Pith. sign in

REVIEW 3 major objections 5 minor 41 references

A lightweight post-adaptation pipeline turns general 3D foundation models into single-image, animation-ready character assets that beat open character generators on geometry, texture, preference, and speed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 17:29 UTC pith:S4JKB2LK

load-bearing objection Solid character-gen systems paper with real metric wins; “product-ready / animation-ready” is ahead of the evidence. the 3 major comments →

arxiv 2607.07817 v1 pith:S4JKB2LK submitted 2026-07-08 cs.CV cs.AI

DreamCharacter-1: From 3D Generative Foundation Models to Product-Ready Character Generation

Weizhe Liu , Yunjie Wu , Xiangqian Shu , Guangwei Wang , Xiangyu Xu , Peng Li , Yujie Li , Hengkai Guo This is my paper
classification cs.CV cs.AI
keywords 3D character generationpost-adaptationgeometry preference optimizationmulti-view texture synthesissparse-voxel inpaintingsingle-image to 3Danimation-ready assetsinference acceleration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most general 3D foundation models still produce oversmoothed, poorly textured, or unriggable characters that fail industrial standards for identity, high-frequency detail, occluded appearance, and skeletal animation. DreamCharacter-1 claims that you do not need to retrain a foundation model from scratch: a lightweight post-adaptation stack—geometry preference optimization on a coarse-to-fine SDF latent, multi-view texture synthesis plus sparse-voxel inpainting for occlusions, and distillation/acceleration—calibrates a pretrained backbone into production-ready character assets from one reference image. On a public single-image character benchmark and a 60-image human study, the method reports higher image-mesh alignment, better front-view texture metrics, stronger multi-criteria preference scores, and lower inference latency than recent DiT-based and character-specific baselines. The practical payoff the paper argues for is a usable path from academic 3D generators to assets that can be rigged, skinned, and driven without obvious collapse.

Core claim

DreamCharacter-1 establishes that targeted post-training of a pretrained 3D foundation model—hierarchical coarse-to-fine geometry with multi-metric geometric preference optimization, dual-stage multi-view texture generation with sparse-voxel occlusion inpainting, and inference acceleration—yields single-image 3D characters that are more identity-consistent, detail-rich, view-stable, and animation-compatible than state-of-the-art open character generation methods, without full backbone retraining.

What carries the argument

The post-adaptation stack: coarse-to-fine Shape-VAE/Shape-DiT geometry in structured SDF latents refined by multi-metric preference RL; Texture-MV multi-view synthesis followed by sparse-voxel Texture-Inpainting for occluded regions; plus distillation and pipeline acceleration for deployment.

Load-bearing premise

That winning open benchmarks, front-view render scores, and a 60-image preference study is enough evidence that the outputs meet full industrial product-ready standards for identity, occluded appearance, and universal rigging.

What would settle it

Run the same single-image test set through automatic rigging and large-range skeletal animation; if DreamCharacter-1 shows higher rates of mesh collapse, texture tearing, or failed skinning than the baselines it claims to beat, the product-ready claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Studios can start from a general 3D foundation and reach usable character assets with post-training rather than full model rebuilds.
  • Coarse-to-fine SDF geometry plus preference rewards can recover thin structures and back-side plausibility that one-stage generators miss.
  • Sparse-voxel inpainting after multi-view projection can complete self-occluded character textures without lighting-entangled artifacts.
  • Distillation and pipeline parallelization make DiT-based character generation fast enough for large-scale asset pipelines.
  • Generated meshes are claimed to be natively compatible with standard rigging, skinning, and motion retargeting workflows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If post-adaptation is sufficient, character generation may converge on modular adapters over ever-larger monolithic 3D foundations.
  • The same dual-stage texture pattern (visible multi-view + 3D native completion) may transfer to other articulated assets with heavy self-occlusion, such as animals or layered garments.
  • Preference RL on anatomical and rigging rewards could become a standard second stage for any human-centric 3D generator that must stay animation-safe.
  • Benchmarking only front-view metrics may systematically under-test the occluded-completion modules that the paper treats as central.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. DreamCharacter-1 is a lightweight post-adaptation framework that calibrates a pretrained 3D foundation backbone (Seed3D-style Shape-VAE + Shape-DiT) for single-image, production-oriented 3D character generation. Geometry uses a hierarchical coarse-to-fine SDF latent pipeline with multi-scale image conditioning, back-view cues, high-quality SFT, and multi-objective RL preference optimization for anatomy, identity, silhouette, and rigging readiness. Texture uses dual-conditional multi-view DiT synthesis, sparse-voxel 3D inpainting for occlusions, plus de-lighting, dual-mesh decoupling, and semantic UV allocation. Acceleration includes distillation, efficient attention, and pipeline parallelism. On the PANIC-3D-style benchmark the method reports SOTA ULIP/Uni3D geometry alignment and SSIM/LPIPS/FID/CLIP-Sim texture metrics versus CharacterGen, StdGEN, Hunyuan3D, TRELLIS, and Pixal3D, plus a 60-image multi-criteria user study and lower wall-clock latency than DiT baselines; qualitative results and a rigged animation example support the claim of animation-ready assets.

Significance. If the results hold under broader validation, the paper offers a practical industrial path: post-adaptation of generic 3D foundations rather than full retraining, with concrete modules (geometry preference RL, sparse-voxel inpainting, dual-mesh texturing, semantic UV) that address character-specific failure modes (thin structures, occlusions, lighting entanglement, rigging). Clear quantitative margins on standard open baselines, multi-axis human preference, and explicit efficiency gains are strengths. The work is timely for game/avatar pipelines. Significance is tempered by the gap between front-view/preference evidence and the stronger 'product-ready / universal skeletal articulation' claim, and by heavy dependence on the authors' own foundation backbone.

major comments (3)
  1. The Abstract and §1 claim 'product-ready' assets satisfying identity, high-frequency geometry, occluded appearance, and 'universal skeletal articulation' / animation-ready topology. §5.1–5.3 and Tables 1–2 report only ULIP/Uni3D, front-view texture metrics, a 60-image preference study, and latency; Fig. 9 is qualitative. No quantitative rigging success rate, skinning-weight quality, deformation error under large articulations, multi-view identity consistency, or thin-structure failure rates are given. This is the load-bearing leap from 'beats open baselines on [3]' to industrial-grade animation readiness and should be closed with measurable auto-rigging / deformation metrics or clearly scoped down.
  2. §5.1–5.2 and Tables 1–2 give point estimates without error bars, confidence intervals, or statistical tests on the fixed PANIC-3D-style split [3]. With free parameters (RL reward aggregation, SDF resolution schedule, dense-view count, distillation NFE) and a stylized/anime-oriented test set, the 'consistently surpassing SOTA' claim needs at least multi-seed variance or a second, more photorealistic hold-out to support the strength of the conclusion.
  3. §2.1 and §3.1.1 describe multi-metric geometric reward models (anatomy, facial identity, silhouette, rigging readiness, image–mesh alignment) and joint RL, yet no reward definitions, aggregation weights, human-annotation protocol, or ablation of RL versus SFT-only appear. Because preference optimization is listed as a core geometry component (Abstract; §1), its contribution to Tables 1–2 and Fig. 1 remains unquantified and should be isolated.
minor comments (5)
  1. §5.1 lists baselines inconsistently (ChaGen/stdGEN/HY in Fig. 7 vs full names in Table 1); standardize naming and cite exact model versions/checkpoints.
  2. Fig. 1 and Fig. 9 captions are informative but raw mesh vs textured views in Figs. 7–8 would benefit from consistent lighting/camera so micro-detail claims are easier to verify.
  3. §4.2–4.3 describe GPU remeshing, thin-shell correction, and stylized augmentation; approximate dataset scale and filter thresholds would aid reproducibility.
  4. Limitations §7 correctly notes SDF watertightness and pipeline length; a short forward pointer from §2.1 would help readers anticipate the non-watertight failure mode.
  5. Date line 'July 10, 2026' and arXiv stamp appear future-dated; correct for archival consistency.

Circularity Check

0 steps flagged

No significant circularity: empirical post-adaptation claims are validated on external open baselines and standard metrics, not by construction from inputs.

full rationale

DreamCharacter-1 is an engineering post-adaptation system (geometry preference RL + Texture-MV/inpainting + acceleration) on a pretrained 3D foundation backbone. Its load-bearing claims are empirical superiority on ULIP/Uni3D, SSIM/LPIPS/FID/CLIP-Sim, a 60-image user study, and wall-clock speed versus external open methods (CharacterGen, StdGEN, Hunyuan3D-2.x, TRELLIS, Pixal3D) on the PANIC-3D-style benchmark. These comparisons use independent metrics and baselines; nothing reduces by definition or by fitting a parameter then re-predicting a near-identical quantity. Self-citations to Seed3D ([25]/26]) identify the backbone being calibrated, which is normal and not load-bearing for the superiority claim (the claim is the post-training gains, measured externally). Reward models and high-quality SFT subsets are internal training signals, but final evaluation is separate human preference and public metrics. No uniqueness theorems, ansatz smuggling, or self-definitional equations appear. The product-ready leap is a validation-gap issue (correctness risk), not circularity. Score 1 only for the mild, non-load-bearing self-backbone dependence; steps empty as no reduction by construction exists.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central claim is empirical systems performance, not a closed-form derivation. Load-bearing premises are domain modeling choices (SDF latents, multi-view→3D texture, preference RL for 'drivability'), reliance on an internal foundation backbone and curated data, and the evaluation protocol as a proxy for industrial readiness. Free parameters are the usual training/inference knobs and reward aggregation weights; invented 'entities' are pipeline modules rather than physical objects.

free parameters (5)
  • Multi-objective RL reward weights / joint geometric preference aggregation
    Anatomical proportions, facial identity, silhouette, rigging readiness, and image-mesh alignment rewards are combined in one RL loop (§3.1.1); relative weights and human/VLM reward calibration are not fixed by theory and directly affect the 'aesthetic and drivability' claim.
  • SDF resolution schedule and coarse-to-fine latent sizes
    Continued training at higher SDF resolution and structured voxel latents (§2.1, §3.1.1) are design choices that control detail vs. stability; values are engineering fits, not derived constants.
  • Dense multi-view count, RoPE offsets, and dual-reference masking probability for Texture-MV
    Post-training densifies views and randomly masks front/back references (§3.2.1); these hyperparameters shape consistency and back-view control.
  • Student distillation NFE / guidance removal and attention-skip policy
    Inference acceleration (§3.1.2, §3.2.3) trades quality for speed via distillation and KV-cache-like skips; operating point is chosen for deployment, not uniquely determined.
  • Quality-filter thresholds and high-quality SFT subset selection
    Hierarchical filtering via mesh stats and VLM tags (§4.1) defines which assets enter high-res SFT; thresholds are free and affect reported fidelity.
axioms (5)
  • domain assumption Signed-distance / watertight mesh latents are an adequate geometry representation for production character assets including thin structures after thin-shell correction.
    Pipeline is built on SDF Shape-VAE/DiT (§2.1); Limitations §7 admits non-watertight and very thin structures are hard—yet product-ready claims assume SDF sufficiency for industrial characters.
  • domain assumption Multi-view 2D texture synthesis plus sparse-voxel 3D inpainting yields view-consistent, identity-preserving full-surface appearance under self-occlusion.
    Core texture design (§2.2); evaluation mostly front-view metrics (§5.2), so full-surface consistency is partly assumed.
  • ad hoc to paper Post-adaptation (SFT + preference RL + light texture post-training) of a generic 3D foundation is enough to reach production character standards without full backbone retraining.
    Central framing in Abstract/§1; success is demonstrated only relative to open baselines on a specific test set.
  • domain assumption Image-mesh metrics (ULIP, Uni3D) and front-view SSIM/LPIPS/FID/CLIP-Sim plus a 60-image user study proxy industrial geometric rationality, texture rationality, and animation readiness.
    §5 evaluation protocol underwrites the 'surpassing SOTA' and product-ready language.
  • standard math Standard rectified-flow / DiT generative modeling and preference optimization transfer from 2D/generic 3D to character geometry and texture as implemented.
    Uses established flow-matching DiT and RL preference machinery (§2–3) as background technology.
invented entities (3)
  • DreamCharacter-1 post-adaptation stack (geometry preference RL + Texture-MV/Inpainting + acceleration suite) no independent evidence
    purpose: Name and organize the end-to-end character calibration pipeline claimed to deliver product-ready assets.
    System-level construct; evidence is internal experiments, not an independently measured physical entity.
  • Multi-metric geometric reward models for anatomy, identity, silhouette, rigging readiness, and image-mesh alignment no independent evidence
    purpose: Drive RL geometry optimization toward aesthetic and animatable characters without extra inference cost.
    Task-specific reward models (§3.1.1) are paper-defined evaluators; no external validation suite is released.
  • Decoupled dual-mesh texturing for overlapping semantic layers (e.g., clothing vs body) no independent evidence
    purpose: Reduce cross-surface texture leakage in occluded layered regions.
    Engineering module (§2.2.3); usefulness shown qualitatively within the paper.

pith-pipeline@v1.1.0-grok45 · 21264 in / 4378 out tokens · 60365 ms · 2026-07-10T17:29:32.535763+00:00 · methodology

0 comments
read the original abstract

We present DreamCharacter-1, a lightweight post-adaptation framework that calibrates pretrained 3D foundation models toward high-fidelity, production-ready 3D character generation. Building upon a 3D foundation backbone, our pipeline incorporates three task-oriented components: (1) geometry post-training, which enhances fine-grained surface details through geometric preference optimization; (2) texture post-training, which synthesizes high-resolution textures and refines the appearance of occluded regions; and (3) inference acceleration, which enables scalable deployment. Extensive quantitative and qualitative experiments demonstrate that DreamCharacter-1 produces visually compelling and structurally robust 3D character assets, consistently surpassing state-of-the-art character generation methods.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 13 internal anchors

  1. [1]

    Lafite: A generative latent field for 3d native texturing

    Chia-Hao Chen, Yuan-Chen Guo, Zi-Xin Zou, Ze Yuan, Guan Luo, Xiaojuan Qi, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Lafite: A generative latent field for 3d native texturing. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19960–19971, 2026

    work page 2026

  2. [2]

    Dora: Sampling and benchmarking for 3d shape variational auto-encoders

    Rui Chen, Jianfeng Zhang, Yixun Liang, Guan Luo, Weiyu Li, Jiarui Liu, Xiu Li, Xiaoxiao Long, Jiashi Feng, and Ping Tan. Dora: Sampling and benchmarking for 3d shape variational auto-encoders. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16251–16261, 2025

    work page 2025

  3. [3]

    Panic-3d: Stylized single-view 3d reconstruction from portraits of anime characters

    Shuhong Chen, Kevin Zhang, Yichun Shi, Heng Wang, Yiheng Zhu, Guoxian Song, Sizhe An, Janus Kristjansson, Xiao Yang, and Matthias Zwicker. Panic-3d: Stylized single-view 3d reconstruction from portraits of anime characters. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    work page 2023

  4. [4]

    Neusdfusion: A spatial-aware generative model for 3d shape completion, reconstruction, and generation

    Ruikai Cui, Weizhe Liu, Weixuan Sun, Senbo Wang, Taizhang Shang, Yang Li, Xibin Song, Han Yan, Zhennan Wu, Shenzhou Chen, et al. Neusdfusion: A spatial-aware generative model for 3d shape completion, reconstruction, and generation. InEuropean Conference on Computer Vision, pages 1–18. Springer, 2024

    work page 2024

  5. [5]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-firstinternational conference on machine learning, 2024

    work page 2024

  6. [6]

    MARS: Mesh AutoRegressive Model for 3D Shape Detailization

    Jingnan Gao, Weizhe Liu, Weixuan Sun, Senbo Wang, Xibin Song, Taizhang Shang, Shenzhou Chen, Hongdong Li, Xiaokang Yang, Yichao Yan, et al. Mars: Mesh autoregressive model for 3d shape detailization.arXiv preprint arXiv:2502.11390, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Hitem3D 2.0: Multi-View Guided Native 3D Texture Generation

    Huiang He, Shengchu Zhao, Jianwen Huang, Jie Li, Jiaqi Wu, Hu Zhang, Pei Tang, Heliang Zheng, Yukun Li, and Rongfei Jia. Hitem3d 2.0: Multi-view guided native 3d texture generation.arXiv preprint arXiv:2604.09231, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  8. [8]

    Stdgen: Semantic-decomposed 3d character generation from single images

    Yuze He, Yanning Zhou, Wang Zhao, Zhongkai Wu, Kaiwen Xiao, Wei Yang, Yong-Jin Liu, and Xiao Han. Stdgen: Semantic-decomposed 3d character generation from single images. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26345–26355, 2025

    work page 2025

  9. [9]

    Stdgen++: A comprehensive system for semantic-decomposed 3d character generation.arXiv preprint arXiv:2601.07660, 2026

    Yuze He, Yanning Zhou, Wang Zhao, Jingwen Ye, Zhongkai Wu, Ran Yi, and Yong-Jin Liu. Stdgen++: A comprehensive system for semantic-decomposed 3d character generation.arXiv preprint arXiv:2601.07660, 2026

    work page arXiv 2026

  10. [10]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advancesin neural information processing systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advancesin neural information processing systems, 30, 2017

    work page 2017

  11. [11]

    Lrm: Large reconstruction model for single image to 3d

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. InInternational Conference on Learning Representations, volume 2024, pages 50678–50702, 2024

    work page 2024

  12. [12]

    Mv-adapter: Multi-view consistent image generation made easy

    Zehuan Huang, Yuan-Chen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 16377–16387, 2025

    work page 2025

  13. [13]

    Hunyuan3D 2.1: From Images to High-Fidelity 3D Assets with Production-Ready PBR Material

    Team Hunyuan3D, Shuhui Yang, Mingxin Yang, Yifei Feng, Xin Huang, Sheng Zhang, Zebin He, Di Luo, Haolin Liu, Yunfei Zhao, et al. Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material. arXiv preprint arXiv:2506.15442, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Ultrashape 1.0: High-fidelity 3d shape generation via scalable geometric refinement.arXiv preprint arXiv:2512.21185, 2025

    Tanghui Jia, Dongyu Yan, Dehao Hao, Yang Li, Kaiyi Zhang, Xianyi He, Lanjiong Li, Yuhan Wang, Jinnan Chen, Lutao Jiang, et al. Ultrashape 1.0: High-fidelity 3d shape generation via scalable geometric refinement.arXiv preprint arXiv:2512.21185, 2025

    work page arXiv 2025

  15. [15]

    Auto-Encoding Variational Bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  16. [16]

    Hunyuan3D 2.5: Towards High-Fidelity 3D Assets Generation with Ultimate Details

    Zeqiang Lai, Yunfei Zhao, Haolin Liu, Zibo Zhao, Qingxiang Lin, Huiwen Shi, Xianghui Yang, Mingxin Yang, Shuhui Yang, Yifei Feng, et al. Hunyuan3d 2.5: Towards high-fidelity 3d assets generation with ultimate details. arXiv preprint arXiv:2506.16504, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Lattice: Democratize high-fidelity 3d generation at scale.arXiv preprint arXiv:2512.03052, 2025

    Zeqiang Lai, Yunfei Zhao, Zibo Zhao, Haolin Liu, Qingxiang Lin, Jingwei Huang, Chunchao Guo, and Xiangyu Yue. Lattice: Democratize high-fidelity 3d generation at scale.arXiv preprint arXiv:2512.03052, 2025. 20

    work page arXiv 2025

  18. [18]

    Pixal3D: Pixel-Aligned 3D Generation from Images

    Dong-Yang Li, Wang Zhao, Yuxin Chen, Wenbo Hu, Meng-Hao Guo, Fang-Lue Zhang, Ying Shan, and Shi-Min Hu. Pixal3d: Pixel-aligned 3d generation from images.arXiv preprint arXiv:2605.10922, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    Craftsman3d: High-fidelity mesh generation with 3d native diffusion and interactive geometry refiner

    Weiyu Li, Jiarui Liu, Hongyu Yan, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman3d: High-fidelity mesh generation with 3d native diffusion and interactive geometry refiner. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5307–5317, 2025

    work page 2025

  20. [20]

    Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models

    Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, et al. Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    work page 2025

  21. [21]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  22. [22]

    Dual marching cubes

    Gregory M Nielson. Dual marching cubes. InIEEE visualization 2004, pages 489–496. IEEE, 2004

    work page 2004

  23. [23]

    Charactergen: Efficient 3d character generation from single images with multi-view pose canonicalization.ACM Transactions on Graphics (TOG), 43(4):1–13, 2024

    Hao-Yang Peng, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, and Shi-Min Hu. Charactergen: Efficient 3d character generation from single images with multi-view pose canonicalization.ACM Transactions on Graphics (TOG), 43(4):1–13, 2024

    work page 2024

  24. [24]

    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, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8763. PmLR, 2021

    work page 2021

  25. [25]

    Seed3d 1.0: From images to high-fidelity simulation-ready 3d assets

    ByteDance Seed. Seed3d 1.0: From images to high-fidelity simulation-ready 3d assets. 2025

    work page 2025

  26. [26]

    Seed3d 2.0: Advancing high-fidelity simulation-ready 3d content generation

    ByteDance Seed. Seed3d 2.0: Advancing high-fidelity simulation-ready 3d content generation. 2026

    work page 2026

  27. [27]

    Seedream 4.0: Toward Next-generation Multimodal Image Generation

    Team Seedream, Yunpeng Chen, Yu Gao, Lixue Gong, Meng Guo, Qiushan Guo, Zhiyao Guo, Xiaoxia Hou, Weilin Huang, Yixuan Huang, et al. Seedream 4.0: Toward next-generation multimodal image generation.arXiv preprint arXiv:2509.20427, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    work page 2024

  29. [29]

    TripoSR: Fast 3D Object Reconstruction from a Single Image

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    AssetGen: Deployable 3D Asset Generation at Interactive Speed

    Dilin Wang, Xiaoyu Xiang, Kihyuk Sohn, Tom Monnier, Yu-Ying Yeh, Thu Nguyen-Phuoc, Jiawen Zhang, Yuchen Fan, Antoine Toisoul, Hyunyoung Jung, et al. Assetgen: Deployable 3d asset generation at interactive speed. arXiv preprint arXiv:2605.26137, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  31. [31]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

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

    work page 2004

  32. [32]

    Native and compact structured latents for 3d generation.Tech report, 2025

    Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, and Jiaolong Yang. Native and compact structured latents for 3d generation.Tech report, 2025

    work page 2025

  33. [33]

    Structured 3d latents for scalable and versatile 3d generation

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21469–21480, 2025

    work page 2025

  34. [34]

    Ulip-2: Towards scalable multimodal pre-training for 3d understanding

    Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27091–27101, 2024

    work page 2024

  35. [35]

    Pandora3D: A Comprehensive Framework for High-Quality 3D Shape and Texture Generation

    Jiayu Yang, Taizhang Shang, Weixuan Sun, Xibin Song, Ziang Cheng, Senbo Wang, Shenzhou Chen, Weizhe Liu, Hongdong Li, and Pan Ji. Pandora3d: A comprehensive framework for high-quality 3d shape and texture generation. arXiv preprint arXiv:2502.14247, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  36. [36]

    One-step Diffusion with Distribution Matching Distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation, 2024.URL https://arxiv. org/abs/2311.18828. 21

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  37. [37]

    Textrix: Latent attribute grid for native texture generation and beyond

    Yifei Zeng, Yajie Bao, Jiachen Qian, Shuang Wu, Youtian Lin, Hao Zhu, Buyu Li, Feihu Zhang, Xun Cao, and Yao Yao. Textrix: Latent attribute grid for native texture generation and beyond. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27104–27113, 2026

    work page 2026

  38. [38]

    3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models.ACM TransactionsOn Graphics (TOG), 42(4):1–16, 2023

    Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models.ACM TransactionsOn Graphics (TOG), 42(4):1–16, 2023

    work page 2023

  39. [39]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    work page 2018

  40. [40]

    Hunyuan3D 2.0: Scaling Diffusion Models for High Resolution Textured 3D Assets Generation

    Zibo Zhao, Zeqiang Lai, Qingxiang Lin, Yunfei Zhao, Haolin Liu, Shuhui Yang, Yifei Feng, Mingxin Yang, Sheng Zhang, Xianghui Yang, et al. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation. arXiv preprint arXiv:2501.12202, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    Uni3d: Exploring unified 3d representation at scale

    Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. InInternational Conference on Learning Representations, volume 2024, pages 46766–46782, 2024. 22

    work page 2024