Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Text-driven 3D Human Generation via Contrastive Preference Optimization

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Preference-model gradients align 3D humans to long text

desk verdict A plausible but under-specified engineering combination of preference models and negation prompts for text-to-3D humans; the central equations don't type-check as written, so the method is not reproducible from the paper alone. read the letter →

arxiv 2502.08977 v3 pith:6NF356T3 submitted 2025-02-13 cs.CV

Pengfei Zhou , Xukun Shen , Yong Hu This is my paper · ORCID
classification cs.CV
keywords text-to-3Dgeneration3DhumanScoreDistillationSamplingpreferenceoptimizationrewardhackingnegationpromptsGaussianSplattinghuman-levelmodels
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 argues that text-driven 3D human generation fails on long, complex prompts because Score Distillation Sampling entangles attributes such as colors, garments, and accessories, and proposes to fix this by adding contrastive preference gradients from two complementary 2D human-preference models, ImageReward and PickScore, together with static and LLM-written negation prompts. It reports state-of-the-art results on a suite of 100 complex prompts, with higher CLIP, DenseScore, and HPSv2 scores and better user ratings of texture and geometry. The method is built around a preference optimization module with least-common-multiple adaptive weighting and a negation preference module that suppresses reward hacking by penalizing generated objects and attribute shifts not present in the prompt.

What carries the argument

The load-bearing mechanism is the contrastive preference function built from two complementary preference models, ImageReward and PickScore, whose gradients are combined by a least-common-multiple adaptive weighting scheme: $\lambda_i = \sum_{i=1}^2 \text{LCM}(s_{\text{IR}}, s_{\text{PS}})/s_i$, so the model that scores lower at a given step receives more influence. The positive preference term $C_p^+ = \frac{1}{2}\sum_i \lambda_i \nabla s_i$ is added to the SDS loss, and the negative term $C_p^- = \frac{1}{2}\sum_i \nabla s_i^\text{neg}$ is evaluated on static and LLM-generated negation prompts to suppress irrelevant objects, reorder salient attributes, and prevent over-optimization. The framework optimizes a 3D Gaussian Splatting model initialized densely on the SMPL-X mesh, using rendered views as inputs to the preference models.

What would settle it

A direct test would be to evaluate the method with the preference models queried on renderings from extreme camera elevations or top-down views (outside the paper's elevation range of -30 to 30 degrees). If the preference scores and gradients degrade sharply on these views and the final 3D model quality drops, it would confirm the load-bearing assumption is viewpoint-limited. A second falsifier is to replace the LLM-generated dynamic negation prompts with random adjective-attribute reorderings: if the reported DenseScore gains persist under random permutations, the claimed benefit of the negation module's semantic analysis would be called into question.

Watch

Extended reading notes

Core claim

The central claim is that fine-grained semantic alignment in text-to-3D human generation, especially for long and complex textual inputs, improves substantially when the SDS optimization is guided by a contrastive preference objective that combines positive gradients from multiple 2D preference models and negative gradients from negation prompts. The authors show that ImageReward and PickScore have complementary biases—ImageReward captures colors and textures well, PickScore captures object categories such as hats and glasses—so fusing them with an adaptive least-common-multiple weighting produces more faithful attribute mapping. They also introduce a negation preference module that uses static negative phrases plus LLM-generated dynamic negation prompts (reordered adjective-attribute pairs and plausible irrelevant objects) to prevent over-optimization artifacts, a form of reward hacking. The complete objective is $L_{\text{all}} = C_p^+ + C_p^- + L_{\text{sds}}$, where $C_p^+$ is the weighted sum of preference gradients and $C_p^-$ is the average of negation gradients; optimization proceeds on a 3D Gaussian Splatting representation initialized on SMPL-X.

Load-bearing premise

The method assumes that ImageReward and PickScore, which were trained on 2D natural images, produce stable and meaningful preference gradients when applied to rendered 3D human views at arbitrary viewpoints; if those gradients are noisy or misleading on out-of-distribution renders, the preference module could distort geometry or texture instead of aligning it.

Editorial extensions

If this is right

  • If the method is correct, text-to-3D human generation can move beyond single-model alignment: fusing multiple 2D preference models with adaptive weighting reduces the semantic bias of any one reward model, improving both attribute-level and category-level fidelity.
  • The negation preference module gives a practical recipe for suppressing reward hacking without retraining the preference models, by injecting LLM-generated negative prompts that are specific to each textual description.
  • The reported speed (0.8 hours on a single RTX 4090) suggests the contrastive preference gradients add little overhead to the base SDS pipeline, making the approach viable for interactive avatar creation.
  • The quantitative gains on DenseScore and HPSv2 indicate the method is not merely fitting a single metric but improves long-text alignment and texture aesthetics simultaneously.

Reading between the lines

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

  • A testable extension is to apply the same contrastive preference formulation to text-to-3D object generation (non-human) where the semantic entanglement problem also appears; the method's reliance on 2D preference models should transfer whenever rendered views resemble natural images.
  • The LCM weighting strategy is heuristic; a more principled alternative would be to learn the fusion weights from validation data or to use gradient variance to set per-step weights, which could be evaluated against the reported ablations.
  • The dynamic negation prompts generated by an LLM are only as good as the LLM's commonsense judgment of 'irrelevant elements'; a systematic study of how LLM errors (e.g., suggesting a plausible but absent object) affect the final 3D model would clarify the method's robustness.
  • Because ImageReward and PickScore are trained on 2D natural images, their gradients on unusual rendered viewpoints (e.g., extreme elevations or close-ups) may be unreliable; the paper's camera range is constrained, so a stress test with a wider viewpoint distribution would probe this boundary.
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

3 major / 5 minor

Summary. The paper proposes a text-driven 3D human generation framework that combines 3D Gaussian Splatting with SMPL-X and Score Distillation Sampling, augmented by a contrastive preference optimization module. The module uses two preference models, ImageReward and PickScore, with an LCM-based adaptive weighting strategy, together with static and LLM-generated dynamic negation prompts, to improve semantic alignment and suppress reward hacking. The authors report qualitative comparisons and quantitative evaluations against DreamWaltz, TADA, X-Oscar, HumanGaussian, and HumanNorm, and claim state-of-the-art performance on CLIP, DenseScore, HPSv2, and a user study, with ablations for each module.

Significance. The underlying idea is relevant and timely: using contrastive positive/negative preference signals to refine SDS for long and complex text-to-3D human generation addresses a real failure mode of existing SDS-based methods. The dynamic LLM-driven negation prompting mechanism is a plausible way to mitigate over-optimization and attribute entanglement. However, the paper does not ship code or exact checkpoints, the quantitative results are single-run numbers without error bars or significance tests, and the central optimization objective is not mathematically specified in a reproducible way. If the equations are corrected and the evaluation is made statistically rigorous and less dependent on preference-aligned metrics, the framework could be a useful contribution to text-to-3D human generation.

major comments (3)
  1. [Sec. III-C/D, Eqs. (7)-(12)] The preference optimization objective is internally inconsistent. Eq. (7) defines λ_i via LCM(s_IR, s_PS), but LCM is undefined for real-valued preference scores, since ImageReward scores and PickScore log-probabilities are not integers; the printed formula also writes λ_i on the left and a sum over i on the right, which is not a well-formed definition. Eqs. (8)-(11) define C_p^+, C_p^-, and C_p^all as gradient vectors, while L_sds in Eq. (3) is a scalar loss. Eq. (12) then sets L_all = C_p_all + L_sds without stating whether C_p_all is a loss to be minimized, an update direction to be added to ∇L_sds, or a gradient with an implicit sign and scale. Because the sign of the preference term determines whether the optimizer increases or decreases ImageReward/PickScore scores, the central alignment mechanism is not reproducible from the paper as written.
  2. [Sec. IV-C, Table I] The quantitative evaluation is partially circular. CLIP, DenseScore, and HPSv2 are themselves preference-aligned models in the same family as the ImageReward and PickScore models used as optimization signals, so score improvements can reflect overfitting to preference models rather than improved human alignment. The user study is the only independent check, but it is under-reported: no protocol details are given (viewpoint selection, side-by-side or sequential presentation, randomization, number of ratings per participant, instructions, inter-rater reliability), and no confidence intervals, standard deviations, or statistical tests are reported. The claim of state-of-the-art performance therefore lacks statistical support.
  3. [Sec. IV-D, Table II] The ablation table does not establish the contribution of the proposed components. There are no error bars or multiple runs, and the row labels are ambiguous: 'w/ NPOM(DSNP)' scores 30.78 on CLIP while 'Ours' scores 32.09, but the paper never states which components are active in 'w/ NPOM(DSNP)' versus 'Ours'. Since the central claim is that both POM and NPOM are needed, the table should show base, +POM, +NPOM, and +POM+NPOM explicitly, with variances. Additionally, the DenseScore for 'w/ POM (MW)' (28.41) is lower than that for 'w/o POM' (28.84), which suggests the adaptive weighting result is not robust as presented.
minor comments (5)
  1. [Sec. III-A, Eq. (4)] Equation (4) is missing the covariance matrix in the exponent: it reads exp(-1/2 (p - μ_i)) rather than the full anisotropic Gaussian form used in 3DGS, which makes the equation mathematically incomplete.
  2. [Sec. IV-A / Sec. III-D] The static negation prompt is given only as 'blurry, oversaturated, noisy, ..., jpeg artifacts'; the ellipsis hides the actual prompt set, and the full list should be provided for reproducibility.
  3. [Sec. II-A] The related-work paragraph mentions 'ISD refines the SDS framework...' without a citation; if ISD refers to a specific published method, it should be cited.
  4. [Sec. IV-C, Table I] Table I mixes user-study scores (on a 1-10 scale) with automated metric scores (CLIP, DenseScore, HPSv2) in one table without clarifying that they are on different scales, and the header 'DenseScore Score' contains a typo.
  5. [Sec. V, Limitations] The limitations paragraph is too terse: it only mentions texture resolution and Gaussian artifacts, and does not address the preference-model domain gap, the ambiguity in the optimization equations, or the reliance on preference-aligned evaluation metrics.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the preference-optimization training signal and the evaluation metrics are distinct, and no central claim reduces by construction to its own inputs.

full rationale

The paper's core contribution is a training-time objective (Sec. III-C/D) that adds gradients from ImageReward and PickScore to the SDS gradient. Neither of these preference models appears in the evaluation metrics (CLIP, DenseScore, HPSv2), and the metrics are not used in the loss, so improvements are not forced by construction. The LCM-based weighting in Eqs. 7-8 is a heuristic computed from the same preference scores being optimized, but it is part of the optimization loop rather than a fitted parameter later renamed as a prediction. The LLM prompt template is an input, not a result derived from the outputs. The paper does not rely on self-citations: the cited preference models and baselines are external prior work by other author groups. Some technical concerns exist, such as the vector/scalar inconsistency in Eq. 12, the undefined LCM for real-valued scores in Eq. 7, and the underspecified update rule, but these are reproducibility/correctness issues rather than circularity. The claimed state-of-the-art results are therefore self-contained empirical claims, not definitional equivalences.

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

The method rests on standard domain assumptions about SDS and 3DGS, plus several untested assumptions about preference model transferability and the LCM weighting scheme. No new entities are introduced.

free parameters (4)
  • Inverse LCM preference weights lambda_i = undefined (not specified)
    Eq 7 attempts to weight preference gradients by LCM(s_IR, s_PS)/s_i, but LCM is defined for integers, not real-valued scores; the actual weights in the implementation are unstated.
  • Classifier-free guidance scale = 7.5
    Fixed at 7.5 across all experiments; chosen by hand from prior SDS practice, not tuned or fitted.
  • SDS timestep sampling range = U(0.02, 0.50)
    Time steps uniformly sampled in this range; a training hyperparameter that affects texture and geometry balance.
  • Static negative prompt set = hand-crafted phrase list
    Set includes phrases like blurry, oversaturated, noisy, jpeg artifacts; the exact contents and their weights are not fully specified.
assumptions (5)
  • domain assumption The pretrained diffusion model epsilon_phi provides valid score estimates for SDS optimization of 3D representations.
    Used in Eq 3; the entire SDS framework assumes the 2D diffusion prior transfers to 3D rendered views.
  • domain assumption ImageReward and PickScore produce meaningful, differentiable preference scores on rendered 3D human images.
    Sec III-C applies these 2D-trained models to rendered images X from 3DGS; the gradient of the scalar score is assumed to guide 3D parameters correctly.
  • ad hoc to paper The inverse LCM weighting in Eq 7 is a valid adaptive weighting scheme.
    No derivation or empirical validation is given; LCM of real scores is not defined in standard mathematics.
  • domain assumption LLM-generated dynamic negation prompts identify truly irrelevant attributes and recombined adjective-attribute pairs.
    Sec III-B and III-D rely on the LLM to estimate saliency and irrelevant objects; errors in this estimation would propagate to the negation loss.
  • domain assumption SMPL-X provides a sufficiently accurate human body prior for initialization and pose control.
    Used throughout the pipeline; if the template does not match the described human, generation will be constrained.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text-driven 3D Human Generation via Contrastive Preference Optimization." pith.science (2026). https://pith.science/paper/6NF356T3

@misc{pith2026250208977,
  author       = {Pith},
  title        = {Pith review of: Text-driven 3D Human Generation via Contrastive Preference Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NF356T3}},
  note         = {Machine review of arXiv:2502.08977}
}
read the original abstract

Recent advances in Score Distillation Sampling (SDS) have improved 3D human generation from textual descriptions. However, existing methods still face challenges in accurately aligning 3D models with long and complex textual inputs. To address this challenge, we propose a novel framework that introduces contrastive preferences, where human-level preference models, guided by both positive and negative prompts, assist SDS for improved alignment. Specifically, we design a preference optimization module that integrates multiple models to comprehensively capture the full range of textual features. Furthermore, we introduce a negation preference module to mitigate over-optimization of irrelevant details by leveraging static-dynamic negation prompts, effectively preventing ``reward hacking". Extensive experiments demonstrate that our method achieves state-of-the-art results, significantly enhancing texture realism and visual alignment with textual descriptions, particularly for long and complex inputs.

Figures

Figures reproduced from arXiv: 2502.08977 by the authors.

Figure 1
Figure 1. Comparison of Long-Text 3D Human Generation Methods [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Proposed Framework. We first leverage [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. PickScore [10] demonstrates a stronger capability i [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons with the SOTA Text-to-3D hu [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparisons with the SOTA Text-to-3D me [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: More examples of 3D huamn produced by our approach. Th [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Ablation Study on Negation Prompts and Adaptive Weig [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 9
Figure 9. Figure 9: Applications of our method. capability by modifying Curry’s appearance. This flexibility ensures high visual fidelity and personalization, benefiting applications in content creation, virtual avatars, and digital entertainment. Text-based Style Creation. Our method ena…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 39 canonical work pages

  1. [1]

    Dream Fusion: Text- to-3D using 2D Diffusion,

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall, “Dream Fusion: Text- to-3D using 2D Diffusion,” in Proceedings of the Eleventh In ternational Conference on Learning Representations, 2023

    work page 2023

  2. [2]

    TADA! Text to Animatable Digital Avatars,

    T. Liao, H. Yi, Y . Xiu, J. Tang, Y . Huang, J. Thies, and M. J. Black, “TADA! Text to Animatable Digital Avatars,” in Inter national Conference on 3D Vision (3DV), 2024. 1, 3, 8, 9, 11

    work page 2024

  3. [3]

    Expressive body capture: 3d hands , face, and body from a single image,

    G. Pavlakos, V . Choutas, N. Ghorbani, T. Bolkart, A. A. Os man, D. Tzionas, and M. J. Black, “Expressive body capture: 3d hands , face, and body from a single image,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 201 9, pp. 10 975–10 985. 1, 2, 4, 5, 6, 7, 12, 15

    work page

  4. [4]

    DreamWaltz: Make a Scene with Complex 3D Animatab le Avatars,

    Y . Huang, J. Wang, A. Zeng, H. Cao, X. Qi, Y . Shi, Z.-J. Zha, and L. Zhang, “DreamWaltz: Make a Scene with Complex 3D Animatab le Avatars,” in Advances in Neural Information Processing Sys tems, 2023. 1, 2, 3, 5, 6, 7, 8, 9, 15

    work page 2023

  5. [5]

    Adding conditional co ntrol to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional co ntrol to text-to-image diffusion models,” in Proceedings of the I EEE/CVF International Conference on Computer Vision, 2023, pp. 383 6–3847

    work page 2023

  6. [6]

    HumanGaussian: Text-Driven 3D Human Generation with Gaus sian Splatting,

    X. Liu, X. Zhan, J. Tang, Y . Shan, G. Zeng, D. Lin, X. Liu, an d Z. Liu, “HumanGaussian: Text-Driven 3D Human Generation with Gaus sian Splatting,” in Proceedings of the IEEE/CVF Conference on Co mputer Vision and Pattern Recognition, 2024, pp. 6646–6657. 1, 3, 5 , 8, 9, 11

    work page 2024

  7. [7]

    3 D Gaussian Splatting for Real-Time Radiance Field Rendering,

    B. Kerbl, G. Kopanas, T. Leimkuhler, and G. Drettakis, “3 D Gaussian Splatting for Real-Time Radiance Field Rendering,” ACM Tra nsactions on Graphics, vol. 42, no. 4, July 2023. 1, 2, 3, 5, 6, 7, 8, 11, 14 , 15

    work page 2023

  8. [8]

    HumanNorm: Learning Normal Diffusion Model for High - quality and Realistic 3D Human Generation,

    X. Huang, R. Shao, Q. Zhang, H. Zhang, Y . Feng, Y . Liu, and Q . Wang, “HumanNorm: Learning Normal Diffusion Model for High - quality and Realistic 3D Human Generation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2024. 3

    work page 2024

Show all 42 references
  1. [9]

    Image reward: Learning and evaluating human preferences for text-to-ima ge genera- tion,

    J. Xu, X. Liu, Y . Wu, Y . Tong, Q. Li, M. Ding, J. Tang, “Image reward: Learning and evaluating human preferences for text-to-ima ge genera- tion,” In Advances in Neural Information Processing System s, 2023, pp.15903-15935

    2023

  2. [10]

    Pick-a-pic: An open dataset of user preferences for text-t o-image generation,

    Y . Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, O. Levy, “Pick-a-pic: An open dataset of user preferences for text-t o-image generation,” In Advances in Neural Information Processing Systems, 36, pp.36652-36663

  3. [11]

    X. Wu, Y . Hao, K. Sun, Y . Chen, F. Zhu, R. Zhao, H. Li, “Hu- man Preference Score v2: A Solid Benchmark for Evaluating Hu - man Preferences of Text-to-Image Synthesis, 2023, arXiv pr eprint arXiv:2306.09341

    2023 arXiv

  4. [12]

    Magic3D: High-Resolut ion Text- to-3D Content Creation,

    C.-H. Lin, J. Gao, L. Tang, T. Takikawa, X. Zeng, X. Huang , K. Kreis, S. Fidler, M.-Y . Liu, and T.-Y . Lin, “Magic3D: High-Resolut ion Text- to-3D Content Creation,” in Proceedings of the IEEE/CVF Con ference On Computer Vision and Pattern Recognition, 2023, pp. 300-3 09

    2023

  5. [13]

    Proli ficdreamer: High-fidelity and diverse text-to-3d generation with varia tional score distillation,

    Z. Wang, C. Lu, Y . Wang, F. Bao, C. Li, H. Su, J. Zhu, “Proli ficdreamer: High-fidelity and diverse text-to-3d generation with varia tional score distillation,” in Advances in Neural Information Processi ng Systems, 2023, 8406-8441

    2023

  6. [14]

    R. Chen, Y . Chen, N. Jiao and K. Jia, ”Fantasia3D: Disent angling Geometry and Appearance for High-quality Text-to-3D Conte nt Cre- ation,” in Proceedings of the IEEE/CVF International Confe rence on Computer Vision, 2023, pp. 22189-22199

    2023

  7. [15]

    DreamGauss ian: Generative Gaussian Splatting for Efficient 3D Content Crea tion,

    J. Tang, J. Ren, H. Zhou, Z. Liu, and G. Zeng, “DreamGauss ian: Generative Gaussian Splatting for Efficient 3D Content Crea tion,” in International Conference on Learning Representations, 20 24. 3

  8. [16]

    Mvdream: Multi-view diffusion for 3d generation,

    Y . Shi, P . Wang, J. Y e, M. Long, K. Li, X. Y ang, “Mvdream: Multi-view diffusion for 3d generation,” 2023, arXiv preprint arXiv:2 308.16512

    2023

  9. [17]

    MVGaussian : High- Fidelity text-to-3D Content Generation with Multi-View Gu idance and Surface Densification,

    P . Pham, A. Mathur, N. A, O. Sharma, A. Bera, “MVGaussian : High- Fidelity text-to-3D Content Generation with Multi-View Gu idance and Surface Densification,” 2024, arXiv preprint arXiv:2409.0 6620

    2024

  10. [18]

    DreamCouple: Exploring High Qual ity Text- to-3D Generation Via Rectified Flow,

    H. Li, X. Chu, D. Shi, “DreamCouple: Exploring High Qual ity Text- to-3D Generation Via Rectified Flow,” 2024, arXiv: 2408.050 08

    2024

  11. [19]

    DreamMapping: High-Fidelity Text-to-3D Generation via V ariational Distribution Mapping,

    Z. Cai, D. Wang, Y . Liang, Z. Shao, Y . Chen, X. Zhan, Z. Wan g, “DreamMapping: High-Fidelity Text-to-3D Generation via V ariational Distribution Mapping,” 2024, arXiv preprint arXiv:2409.0 5099

    2024

  12. [20]

    Deep Marching Tetrahedra: A Hybrid Representation for High-resolution 3D Shape Synth esis,

    T. Shen, J. Gao, K. Yin, M. Liu, S. Fidler, “Deep Marching Tetrahedra: A Hybrid Representation for High-resolution 3D Shape Synth esis,” in Advances in Neural Information Processing Systems, 34:608 7–6101,

  13. [21]

    Ava tarCLIP: Zero-Shot Text-Driven Generation and Animation of 3D Avata rs,

    F. Hong, M. Zhang, L. Pan, Z. Cai, L. Y ang, and Z. Liu, “Ava tarCLIP: Zero-Shot Text-Driven Generation and Animation of 3D Avata rs,” ACM Transactions on Graphics (TOG), vol. 41, no. 4, pp. 1–19, 2022. 1, 3, 4

    2022

  14. [22]

    CLIPMesh: Generating textured meshes from text using pretrained imag e-text models,

    N. Mohammad Khalid, T. Xie, E. Belilovsky, and T. Popa, “ CLIPMesh: Generating textured meshes from text using pretrained imag e-text models,” in SIGGRAPH Asia 2022 Conference Papers, 2022, pp. 1–8. 3

    2022

  15. [23]

    Sanghi, C

    A. Sanghi, C. Hang, GL Joseph, Y . Wang, et al., ”CLIP-For ge: Towards Zero-Shot Text-to-Shape Generation,” in Proceedings of th e IEEE/CVF Conference on Computer Vision and Pattern Recognition, 202 2, pp. 18582-18592

  16. [24]

    Building efficient, accurate c haracterskins from examples,

    A. Mohr and M. Gleicher, “Building efficient, accurate c haracterskins from examples,” ACM Transactions on Graphics (TOG), vol. 22 , no. 3, pp. 562–568, 2003. 4, 6

    2003

  17. [25]

    SMPL: a skinned multi-person linear mode,

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. B lack, “SMPL: a skinned multi-person linear mode,” ACM transactio ns on graphics (TOG), vol. 34, no. 6, pp. 1–16, 2015. 1, 2, 3, 4, 6

    2015

  18. [26]

    Photoreali stic text-to-image diffusion models with deep language underst anding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, EL. Dento n, K. Ghasemipour, R. G. Lopes, B. Ayan, T. Salimans, “Photoreali stic text-to-image diffusion models with deep language underst anding,” in Advances in Neural Information Processing Systems, 35:364 79–36494,

  19. [27]

    NeRF: Representing Scenes as NeuralRadiance Fields fo r View Synthesis,

    B. Mildenhall, P . Srinivasan, M. Tancik, J. Barron, R. R amamoorthi. R. Ng, “NeRF: Representing Scenes as NeuralRadiance Fields fo r View Synthesis,” Communications of the ACM, V ol. 65, no. 1, pp. 99 -106, 2021, 1, 2, 4, 5, 15

    2021

  20. [28]

    NeuS: Learning Neural Implicit Surfaces by V olume Rendering for Multi-view Reconstruction,

    P . Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wan g, “NeuS: Learning Neural Implicit Surfaces by V olume Rendering for Multi-view Reconstruction,” Advances in Neural Information Processi ng Systems, vol. 34, pp. 27 171–27 183, 2021. 1, 2

    2021

  21. [29]

    DreamAvatar: Text-and-Shape Guided 3D Human Avatar Generation via Diffu sion Models,

    Y . Cao, Y .-P . Cao, K. Han, Y . Shan, and K.-Y . K. Wong, “DreamAvatar: Text-and-Shape Guided 3D Human Avatar Generation via Diffu sion Models,” in Proceedings of the IEEE/CVF Conference on Compu ter Vision and Pattern Recognition, 2024, pp. 958-968

    2024

  22. [30]

    AvatarCraft: Transforming Text into Neural Human Av atars with Parameterized Shape and Pose Control,

    R. Jiang, C. Wang, J. Zhang, M. Chai, M. He, D. Chen, and J. Liao, “AvatarCraft: Transforming Text into Neural Human Av atars with Parameterized Shape and Pose Control,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 202 3, pp. 14325-14336

  23. [31]

    DreamWaltz-G: Expre ssive 3D Gaussian Avatars from Skeleton-Guided 2D Diffusion,

    Y . Huang, J. Wang, A. Zeng, et al., “DreamWaltz-G: Expre ssive 3D Gaussian Avatars from Skeleton-Guided 2D Diffusion,” 2024 , arXiv preprint arXiv:2409.17145

    2024 arXiv

  24. [32]

    DreamHuman: Animatable 3D Avatars from T ext,

    N. Kolotouros, T. Alldieck, A. Zanfir, E. Bazavan, M. Fie raru, and C. Sminchisescu, “DreamHuman: Animatable 3D Avatars from T ext,” Advances in Neural Information Processing Systems, vol. 36 , 2024. 1, 3, 8, 9

    2024

  25. [33]

    imGHUM: Impli cit Genera- tive Models of 3D Human Shape and Articulated Pose,

    T. Alldieck, H. Xu, and C. Sminchisescu, “imGHUM: Impli cit Genera- tive Models of 3D Human Shape and Articulated Pose,” in Proce edings of the IEEE/CVF International Conference on Computer Visio n, 2021, pp. 5461–5470. 3

    2021

  26. [34]

    Avatarbooth : High- quality and Customizable 3D Human Avatar Generation,

    Y . Zeng, Y . Lu, X. Ji, Y . Y ao, H. Zhu, X. Cao, “Avatarbooth : High- quality and Customizable 3D Human Avatar Generation,” 2023 , arXiv preprint arXiv:2306.09864

    2023 arXiv

  27. [35]

    AvatarV erse: High-quality & Sta ble 3D Avatar Creation from Text and Pose,

    H. Zhang, B. Chen, H. Y ang, L. Qu, X. Wang, L. Chen, C. Long , F. Zhu, D. Du, and M. Zheng, “AvatarV erse: High-quality & Sta ble 3D Avatar Creation from Text and Pose,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 7, 2024, pp. 7124-

    2024

  28. [36]

    DensePose: D ense Human Pose Estimation in the Wild,

    R. A. Guler, N. Neverova, and I. Kokkinos, “DensePose: D ense Human Pose Estimation in the Wild,” in Proceedings of the IEE E/CVF Conference on Computer Vision and Pattern Recognition, 201 8, pp. 7297–7306. 3

  29. [37]

    X-oscar: A Progressive Frame work for High- quality Text-guided 3d Animatable Avatar Generation,

    Y . Ma, Z. Lin, J. Ji, et al. “X-oscar: A Progressive Frame work for High- quality Text-guided 3d Animatable Avatar Generation,” 202 4, arXiv preprint arXiv:2405.00954

    arXiv

  30. [38]

    GAvatar: Animatable 3D Gaussian Avatars with Impli cit Mesh Learning,

    Y . Y uan, X. Li, Y . Huang, S. De Mello, K. Nagano, J. Kautz, and U. Iqbal, “GAvatar: Animatable 3D Gaussian Avatars with Impli cit Mesh Learning,” in Proceedings of the IEEE/CVF Conference on Com puter Vision and Pattern Recognition, 2024. 1, 3, 5, 8, 9

    2024

  31. [39]

    Multimodal large language mode l is a human-aligned annotator for text-to-image generation,

    X. Wu, S. Huang, F. Wei, “Multimodal large language mode l is a human-aligned annotator for text-to-image generation,” 2 024, arXiv preprint arXiv:2404.1510

  32. [40]

    Longal ign: A recipe for long context alignment of large language models,

    Y . Bai, X. Lv, J. Zhang, Y . He, J. Qi, L. Hou, J. Li, “Longal ign: A recipe for long context alignment of large language models, ” 2024, arXiv preprint arXiv:2401.18058

    2024 arXiv

  33. [41]

    Learning T ransferable Visual Models From Natural Language Supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Ag arwal, G. Sastry, A. Askell, P . Mishkin, J. Clark et al., “Learning T ransferable Visual Models From Natural Language Supervision,” in Inter national Conference on Machine Learning. PMLR, 2021, pp. 8748–8763. 3

    2021

  34. [7132]

    3 JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 12

    2020

Pith tools

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