, colors) with classifier-free guidance generates new humans from the same text prompts. The paper reports that this approach renders more realistic hands and faces than competitors such as HumanGaussian, DreamWaltz-G, and TADA, and reduces generation time from tens of minutes or hours to under a minute, while also outperforming them on LLM-based text-alignment and aesthetic scores.

Load-bearing premise

The whole pipeline depends on the assumption that the 10,000 synthetic images from the fixed template prompts (plain clothing, T-pose, front/back/side views, short hair) cover the appearance and pose space well enough that a model trained only on them can faithfully follow arbitrary text prompts from real users.

Editorial extensions

If this is right

  • A text prompt can generate a fully textured, animatable 3D human in about 0.45 minutes, making interactive text-to-avatar creation practical for games, VR, and telepresence.
  • The UV-free mapping means reconstruction and generation extend to any body pose and viewpoint without precomputed UV coordinates, simplifying real-world deployment.
  • Classifier-free guidance gives users a tunable balance between sticking closely to the text prompt and producing varied appearances.
  • Because generated avatars are anchored to SMPL-X, they can be re-articulated into new poses, enabling animation from a single generated asset.
  • The released code and synthetic dataset would give other groups a ready-made base for 3D human generative models without expensive real-world human data collection.

Reading between the lines

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

  • The diffusion model can only reproduce the appearance distribution present in the synthetic FLUX images; asking for clothing patterns, long hair, or poses outside the fixed template list would likely degrade prompt alignment, since the paper's own limitations note that patterns and long hair are not represented.
  • The same image-feature-to-point-cloud cross-attention design is not human-specific beyond the SMPL-X anchor mesh, so it could plausibly be retargeted to other articulated categories (animals, hands, avatars without bodies) by swapping the canonical point set.
  • The speed claim relies on a single fixed 3DGS parameterization and a small number of evaluation samples; a larger, standardized benchmark with human raters and a much broader prompt set would be needed to confirm that the quality edge over competitors holds outside the tested distribution.
  • If the synthetic dataset grows, the pipeline could become a seed for a foundational 3D human model, since the weakly supervised loop decouples data generation from real-world collection.
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

5 major / 8 minor

Summary. The paper presents HuGeDiff, a weakly supervised three-stage pipeline for text-to-3D human generation. The pipeline first uses FLUX to synthesize a dataset of ~10,000 human images from template prompts with controlled attributes (race, gender, hair, clothing, color, view). It then fits SMPL-X and 3D Gaussian Splatting (3DGS) parameters to those images and trains a transformer-based model that maps Sapiens image features to 3DGS parameters anchored to a canonical SMPL-X mesh, using a UV-free cross-attention mechanism with an upsampling module. Finally, it trains a text-conditioned diffusion model directly on the 3DGS parameters (scales, displacements, colors) using classifier-free guidance. The authors claim state-of-the-art rendered quality, speed, and appearance control, with orders-of-magnitude speed-ups relative to existing text-driven methods.

Significance. If the claims were substantiated, HuGeDiff would offer a practical and fast text-to-3D human generation approach that avoids UV-template dependency and integrates synthetic data generation, reconstruction, and generative modeling in one loop. The conceptual pipeline is interesting, and the authors are to be credited for a fully integrated system and for committing to release code and data. However, the current evidence is insufficient to support the central claims: the only quantitative evaluation uses 10 LLM-scored renders per method without error bars; the speed comparison is not controlled for conditioning; and the model is trained and evaluated on a single in-distribution prompt template, so reported alignment may reflect memorization. These issues are addressable, but until they are fixed the conclusions in Section 6 outrun the evidence.

major comments (5)
  1. [§4.2, Table 2] The central claim that HuGeDiff 'outperforms the current state-of-the-art approaches in 3D human generation regarding rendered quality' rests on Table 2, which reports only 10 rendered images per method scored by four LLMs on a 0-1 scale. No standard deviations, confidence intervals, inter-rater agreement, or per-prompt scores are provided, and no human validation is included. This evidence is too thin to support the strong comparative claim. Please report results across multiple seeds and prompts with variance measures, and include either a human evaluation or a validated automatic metric with known correlation to human judgment.
  2. [§4.1, Table 1; Abstract] The abstract's claim of 'orders-of-magnitude speed-ups in 3D human generation compared to the state-of-the-art approaches' is not supported by Table 1: Ours (0.45 min) is slower than SMPLitex (0.29 min) and slower than the unconditional baselines StructLDM (0.10), GSM (0.12), and EVA3D (0.16). The table mixes appearance-controlled and non-controlled methods, and the speed advantage is only relative to optimization-based text-conditioned methods. Please restrict the speed claim to that comparison class, state the exact hardware and implementation details for all methods, and report timings that are directly comparable in terms of conditioning and output format.
  3. [§3.3, §4.3, Supp. §9] The text-conditioned diffusion model is trained and evaluated on prompts drawn from a single template distribution ('a real {race} {gender} with {hair} hair ... tight {color} {top} and tight {color} trousers ... T-pose'). The paper does not state whether any evaluation prompt falls outside this template, and the limitations in Section 5 and Supp. Section 9 confirm that the model is constrained to this template. Consequently, the reported text-alignment scores may reflect in-distribution memorization rather than general text-to-3D control. Please provide quantitative evaluation on out-of-template prompts, for example prompts with different poses, clothing types, accessories, free-form descriptions, or prompts combining attributes not seen during training, and report whether the model generalizes.
  4. [§3.2, Multi-view Regularization] The multi-view supervision used to regularize the image-conditioned reconstruction model is obtained by rendering outputs of a text-conditioned reconstruction model that is itself trained on the same synthetic dataset and is part of the same closed loop. No external ground truth is introduced anywhere in the pipeline, so errors from FLUX synthesis or SMPL-X fitting can propagate through reconstruction and diffusion without correction. This is a load-bearing concern for the validity of the generated 3D humans. Please validate the reconstruction stage against real multi-view human datasets (e.g., THuman, MVHumanNet) with standard novel-view metrics (PSNR, SSIM, LPIPS) and report quantitative results on held-out real images, not only qualitative examples from SHHQ.
  5. [§4.2] The paper dismisses CLIP as unreliable but does not validate the LLM-based evaluation protocol it replaces it with. The evaluation prompt given to the LLMs is vague (asking for 'prompt alignment' and 'aesthetics' on a 0-1 scale), and different LLMs may interpret the scale inconsistently. Please report inter-rater agreement (e.g., Krippendorff's alpha or pairwise correlations), show score distributions per prompt, and discuss how these LLM scores correlate with human preference. Without this, the quantitative results in Table 2 are difficult to interpret.
minor comments (8)
  1. [§3.2] The sentence 'For displacements and scales, we found that sigmoid can suffer from vanishing gradients, leading' is incomplete; it appears to be cut off at the end of the subsection. Please complete the sentence or rephrase.
  2. [§3.2] The phrase "the model tends to generate an 'average'; human appearance" contains a stray semicolon; it should read "an 'average' human appearance".
  3. [§4.1, Table 1] The method name 'EV A3d' is inconsistent with 'EVA3D' used in the text; please use one consistent spelling throughout.
  4. [§7 and Abstract] The paper states 'We will make the code and dataset available' but provides no release plan, repository link, or license information. Please add these details or a clear timeline for release.
  5. [§5] The limitation 'our model is still limited by certain combinations of the text prompts discussed in the Supplementary Materials' is not actually elaborated in Supp. Section 9, which only lists the template options. Please specify which combinations are problematic and why.
  6. [Figure 2 and Figure 3] The red circles highlighting artifacts are small and hard to discern in the rendered images; please provide zoomed crops or higher-resolution figures to make the comparisons visible.
  7. [§4.2] The statement 'Almost 6 days of continuous GPU computing were needed to generate competitor results' would be more informative if the hardware and the number of generated samples per method are specified.
  8. [References] Reference [1] for Gemini is incomplete (no authors, no technical report identifier); please update the citation. Also, in the Related Work, the citation '[4, 4]' duplicates reference [4].

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the pipeline composes independent stages; evaluation and data are in-distribution, which is a generalization weakness rather than a definitional circularity.

full rationale

The paper's derivation chain is not circular under the definitions used here. Stage 1 (Sec. 3.1) generates a synthetic dataset with FLUX from template prompts, then fits SMPL-X/3DGS parameters to those images; Stage 2 (Sec. 3.2) trains a transformer to regress 3DGS parameters from Sapiens image features; Stage 3 (Sec. 3.3) trains a text-conditioned diffusion model on those 3DGS parameters. Each stage has its own objective (rendering loss, cross-attention regression, denoising L2), and no equation in the paper makes an output equal to an input by construction. The phrase "we close the loop by training a point-cloud diffusion model that is conditioned on the same text prompts used to generate the original samples" describes a training-distribution overlap, not a formal reduction. The acknowledged limitations in Sec. 5 and Supp. Sec. 9 (template-only prompts, missing clothing patterns, short hair) are external-validity concerns: the model may not generalize beyond the template, and the Sec. 6 SOTA claim rests on 10 LLM-scored renders without error bars plus an uncontrolled speed comparison. These are evidence-quality issues, not circularity. The Sec. 3.2 multi-view regularizer uses renderings from the authors' own text-conditioned reconstruction model as pseudo-ground-truth; while this is a self-referential supervision loop, the paper discloses it and the final generative model is not defined by that supervision alone. Consequently there is no load-bearing reduction that would warrant a circularity score above 0.

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

The pipeline relies on a chain of unverified assumptions: the SMPL-X prior and pose estimators, the opacity/rotation restrictions, the hallucinated multi-view supervision, and the representativeness of the synthetic template data. No external ground truth is used, so the central claim rests entirely on these internally defined components.

free parameters (5)
  • Dataset min/max normalization bounds (x_min, x_max) for 3DGS parameters = computed from the synthetic training set, not stated
    Section 3.3 normalizes scales/displacements/colors to [-1,1] using per-parameter min/max; the model's generative distribution is anchored to these bounds.
  • Classifier-free guidance scale = not stated; adjustable
    Section 3.3 says the guidance scale controls alignment vs diversity, but no specific value or sensitivity analysis is reported.
  • Loss weights (LPIPS 0.15, L1 1.0, SSIM 0.25) = 0.15 / 1.0 / 0.25
    Section 4.1 lists these weights without justification or ablation.
  • Mesh densification thresholds = thresholds that produce 84,317 points
    Section 4.1 densifies the SMPL-X mesh by recursively splitting edges/faces based on length/area thresholds; these thresholds are hand-set.
  • Template prompt attributes and uniform sampling probabilities = uniform over listed options
    Supplementary Section 9 defines the prompt template; the choice of attribute vocabulary is ad hoc and shapes the whole data distribution.
assumptions (4)
  • domain assumption SMPL-X is an adequate body prior and ExPose/HaMeR provide sufficiently accurate poses
    Section 3.1 uses off-the-shelf SMPL-X fitting; if poses are misaligned, the 3DGS fitting inherits the error.
  • ad hoc to paper Setting all opacities to 1 and aligning Gaussian rotations to surface normals does not remove important appearance information
    Section 3.1 introduces this regularization as 'without loss of generality', but 3DGS opacity is a key degree of freedom; the paper provides no experiments demonstrating that this restriction preserves appearance.
  • ad hoc to paper Rendering the authors' own text-conditioned reconstruction model from four views provides valid multi-view supervision
    Section 3.2 uses hallucinated back/left/right views to regularize the feature-based model; if these renderings are inaccurate, the model is trained on false targets.
  • standard math Standard diffusion and attention math (softmax attention, classifier-free guidance, positional encoding) is correct
    Equations in Section 3.2 and 3.3 rely on standard attention and diffusion formulations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HuGeDiff: 3D Human Generation via Diffusion with Gaussian Splatting." pith.science (2026). https://pith.science/paper/R2YGLLIY

@misc{pith2026250604351,
  author       = {Pith},
  title        = {Pith review of: HuGeDiff: 3D Human Generation via Diffusion with Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R2YGLLIY}},
  note         = {Machine review of arXiv:2506.04351}
}
read the original abstract

3D human generation is an important problem with a wide range of applications in computer vision and graphics. Despite recent progress in generative AI such as diffusion models or rendering methods like Neural Radiance Fields or Gaussian Splatting, controlling the generation of accurate 3D humans from text prompts remains an open challenge. Current methods struggle with fine detail, accurate rendering of hands and faces, human realism, and controlability over appearance. The lack of diversity, realism, and annotation in human image data also remains a challenge, hindering the development of a foundational 3D human model. We present a weakly supervised pipeline that tries to address these challenges. In the first step, we generate a photorealistic human image dataset with controllable attributes such as appearance, race, gender, etc using a state-of-the-art image diffusion model. Next, we propose an efficient mapping approach from image features to 3D point clouds using a transformer-based architecture. Finally, we close the loop by training a point-cloud diffusion model that is conditioned on the same text prompts used to generate the original samples. We demonstrate orders-of-magnitude speed-ups in 3D human generation compared to the state-of-the-art approaches, along with significantly improved text-prompt alignment, realism, and rendering quality. We will make the code and dataset available.

Figures

Figures reproduced from arXiv: 2506.04351 by the authors.

Figure 1
Figure 1. Overview of the proposed pipeline. The left side shows the data generation part, the middle part is the 3DGS reconstruction from image features. The right side highlights the text-conditioned diffusion model. age features. We estimate initial SMPL-X parameters using ExPose [9], and refine hand poses with HaMeR [52]. Since ExPose reconstructions are often misaligned, we initialize our Gaussian splatting model with th… view at source ↗
Figure 2
Figure 2. Comparison of rendered images generated with text prompt conditioning. The rendering artifacts and problems on hands and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Comparison to images rendered by state-of-the-art models that do not support text conditioning (we used random prompts for our method). Artifacts and issues, particularly on hands and faces, are circled in red [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Novel view reconstruction on out-of-distribution image from the SHHQ [ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Examples of novel view rendering of a 3D human generated via the proposed diffusion model. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: A generated human via our diffusion model rendered in various SMPL-X articulation poses from the AGORA [ [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Examples of images generated by the FLUX diffusion [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Same prompt with different seed diffusion sampling [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 10
Figure 10. Figure 10: Novel view reconstruction. The first two images show the input image (left) and its 3D reconstruction in the same articulation [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Novel view reconstruction on out-of-distribution images from the SHHQ [ [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

100 extracted references · 62 canonical work pages

  1. [1]

    Gemini: A family of highly capable multimodal models,

  2. [2]

    Gaussian Shell Maps for Efficient 3D Human Generation

    Rameen Abdal, Wang Yifan, Zifan Shi, Yinghao Xu, Ryan Po, Zhengfei Kuang, Qifeng Chen, Dit-Yan Yeung, and Gor- don Wetzstein. Gaussian Shell Maps for Efficient 3D Human Generation . In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9441–9451, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2, 5, 7

    2024

  3. [3]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid- based neural radiance fields.ICCV, 2023. 2

    2023

  4. [4]

    Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and Kwan- Yee K. Wong. Dreamavatar: Text-and-shape guided 3d hu- man avatar generation via diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 958–968, 2024. 2, 5, 6

    2024

  5. [5]

    SMPLitex: A Gener- ative Model and Dataset for 3D Human Texture Estimation from Single Image

    Dan Casas and Marc Comino-Trinidad. SMPLitex: A Gener- ative Model and Dataset for 3D Human Texture Estimation from Single Image. InBritish Machine Vision Conference (BMVC), 2023. 2, 5, 6

    2023

  6. [6]

    Qi Charles, Hao Su, Mo Kaichun, and Leonidas J

    R. Qi Charles, Hao Su, Mo Kaichun, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77–85, 2017. 5

    2017

  7. [7]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InEuropean Conference on Computer Vision (ECCV), 2022. 2

    2022

  8. [8]

    Learning 3d shape latent for point cloud completion.IEEE Transactions on Multime- dia, 26:8717–8729, 2024

    Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wen- gang Zhou, Jiebo Luo, and Tao Mei. Learning 3d shape latent for point cloud completion.IEEE Transactions on Multime- dia, 26:8717–8729, 2024. 5

    2024

Show all 100 references
  1. [9]

    Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dim- itrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. InEuropean Conference on Computer Vision (ECCV), 2020. 3

    2020

  2. [10]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. InProceedings of the 35th Inter- national Conference on Neural Information Processing Sys- tems, Red Hook, NY , USA, 2021. Curran Associates Inc. 5

    2021

  3. [11]

    An image is worth 16x16 words: Transformers for image recognition at scale.ICLR, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  4. [12]

    Scaling rec- tified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rec- tified flow transformers for high-resolution image syn...

    2024

  5. [13]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. InCVPR, 2022. 2

    2022

  6. [14]

    Stylegan-human: A data-centric odyssey of human genera- tion

    Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human genera- tion. InComputer Vision – ECCV 2022: 17th European Con- ference, Tel Aviv, Israel, October 23–27, 2022, Proceedings,...

    2022

  7. [15]

    Text-guided 3d human generation from 2d collections

    Tsu-Jui Fu, Wenhan Xiong, Yixin Nie, Jingyu Liu, Barlas Ouguz, and William Yang Wang. Text-guided 3d human generation from 2d collections. InConference on Empiri- cal Methods in Natural Language Processing, 2023. 2

    2023

  8. [16]

    Text2avatar: text to 3d human avatar generation with codebook-driven body controllable attribute

    Chaoqun Gong, Yuqin Dai, Ronghui Li, Achun Bao, Jun Li, Jian Yang, Yachao Zhang, and Xiu Li. Text2avatar: text to 3d human avatar generation with codebook-driven body controllable attribute. InICASSP 2024-2024 IEEE Interna- tional Conference on Acoustics, Speech and Signal Pro...

    2024

  9. [17]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. InAdvances in neural information processing systems, pages 2672–2680,

  10. [18]

    Classifier-free diffusion guidance.ArXiv, abs/2207.12598, 2022

    Jonathan Ho. Classifier-free diffusion guidance.ArXiv, abs/2207.12598, 2022. 5

    2022 arXiv

  11. [19]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. InProceedings of the 34th Inter- national Conference on Neural Information Processing Sys- tems, Red Hook, NY , USA, 2020. Curran Associates Inc. 5

    2020

  12. [20]

    EV A3d: Compositional 3d human generation from 2d image collections

    Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EV A3d: Compositional 3d human generation from 2d image collections. InInternational Conference on Learning Representations, 2023. 2, 5, 7

    2023

  13. [21]

    GaussianAvatar: Towards realistic human avatar modeling from a single video via animatable 3D gaussians.arXiv preprint arXiv:2312.02134, 2023

    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.arXiv preprint arXiv:2312.02134, 2023. 2

    2023 arXiv

  14. [22]

    Structldm: Struc- tured latent diffusion for 3d human generation

    Tao Hu, Fangzhou Hong, and Ziwei Liu. Structldm: Struc- tured latent diffusion for 3d human generation. InComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part LI, page 363–381, Berlin, Heidelberg, 2024. Springer-Verlag. 5

    2024

  15. [23]

    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 Advances in Neural Information Processing Systems, pages 4566–4584, 2023. 6

    2023

  16. [24]

    DreamWaltz-G: Expressive 3D Gaussian Avatars from Skeleton-Guided 2D Diffusion

    Yukun Huang, Jianan Wang, Ailing Zeng, Zheng-Jun Zha, Lei Zhang, and Xihui Liu. DreamWaltz-G: Expressive 3D Gaussian Avatars from Skeleton-Guided 2D Diffusion. 2024. 2, 5, 6

    2024

  17. [25]

    Humanrf: High-fidelity neural radiance fields for humans in motion.ACM Transactions on Graphics (TOG), 42(4):1–12, 2023

    Mustafa Is ¸ık, Martin R ¨unz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion.ACM Transactions on Graphics (TOG), 42(4):1–12, 2023. 2 9

    2023

  18. [26]

    Human3.6m: Large scale datasets and predic- tive methods for 3d human sensing in natural environments

    Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predic- tive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 36(7):1325–1339, 2014. 2

    2014

  19. [27]

    Black, David W

    Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. InComputer Vision and Pattern Recognition (CVPR),

  20. [28]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023

    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

    2023

  21. [29]

    Sapiens: Foundation for human vision mod- els.arXiv preprint arXiv:2408.12569, 2024

    Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision mod- els.arXiv preprint arXiv:2408.12569, 2024. 2

    2024 arXiv

  22. [30]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-Encoding Vari- ational Bayes. In2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14- 16, 2014, Conference Track Proceedings, 2014. 5

    2014

  23. [31]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-Supervised Classi- fication with Graph Convolutional Networks. InProceedings of the 5th International Conference on Learning Representa- tions, 2017. 4

    2017

  24. [32]

    Avat3r: Large an- imatable gaussian reconstruction model for high-fidelity 3d head avatars, 2025

    Tobias Kirschstein, Javier Romero, Artem Sevastopolsky, Matthias Nießner, and Shunsuke Saito. Avat3r: Large an- imatable gaussian reconstruction model for high-fidelity 3d head avatars, 2025. 4

    2025

  25. [33]

    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 Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY , USA...

    2023

  26. [34]

    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. InComputer Vision – ECCV 2024: 18th European Con- ference, Milan, Italy, September 29–October 4, 2024, Pro- ce...

    2024

  27. [35]

    Flux.https://github.com/ black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/ black-forest-labs/flux, 2024. 1, 2

    2024

  28. [36]

    Gaussiananything: Interactive point cloud latent diffu- sion for 3d generation

    Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffu- sion for 3d generation. InICLR, 2025. 5

    2025

  29. [37]

    Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion.arXiv preprint arXiv:2409.10141, 2024

    Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yang- guang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion.arXiv preprint arXiv:2409.10141, 2024. 2

    2024 arXiv

  30. [38]

    Tingting Liao, Hongwei Yi, Yuliang Xiu, Jiaxiang Tang, Yangyi Huang, Justus Thies, and Michael J. Black. TADA! Text to Animatable Digital Avatars. InInternational Confer- ence on 3D Vision (3DV), 2024. 2, 5, 6

    2024

  31. [39]

    One-stage 3d whole-body mesh recovery with component aware transformer

    Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 21159–21168, 2023. 2, 3

    2023

  32. [40]

    Humangaus- sian: Text-driven 3d human generation with gaussian splat- ting.arXiv preprint arXiv:2311.17061, 2023

    Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaus- sian: Text-driven 3d human generation with gaussian splat- ting.arXiv preprint arXiv:2311.17061, 2023. 2, 5, 6

    2023 arXiv

  33. [41]

    Deepfashion: Powering robust clothes recog- nition and retrieval with rich annotations

    Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xi- aoou Tang. Deepfashion: Powering robust clothes recog- nition and retrieval with rich annotations. InProceedings of IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2016. 7

    2016

  34. [42]

    Matthew Loper, Naureen Mahmood, Javier Romero, Ger- ard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model.ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1–248:16, 2015. 2

    2015

  35. [43]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. InInternational Conference on Learning Representations, 2017. 5

    2017

  36. [44]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. InECCV, 2020. 5

    2020

  37. [45]

    Expressive whole-body 3D gaussian avatar

    Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. InECCV, 2024. 2

    2024

  38. [46]

    Instant neural graphics primitives with a multires- olution hash encoding.ACM Trans

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- olution hash encoding.ACM Trans. Graph., 41(4):102:1– 102:15, 2022. 5

    2022

  39. [47]

    Point-e: A system for generating 3d point clouds from complex prompts, 2022

    Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts, 2022. 2

    2022

  40. [48]

    Yang, Zach DeVito, Mar- tin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zem- ing Lin, Natalia Gimelshein, Luca Antiga, Alban Desmai- son, Andreas K ¨opf, Edward Z. Yang, Zach DeVito, Mar- tin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Stei...

    1912 arXiv

  41. [49]

    Huang, Joachim Tesch, David T

    Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regres- sion analysis. InProceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 8

    2021

  42. [50]

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

    2019

  43. [51]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, 10 and body from a single image. InProceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2

    2019

  44. [52]

    Reconstruct- ing hands in 3D with transformers

    Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstruct- ing hands in 3D with transformers. InCVPR, 2024. 3

    2024

  45. [53]

    Marco Pesavento, Yuanlu Xu, Nikolaos Sarafianos, Robert Maier, Ziteng Wang, Chun-Han Yao, Marco V olino, Edmond Boyer, Adrian Hilton, and Tony Tung. Anim: Accurate neural implicit model for human reconstruction from a sin- gle rgb-d image.2024 IEEE/CVF Conference on Computer V...

    2024

  46. [54]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 1, 2

    2023

  47. [55]

    Qi, Li Yi, Hao Su, and Leonidas J

    Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point- net++: deep hierarchical feature learning on point sets in a metric space. InProceedings of the 31st International Conference on Neural Information Processing Systems, page 5105–5114, Red Hook, NY , USA, 2017. Curran...

    2017

  48. [56]

    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

  49. [57]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InInternational Conf...

    2021

  50. [58]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junt- ing Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao- Yuan Wu, Ross Girshick, Piotr Doll´ar, and Christoph Feic...

    2024 arXiv

  51. [59]

    High-resolution image syn- thesis with latent diffusion models.CoRR, abs/2112.10752,

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models.CoRR, abs/2112.10752,

    arXiv

  52. [60]

    High-Resolution Image Synthesis with Latent Diffusion Models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models . In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674–10685, Los Alamitos, CA, USA,

  53. [61]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 5

    2015 arXiv

  54. [62]

    Tensor4d: Efficient neu- ral 4d decomposition for high-fidelity dynamic reconstruc- tion and rendering

    Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neu- ral 4d decomposition for high-fidelity dynamic reconstruc- tion and rendering. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2

    2023

  55. [63]

    X- Avatar: Expressive human avatars

    Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X- Avatar: Expressive human avatars. InCVPR, 2023. 2

    2023

  56. [64]

    Difu: Depth-guided implicit function for clothed human reconstruction

    Dae-Young Song, HeeKyung Lee, Jeongil Seo, and Donghyeon Cho. Difu: Depth-guided implicit function for clothed human reconstruction. In2023 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 8738–8747, 2023. 2

    2023

  57. [65]

    A-nerf: articulated neural radiance fields for learn- ing human shape, appearance, and pose

    Shih-Yang Su, Frank Yu, Michael Zollh ¨ofer, and Helge Rhodin. A-nerf: articulated neural radiance fields for learn- ing human shape, appearance, and pose. InProceedings of the 35th International Conference on Neural Information Processing Systems, Red Hook, NY , USA, 2024. Cu...

    2024

  58. [66]

    Deep- cloth: Neural garment representation for shape and style editing.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1581–1593, 2023

    Zhaoqi Su, Tao Yu, Yangang Wang, and Yebin Liu. Deep- cloth: Neural garment representation for shape and style editing.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1581–1593, 2023. 2

    2023

  59. [67]

    Graph at- tention networks.6th International Conference on Learning Representations, 2017

    Petar Veli ˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li `o, and Yoshua Bengio. Graph at- tention networks.6th International Conference on Learning Representations, 2017. 4

    2017

  60. [68]

    Sarma, Michael M

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds.ACM Transactions on Graphics (TOG), 2019. 5

    2019

  61. [69]

    Hu- manNeRF: Free-viewpoint rendering of moving people from monocular video

    Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Hu- manNeRF: Free-viewpoint rendering of moving people from monocular video. InCVPR, 2022. 2

    2022

  62. [70]

    Srinivasan, Jonathan T

    Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Hu- manNeRF: Free-viewpoint rendering of moving people from monocular video. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pa...

    2022

  63. [71]

    Pointconv: Deep convolutional networks on 3d point clouds

    Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621–9630, 2019. 5

    2019

  64. [72]

    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. InProceedings of the IEEE/CVF Conference on Computer...

    2024

  65. [73]

    Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13296–13306, 2022. 2

    2022

  66. [74]

    H- nerf: neural radiance fields for rendering and temporal recon- struction of humans in motion

    Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H- nerf: neural radiance fields for rendering and temporal recon- struction of humans in motion. InProceedings of the 35th 11 International Conference on Neural Information Processing Systems, Red Hook, NY , USA, 2024. Cur...

    2024

  67. [75]

    Pons-Moll

    Yuxuan Xue, Xianghui Xie, Riccardo Marin, and Gerard. Pons-Moll. Human 3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models. 2024. 2

    2024

  68. [76]

    Sigman:scaling 3d human gaussian gen- eration with millions of assets, 2025

    Yuhang Yang, Fengqi Liu, Yixing Lu, Qin Zhao, Pingyu Wu, Wei Zhai, Ran Yi, Yang Cao, Lizhuang Ma, Zheng-Jun Zha, and Junting Dong. Sigman:scaling 3d human gaussian gen- eration with millions of assets, 2025. 1, 2, 3, 6

    2025

  69. [77]

    Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models

    Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. InCVPR,

  70. [78]

    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 Conference on Computer Vision and Pattern Recogni- tion (CVPR2021), 2021. 2

    2021

  71. [79]

    Dwnet: Dense warp-based network for pose- guided human video generation

    Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose- guided human video generation. In30th British Machine Vi- sion Conference 2019, BMVC 2019, Cardiff, UK, September 9-12, 2019, page 51. BMV A Press, 2019. 7

    2019

  72. [80]

    Lion: Latent point diffusion models for 3d shape generation

    xiaohui zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. InAdvances in Neural Information Processing Systems, pages 10021– 10039. Curran Associates, Inc., 2022. 5

    2022

  73. [81]

    Adding conditional control to text-to-image diffusion models, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 1, 2, 7

    2023

  74. [82]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In2018 IEEE Con- ference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, p...

    2018

  75. [83]

    E3gen: Efficient, expressive and ed- itable avatars generation

    Weitian Zhang, Yichao Yan, Yunhui Liu, Xingdong Sheng, and Xiaokang Yang. E3gen: Efficient, expressive and ed- itable avatars generation. InProceedings of the 32nd ACM International Conference on Multimedia, page 6860–6869,

  76. [84]

    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. InAdvances in Neural Infor- mation Processing Systems (NeurIPS), 2023. 1, 2, 4

    2023

  77. [85]

    Point transformer

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 16259–16268, 2021. 5

    2021

  78. [86]

    Structured local radiance fields for human avatar modeling

    Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yan- dong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  79. [87]

    Diffgs: Functional gaussian splatting diffusion

    Junsheng Zhou, Weiqi Zhang, and Yu-Shen Liu. Diffgs: Functional gaussian splatting diffusion. InAdvances in Neu- ral Information Processing Systems (NeurIPS), 2024. 5

    2024

  80. [88]

    Champ: Controllable and consistent human image animation with 3d parametric guidance

    Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. InEuropean Conference on Computer Vision (ECCV), 2024. 2

    2024

  81. [89]

    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. 1, 2, 3, 4 12 Supplementary Material: HuGeDiff: 3D Human Generation via Diffusion with Ga...

    2024

  82. [93]

    With query (Q)being the point cloud of sizeN, andMimage features with dimensionCare both keys(K)and values(V),his number of heads: FX =softmax(Q K⊤)V,(9) Figure 9

    Description of Equations Decoupling of image features to the point cloud. With query (Q)being the point cloud of sizeN, andMimage features with dimensionCare both keys(K)and values(V),his number of heads: FX =softmax(Q K⊤)V,(9) Figure 9. Novel view rendering of 3D humans gener...

  83. [94]

    a real{race} {gender}with{hair}hair standing upright wearing tight{color} {top}and tight{color}trousers and {color}trainers,{view}view, arms stretched horizontally in a T-pose

    Text Prompts Description The prompt given to the FLUX diffusion model was the fol- lowing: “a real{race} {gender}with{hair}hair standing upright wearing tight{color} {top}and tight{color}trousers and {color}trainers,{view}view, arms stretched horizontally in a T-pose” The opti...

  84. [95]

    Humans generated by the diffusion model using the same prompt but different seeds are shown in Fig. 8

  85. [96]

    3D human reconstructions from single images are shown in Fig. 10. 14

  86. [97]

    Reconstruction results on out-of-distribution data are shown in Fig. 11

  87. [98]

    Novel views of the 3D humans generated via the diffu- sion model are shown in Fig. 9

  88. [99]

    The prompt we gave to the LLM for the evalua- tion is the following: “Evaluate the input images in terms of prompt alignment and aesthetic quality

    LLM Evaluation The large language models (LLM) we used to evaluate text- prompt alignment and image aesthetics are Gemini 2.5 Pro Preview [1], Claude 3.7 Sonnet, Microsoft Copilot, and Grok 3. The prompt we gave to the LLM for the evalua- tion is the following: “Evaluate the i...

  89. [100]

    The idea of this work is to provide a theoretical background for generating diverse avatars and supporting anonymity by replacing a real appearance with a generated one

    Ethical Statement The generated and rendered images, as well as manually de- fined text-prompt combinations, do not intend to discrimi- nate or exclude other races, nationalities, or any other iden- tities. The idea of this work is to provide a theoretical background for gener...

  90. [595]

    Computer Vision Foundation / IEEE Computer Society,

  91. [2022]

    IEEE Computer Society. 1, 2, 6

  92. [2024]

    Springer-Verlag. 2, 6

Pith tools

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