Pith. sign in

REVIEW 4 major objections 6 minor 62 references

TGRHuman: Text-Guided Realistic 3D Human Generation via Diffusion Renderer

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read TGRHuman generates realistic 3D humans from text in about five minutes by generating four-view 1024-pixel normal maps, carving the SMPL template, and painting texture with a diffusion renderer.

desk verdict TGRHuman is a solid, practical text-to-3D human pipeline that delivers fast explicit meshes without SDS, but the loose-clothing claim is asserted more than measured and the evaluation needs more transparency. read the letter →

arxiv 2608.12175 v1 pith:BO2P7IB6 submitted 2026-08-12 cs.CV

classification cs.CV
keywords text-to-3Dgeneration3Dhumandiffusionmodelsmulti-viewnormalmapsgeometrycarvingtexturesynthesisrendererSMPL
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

TGRHuman sets out to show that text-to-3D human generation does not need slow score-distillation optimization to reach high quality. The paper separates geometry from texture: a diffusion module generates four-view, 1024-pixel normal maps, a carving step turns the SMPL template into a clothed mesh with loose garments, and a texture prior plus a diffusion renderer paint the surface from thirty-two surrounding views. The reported result is an explicit, animatable mesh with a texture map in about five minutes, with FID and CLIP scores above existing text-to-3D human methods. If the claim holds, it matters because previous SDS-based systems take hours and often produce over-smoothed or view-inconsistent humans.

What carries the argument

The load-bearing mechanism is the two-stage decoupling. First, a latent diffusion model with cross-view attention produces front, back, left, and right normal maps at 1024 resolution, conditioned on SMPL renderings, text, and camera embeddings. Second, a differentiable rasterizer optimizes vertex displacements from the SMPL mesh against these normals and masks, with continuous remeshing updating topology so loose clothing is not trapped in the template. Third, a shape-guided diffusion model creates a front-view appearance, which is unwrapped into SMPL UV space and completed by a UV inpainting module. Finally, a diffusion renderer with a RenderNet and a ReferenceNet, initialized from a pretrained latent diffusion model, renders dense surrounding views conditioned on coarse vertex colors and normals, from which the final UV texture is optimized with an absolute-difference loss, an SSIM loss, and total-variation regularization.

What would settle it

Render the optimized mesh's normals from the exact four views used as input and compare them pixel-by-pixel with the generated normal maps on prompts describing wide skirts or open coats; if the mean angular error for loose garments substantially exceeds the error for tight clothing while the front-view texture still looks good, the central geometry claim fails for the cases the paper highlights.

Watch

Extended reading notes

Core claim

The paper's central discovery, on its own terms, is that decoupling explicit 2D observation generation from 3D optimization is enough to beat SDS-based pipelines on both geometry and texture. It claims that four high-resolution normal views, when fused by deforming and remeshing an SMPL initialization, recover a consistent clothed surface that supports loose clothing, and that a reference-conditioned diffusion renderer can turn a coarse SMPL UV texture prior into dense, view-consistent RGB observations from which a detailed texture map is optimized. The paper reports lower FID for normal and RGB renderings, higher CLIP scores for text alignment, and better novel-view synthesis (PSNR 28.3, SSIM 0.951, LPIPS 0.043) than the compared methods, with total inference time near five minutes.

Load-bearing premise

The method assumes four generated 1024-pixel normal views, plus SMPL initialization and remeshing, are enough to reconstruct a faithful clothed surface for any prompt, including loose clothing, without a refinement network.

Editorial extensions

If this is right

  • Explicit meshes with texture maps are produced, so downstream graphics pipelines can directly animate, rig, or edit the result.
  • Loose clothing is supported because the mesh topology is updated by remeshing during carving rather than being locked to the SMPL template.
  • Inference drops from hours to about five minutes by avoiding SDS, making interactive or large-scale text-to-human generation feasible.
  • Texture editing can be localized by repainting the SMPL UV or the front view during the texture-prior stage.
  • The same separating-geometry-and-texture design can be extended to image-conditioned generation, since the conditioning signal need not be text.

Reading between the lines

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

  • A natural but untested extension is image-conditioned generation: the pipeline's texture prior and SMPL initialization could be driven by a reference photo instead of a prompt, which the paper flags as possible but does not demonstrate.
  • Because the geometry stage fixes pose and shape before texturing, animation support is likely limited to reposing a static garment rather than cloth that moves with the body.
  • The four-view normal generation is the funnel for all geometry detail; if the method is pushed to higher resolutions or more views, the carving stage would be the place to look for bottlenecks.
  • The diffusion renderer recipe is not inherently human-specific, so testing it on animals or arbitrary objects would reveal how much of the quality gain comes from human priors versus the renderer itself.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. TGRHuman proposes a two-stage text-to-3D human generation pipeline that avoids score-distillation optimization. In the geometry stage, a fine-tuned multi-view latent diffusion model generates four 1024×1024 normal maps (front, back, left, right) conditioned on SMPL renderings and text; these are then fused by optimizing vertex displacements of an initialized SMPL mesh under normal, mask, and smoothness losses, with periodic remeshing, to support loose clothing. In the texture stage, a shape-guided diffusion model generates a front-view appearance, which is unwrapped into SMPL UV space and completed by a UV inpainting model to form a texture prior; this prior is transferred to the carved mesh, and a ReferenceNet-conditioned diffusion renderer synthesizes 32 dense surrounding RGB views, which are finally used to optimize a UV texture map for the human mesh. The paper claims high-quality and consistent geometry and texture, support for loose clothing, and a five-minute inference time, and reports quantitative improvements over Chupa, HumanNorm, TADA, En3D, Joint2Human, SCULPT, and TEXTure on FID and CLIP scores, as well as better PSNR/SSIM/LPIPS than several novel-view synthesis baselines.

Significance. If the reported results hold, TGRHuman is a meaningful contribution: it demonstrates a practical alternative to SDS-based text-to-3D human generation, producing explicit meshes and UV textures in minutes rather than hours, and it introduces a texture-prior-plus-diffusion-renderer recipe for dense-view-consistent appearance synthesis. The decoupling of geometry and texture is principled, and the texture-stage ablations (ReferenceNet, texture prior) show clear drops in PSNR/SSIM/LPIPS when components are removed. The method also targets a practically relevant capability, loose clothing, which topologically fixed SMPL-displacement baselines do not support. However, the main load-bearing claims—especially loose-clothing geometry and superiority over prior art—are supported by relatively weak evidence: the geometry stage is not ablated and is evaluated only qualitatively, the quantitative comparisons lack error bars and a disclosed protocol, and the user study is deferred to a supplementary document. With tightened evaluation and a concrete geometry-stage validation, the contribution could be publishable at a strong venue.

major comments (4)
  1. [Sec. 4.2, Tables 2–4] The quantitative evaluation is statistically thin: only 50 random prompts are used, no error bars or significance tests are reported, the prompt list is not disclosed, and the configuration of each baseline is not specified. FID is computed from 32-view renderings, but the reference image set is not defined (which real images, at what resolution, with what camera distribution?). Please specify the exact protocol, including how baselines are rendered, whether the same prompts and camera setups are used, and how many seeds are averaged. Without this, the claim that TGRHuman 'outperforms existing methods' is not fully supported.
  2. [Sec. 4.3] The statement that removing any module in the geometry and texture-prior stages would break the pipeline, making module-level ablation infeasible, is not convincing. One can ablate the geometry stage by, for example, fixing topology to the initial SMPL mesh, or by using fewer views, or by comparing against a direct SMPL-based baseline with no optimization. The loose-clothing capability is a stated contribution and a central claim, so the absence of any geometry-stage ablation is a substantive deficiency that should be addressed.
  3. [Sec. 4.1, Table 4] The novel-view synthesis comparison to general NVS methods (Wonder3D, SV3D, Stable Zero123) may not be apples-to-apples: TGRHuman is trained on human scans, receives a text prompt and an SMPL pose/shape prior, and generates 32 surround views from a constructed texture prior, whereas the listed baselines perform single-image-conditioned novel-view synthesis on category-agnostic data. Please report the exact input protocol for each baseline (input image, whether text or SMPL is provided, camera conditioning) and discuss the fairness of the comparison. The current table may overstate the advantage of the proposed renderer.
  4. [Sec. 5.1] The limitations section openly admits fine-detail degradation in fingers and hair and failures on out-of-distribution poses, yet the abstract and conclusion claim 'high-quality and consistent 3D human geometry and texture.' The manuscript should bound the headline claim by quantifying the frequency or extent of these failure modes, e.g., by reporting the proportion of generated samples with fused fingers or severe artifacts, rather than only showing selected qualitative examples. This would let readers calibrate the practical scope of the method.
minor comments (6)
  1. [Table 5] The time row reads '2h+1h+30min+5min' without clearly associating each value with a method; please format it as a proper comparison table with one column per method.
  2. [Fig. 4 and Fig. 5 captions] The captions contain typos ('lighxtly', 'layered a red plaid shirtjacket over a gray t-shirt') that should be corrected.
  3. [Eq. (1)] The prediction target is written as 'v_x_t' in the loss, while the definition text uses 'v_n^t' for the v-prediction of normal maps; please unify the notation to avoid confusion.
  4. [Sec. 4.1] The sentence 'we use 10k human scans from the THuman2.1, 2K2K, and Human4DiT datasets' is ambiguous about whether 10k is the total or per-dataset count; please clarify.
  5. [Sec. 4.4] The user study is mentioned but no results are presented in the main paper; please include at least a summary table or the study protocol so the claim of qualitative superiority can be assessed.
  6. [Sec. 3.3] The selection of the initial camera view p0 is not specified; please state how this view is chosen (e.g., fixed front view or randomized).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TGRHuman's geometry and texture generation claims are supported by independently trained diffusion models and external evaluation, not by construction.

full rationale

The paper's central claims are empirical and are not derived from the quantities they aim to predict. The geometry stage trains a multi-view normal diffusion model (Eq. 1) on real and synthetic scans, then optimizes an SMPL-initialized mesh against the resulting normal and mask maps (Eq. 2). This is an optimization with an independent supervision signal, not a fitted parameter later renamed as a prediction. The texture stage constructs a coarse texture prior from a generated front view and UV inpainting, then trains a diffusion renderer on paired data; the final texture is optimized (Eq. 7) to match the renderer's outputs, which is a standard inverse-rendering loop rather than a self-definitional circularity. The evaluation uses external metrics (FID, CLIP score, PSNR/SSIM/LPIPS, user study) against held-out data and external baselines, so the claimed superiority is not forced by construction. The paper's self-citations (e.g., Joint2Human [19], FOF [3]) appear only as background or baselines and are not load-bearing for the central derivation. The acknowledged limitations in Sec. 5.1 and the absence of a geometry-stage ablation in Sec. 4.3 are correctness and robustness concerns, not evidence of circularity. Overall, no circular step could be identified where a prediction reduces to its input by definition or where a fitted value is renamed as a result.

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

The method is a learned pipeline, so it relies on domain assumptions about SMPL initialization, normal-map consistency, UV-prior adequacy, and transferability of diffusion priors. It introduces no new physical entities. Its free parameters are hand-set loss weights and view counts; none is fitted to the reported benchmark in a way that produces a numerical prediction.

free parameters (5)
  • Geometry regularization weight lambda = 1
    In Eq. (2), balances smoothness against normal/mask fidelity for geometry carving; no sensitivity analysis reported.
  • Texture optimization weights lambda_ssim and lambda_tv = 10 and 1
    In Eq. (7), weights for SSIM and total variation losses; chosen by hand without ablated sensitivity in the main text.
  • Number of surround texture views k = 32
    Inference samples 32 yaw views for the diffusion renderer; improves occlusion coverage but no view-count sweep is reported.
  • Number of geometry normal views = 4
    Four views (front, back, left, right) at 1024px chosen to fit VRAM limits; no ablation of 6 or 8 views is provided.
  • Classifier-free guidance drop probability = 10%
    Used when training the multi-view normal model in Sec. 3.1; a standard choice, not analyzed in the paper.
assumptions (5)
  • domain assumption SMPL initialization is sufficient for arbitrary clothed human geometry, including loose clothing.
    Sec. 3.2 starts from SMPL M(theta,beta) and optimizes vertex displacements with remeshing; the search space is anchored to a naked body template.
  • domain assumption Four-view 1024px normal maps generated by the multi-view diffusion model are mutually consistent and detailed enough for geometry carving.
    Sec. 3.1 and Eq. (2) rely on pixel-aligned normal and mask losses; no independent consistency verification is provided beyond qualitative figures.
  • domain assumption Front-view shape-guided diffusion plus UV inpainting produces a faithful global texture prior for occluded regions.
    Sec. 3.3 builds the texture prior from one generated front view and inpainting; errors here propagate to all 32 renderer views.
  • domain assumption Pre-trained Stable Diffusion can be fine-tuned into a view-consistent renderer via RenderNet and ReferenceNet feature injection.
    Sec. 3.4 claims cross-view consistency largely from attention feature summation; this is supported only by the authors' ablation on their own data.
  • standard math Differentiable rasterization and XAtlas UV mapping provide pixel-aligned supervision accurate enough for normal, mask, and texture optimization.
    Eqs. (2) and (7) assume rasterizer gradients align image and mesh coordinates; this is a common and reasonable graphics assumption, not proven in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TGRHuman: Text-Guided Realistic 3D Human Generation via Diffusion Renderer." pith.science (2026). https://pith.science/paper/BO2P7IB6

@misc{pith2026260812175,
  author       = {Pith},
  title        = {Pith review of: TGRHuman: Text-Guided Realistic 3D Human Generation via Diffusion Renderer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BO2P7IB6}},
  note         = {Machine review of arXiv:2608.12175}
}
read the original abstract

Realistic 3D human generation plays a crucial role in many graphics applications. However, current methods still struggle to generate high-quality human geometry and texture while maintaining 3D consistency and inference efficiency. In this work, we address these limitations by introducing TGRHuman, a novel approach for generating realistic 3D humans from text. Our method decouples geometry and texture generation to alleviate the issues commonly encountered in NeRF-based methods. Instead of relying on slow, implicit score-distillation-based optimization, we directly use explicit multi-view observation generation and optimization for efficient 3D synthesis. For geometry generation, we propose a high-resolution generative module for multi-view normals together with a geometry-carving strategy that preserves view consistency and supports loose clothing. For texture generation, we produce spatially consistent RGB observations from densely sampled surrounding views using a carefully designed texture-prior acquisition strategy and a diffusion renderer, enabling detailed human texture synthesis. Experiments show that our method can generate high-quality and consistent 3D human geometry and texture efficiently. TGRHuman outperforms existing text-to-3D human methods in both geometry and texture quality.

Figures

Figures reproduced from arXiv: 2608.12175 by the authors.

Figure 1
Figure 1. Given text descriptions as input, TGRHuman generates diverse and realistic 3D humans with high-quality geometry and texture e [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of TGRHuman. We decouple geometry and texture generation to produce realistic, high-quality humans via explicit 2D observation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Texture optimization results using RGB observations from di [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of human textures. For fair comparison, our method does not perform post-processing during inference, such as replacing hands [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of human geometry. The normal maps are rendered from the generated 3D human meshes. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The detailed architecture of the diffusion renderer. rendered observations, we can easily extract a complete texture map for the human mesh while avoiding the occlusion issues caused by sparse views. Multi-view Rendering Integration for the Texture Map. Given the multi…
Figure 7
Figure 7. Figure 7: Qualitative ablation of the modules in the texture stage. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of novel-view synthesis. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Local-region editing of 3D humans via SMPL UV repainting. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Failure cases involving fingers [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Failure cases involving hair [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Failure cases of generated results under out-of-distribution pose [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison of human texture under complex poses. [PITH_FULL_IMAGE:figures/full_fig_p013_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 53 canonical work pages

  1. [1]

    E. R. Chan, C. Z. Lin, M. A. Chan, K. Nagano, B. Pan, S. De Mello, O. Gallo, L. J. Guibas, J. Tremblay, S. Khamis, et al., Efficient geometry-aware 3d generative adversarial networks, in: CVPR, 2022, pp. 16123–16133

  2. [2]

    J. J. Park, P. Florence, J. Straub, R. Newcombe, S. Love- grove, Deepsdf: Learning continuous signed distance functions for shape representation, in: CVPR, 2019, pp. 165–174

  3. [3]

    Q. Feng, Y . Liu, Y .-K. Lai, J. Yang, K. Li, Fof: Learning fourier occupancy field for monocular real-time human re- construction, 2022

  4. [4]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Om- mer, High-resolution image synthesis with latent diffusion models, in: ICCV , 2022, pp. 10684–10695

  5. [5]

    Y . Men, B. Lei, Y . Yao, M. Cui, Z. Lian, X. Xie, En3d: An enhanced generative model for sculpting 3d humans from 2d synthetic data, in: CVPR, 2024

  6. [6]

    Huang, R

    X. Huang, R. Shao, Q. Zhang, H. Zhang, Y . Feng, Y . Liu, Q. Wang, Humannorm: Learning normal diffusion model for high-quality and realistic 3d human generation, in: CVPR, 2024

  7. [7]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, R. Ng, Nerf: Representing scenes as neural radiance fields for view synthesis, in: ECCV , 2020

  8. [8]

    T. Shen, J. Gao, K. Yin, M.-Y . Liu, S. Fidler, Deep march- ing tetrahedra: a hybrid representation for high-resolution 3d shape synthesis, 2021

Show all 62 references
  1. [9]

    Poole, A

    B. Poole, A. Jain, J. T. Barron, B. Mildenhall, Dream- fusion: Text-to-3d using 2d diffusion, arXiv preprint arXiv:2209.14988 (2022)

  2. [10]

    Huang, H

    Y . Huang, H. Yi, Y . Xiu, T. Liao, J. Tang, D. Cai, J. Thies, TeCH: Text-guided Reconstruction of Lifelike Clothed Humans, 2024

  3. [11]

    Richardson, G

    E. Richardson, G. Metzer, Y . Alaluf, R. Giryes, D. Cohen- Or, Texture: Text-guided texturing of 3d shapes, 2023

  4. [12]

    Kolotouros, T

    N. Kolotouros, T. Alldieck, A. Zanfir, E. G. Bazavan, M. Fieraru, C. Sminchisescu, Dreamhuman: Animatable 3d avatars from text (2023)

  5. [13]

    Cao, Y .-P

    Y . Cao, Y .-P. Cao, K. Han, Y . Shan, K.-Y . K. Wong, Drea- mavatar: Text-and-shape guided 3d human avatar genera- tion via diffusion models, in: ICCV , 2024

  6. [14]

    Y . Shi, P. Wang, J. Ye, M. Long, K. Li, X. Yang, Mv- dream: Multi-view diffusion for 3d generation, arXiv preprint arXiv:2308.16512 (2023)

  7. [15]

    B. Kim, P. Kwon, K. Lee, M. Lee, S. Han, D. Kim, H. Joo, Chupa: Carving 3d clothed humans from skinned shape priors using 2d diffusion probabilistic models, in: ICCV , 2023

  8. [16]

    Sanyal, P

    S. Sanyal, P. Ghosh, J. Yang, M. J. Black, J. Thies, T. Bolkart, SCULPT: Shape-conditioned unpaired learn- ing of pose-dependent clothed and textured human meshes, in: CVPR, 2024

  9. [17]

    T. Wang, B. Zhang, T. Zhang, S. Gu, J. Bao, T. Baltru- saitis, J. Shen, D. Chen, F. Wen, Q. Chen, B. Guo, Rodin: A generative model for sculpting 3d digital avatars using diffusion, in: CVPR, 2023

  10. [18]

    Zhang, J

    X. Zhang, J. Zhang, C. Rohan, H. Xu, G. Song, Y . Yang, J. Feng, Getavatar: Generative textured meshes for ani- matable human avatars, in: ICCV , 2023

  11. [19]

    Zhang, Q

    M. Zhang, Q. Feng, Z. Su, C. Wen, Z. Xue, K. Li, Joint2human: High-quality 3d human generation via com- pact spherical embedding of 3d joints, in: CVPR, 2024

  12. [20]

    Loper, N

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, M. J. Black, Smpl: a skinned multi-person linear model, ACM TOG (2015)

  13. [21]

    M. Shi, W. Feng, L. Gao, D. Zhu, Generating diverse clothed 3d human animations via a generative model, Computational Visual Media 10 (2) (2024) 261–277

  14. [22]

    Wang, H.-Y

    C. Wang, H.-Y . Peng, Y .-T. Liu, J. Gu, S.-M. Hu, Diffu- sion models for 3d generation: A survey, Computational Visual Media 11 (1) (2025) 1–28

  15. [23]

    Z. Wang, Y . Kanamori, Y . Endo, Eg-humannerf: Efficient generalizable human nerf utilizing human prior for sparse view, Computational Visual Media 12 (2) (2026) 355– 379

  16. [24]

    T. Liao, H. Yi, Y . Xiu, J. Tang, Y . Huang, J. Thies, M. J. Black, TADA! Text to Animatable Digital Avatars, 2024

  17. [25]

    Zhang, B

    H. Zhang, B. Chen, H. Yang, L. Qu, X. Wang, L. Chen, C. Long, F. Zhu, K. Du, M. Zheng, Avatarverse: High- quality stable 3d avatar creation from text and pose, in: AAAI, 2024

  18. [26]

    F. Hong, M. Zhang, L. Pan, Z. Cai, L. Yang, Z. Liu, Avatarclip: Zero-shot text-driven generation and anima- tion of 3d avatars, ACM TOG (2022)

  19. [27]

    Jiang, C

    R. Jiang, C. Wang, J. Zhang, M. Chai, M. He, D. Chen, J. Liao, Avatarcraft: Transforming text into neural human avatars with parameterized shape and pose control, arXiv preprint arXiv:2303.17606 (2023)

  20. [28]

    Huang, J

    Y . Huang, J. Wang, A. Zeng, H. Cao, X. Qi, Y . Shi, Z.-J. Zha, L. Zhang, DreamWaltz: Make a Scene with Complex 3D Animatable Avatars, 2023. 14

  21. [29]

    Y . Zeng, Y . Lu, X. Ji, Y . Yao, H. Zhu, X. Cao, Avatar- booth: High-quality and customizable 3d human avatar generation, 2023

  22. [30]

    Saito, Z

    S. Saito, Z. Huang, R. Natsume, S. Morishima, A. Kanazawa, H. Li, Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization, in: ICCV , 2019

  23. [31]

    X. Zhan, J. Yang, Y . Li, J. Guo, Y . Guo, W. Wang, Seman- tic human mesh reconstruction with textures, in: CVPR, 2024

  24. [32]

    AlBahar, S

    B. AlBahar, S. Saito, H.-Y . Tseng, C. Kim, J. Kopf, J.-B. Huang, Single-image 3d human digitization with shape- guided diffusion, 2023, pp. 1–11

  25. [33]

    I. Ho, J. Song, O. Hilliges, et al., Sith: Single-view tex- tured human reconstruction with image-conditioned dif- fusion, in: CVPR, 2024

  26. [34]

    D. Z. Chen, Y . Siddiqui, H.-Y . Lee, S. Tulyakov, M. Nießner, Text2tex: Text-driven texture synthesis via diffusion models, in: ICCV , 2023

  27. [35]

    T. Cao, K. Kreis, S. Fidler, N. Sharp, K. Yin, Texfusion: Synthesizing 3d textures with text-guided image diffusion models, in: ICCV , 2023

  28. [36]

    D. Huo, Z. Guo, X. Zuo, Z. Shi, J. Lu, P. Dai, S. Xu, L. Cheng, Y .-H. Yang, Texgen: Text-guided 3d texture generation with multi-view sampling and resampling, in: ECCV , 2024

  29. [37]

    S. R. K. Perla, Y . Wang, A. Mahdavi-Amiri, H. Zhang, Easi-tex: Edge-aware mesh texturing from single image, ACM TOG 43 (4) (2024). doi:10.1145/3658222. URLhttps://github.com/sairajk/easi-tex

  30. [38]

    X. Zeng, X. Chen, Z. Qi, W. Liu, Z. Zhao, Z. Wang, B. Fu, Y . Liu, G. Yu, Paint3d: Paint anything 3d with lighting- less texture diffusion models, in: CVPR, 2024

  31. [39]

    R. Liu, R. Wu, B. Van Hoorick, P. Tokmakov, S. Zakharov, C. V ondrick, Zero-1-to-3: Zero-shot one image to 3d ob- ject, in: ICCV , 2023

  32. [40]

    Long, Y .-C

    X. Long, Y .-C. Guo, C. Lin, Y . Liu, Z. Dou, L. Liu, Y . Ma, S.-H. Zhang, M. Habermann, C. Theobalt, et al., Won- der3d: Single image to 3d using cross-domain diffusion, in: CVPR, 2024, pp. 9970–9980

  33. [41]

    V oleti, C.-H

    V . V oleti, C.-H. Yao, M. Boss, A. Letts, D. Pankratz, D. Tochilkin, C. Laforte, R. Rombach, V . Jampani, Sv3d: Novel multi-view synthesis and 3d generation from a sin- gle image using latent video diffusion, in: ECCV , 2024, pp. 439–457

  34. [42]

    Jiang, H

    S. Jiang, H. Luo, H. Jiang, Z. Wang, J. Yu, L. Xu, Mvhuman: Tailoring 2d diffusion with multi-view sam- pling for realistic 3d human generation, arXiv preprint arXiv:2312.10120 (2023)

  35. [43]

    X. He, X. Li, D. Kang, J. Ye, C. Zhang, L. Chen, X. Gao, H. Zhang, Z. Wu, H. Zhuang, Magicman: Generative novel view synthesis of humans with 3d-aware diffusion and iterative refinement, arXiv preprint arXiv:2408.14211 (2024)

  36. [44]

    Y . Xue, X. Xie, R. Marin, G. Pons-Moll, Human 3diffu- sion: Realistic avatar creation via explicit 3d consistent diffusion models, Arxiv (2024)

  37. [45]

    P. Li, W. Zheng, Y . Liu, T. Yu, Y . Li, X. Qi, M. Li, X. Chi, S. Xia, W. Xue, et al., Pshuman: Photorealistic single-view human reconstruction using cross-scale diffu- sion, arXiv preprint arXiv:2409.10141 (2024)

  38. [46]

    Huang, Y

    Z. Huang, Y . Guo, H. Wang, R. Yi, L. Ma, Y .-P. Cao, L. Sheng, Mv-adapter: Multi-view consistent image generation made easy, arXiv preprint arXiv:2412.03632 (2024)

  39. [47]

    Y . Xu, Z. Yang, Y . Yang, Seeavatar: Photorealistic text-to- 3d avatar generation with constrained geometry and ap- pearance, arXiv preprint arXiv:2312.08889 (2023)

  40. [48]

    L. Qiu, X. Gu, P. Li, Q. Zuo, W. Shen, J. Zhang, K. Qiu, W. Yuan, G. Chen, Z. Dong, L. Bo, Lhm: Large animat- able human reconstruction model from a single image in seconds, in: ICCV , 2025

  41. [49]

    Xiang, X

    J. Xiang, X. Chen, S. Xu, R. Wang, Z. Lv, Y . Deng, H. Zhu, Y . Dong, H. Zhao, N. J. Yuan, J. Yang, Native and compact structured latents for 3d generation, Tech re- port (2025)

  42. [50]

    Yang, B.-T

    J. Yang, B.-T. Zhang, F.-L. Liu, H. Fu, Y .-K. Lai, L. Gao, Humanlift: Single-image 3d human reconstruction with 3d-aware diffusion priors and facial enhancement, in: ACM SIGGRAPH Asia, 2025

  43. [51]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language supervision, PMLR, 2021, pp. 8748–8763

  44. [52]

    Salimans, J

    T. Salimans, J. Ho, Progressive distillation for fast sam- pling of diffusion models, ICLR (2022)

  45. [53]

    Laine, J

    S. Laine, J. Hellsten, T. Karras, Y . Seol, J. Lehtinen, T. Aila, Modular primitives for high-performance differ- entiable rendering, ACM TOG 39 (6) (2020)

  46. [54]

    Palfinger, Continuous remeshing for inverse rendering, Computer Animation and Virtual Worlds 33 (5) (2022) e2101

    W. Palfinger, Continuous remeshing for inverse rendering, Computer Animation and Virtual Worlds 33 (5) (2022) e2101

  47. [55]

    Y . Xiu, J. Yang, X. Cao, D. Tzionas, M. J. Black, ECON: Explicit Clothed humans Optimized via Normal integra- tion, in: CVPR, 2023

  48. [56]

    Young, Mesh parameterization/uv unwrapping library (2018)

    J. Young, Mesh parameterization/uv unwrapping library (2018). URLhttps://github.com/jpcy/xatlas 15

  49. [57]

    Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli, Im- age quality assessment: from error visibility to structural similarity, IEEE TIP 13 (4) (2004) 600–612

  50. [58]

    L. I. Rudin, S. Osher, E. Fatemi, Nonlinear total varia- tion based noise removal algorithms, Physica D: nonlinear phenomena 60 (1-4) (1992) 259–268

  51. [59]

    T. Yu, Z. Zheng, K. Guo, P. Liu, Q. Dai, Y . Liu, Func- tion4d: Real-time human volumetric capture from very sparse consumer rgbd sensors, in: CVPR, 2021

  52. [60]

    Han, M.-G

    S.-H. Han, M.-G. Park, J. H. Yoon, J.-M. Kang, Y .-J. Park, H.-G. Jeon, High-fidelity 3d human digitization from sin- gle 2k resolution images, in: CVPR, 2023

  53. [61]

    R. Shao, Y . Pang, Z. Zheng, J. Sun, Y . Liu, Human4dit: 360-degree human video generation with 4d diffusion transformer, ACM TOG 43 (6) (2024)

  54. [62]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, O. Wang, The unreasonable effectiveness of deep features as a per- ceptual metric, in: CVPR, 2018, pp. 586–595. 16

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.