Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

EmbodiedGen: Towards a Generative 3D World Engine for Embodied Intelligence

T0 review · 4 major / 6 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 EmbodiedGen claims to be the first toolkit that turns images or text into interactive, physics-ready 3D worlds, outputting URDF assets with real scale and watertight geometry.

desk verdict A useful 3D world generation toolkit whose central claim about physically accurate scale and material properties is not actually validated. read the letter →

arxiv 2506.10600 v2 pith:ZBQ2YYCA submitted 2025-06-12 cs.RO cs.CV

Xinjie Wang , Liu Liu , Yu Cao , Ruiqi Wu , Wenkang Qin , Dehui Wang , Wei Sui , Zhizhong Su This is my paper · ORCID
classification cs.ROcs.CV
keywords 3DworldgenerationembodiedintelligenceURDFassetsphysicssimulationGaussiansplattingdigitaltwinsarticulatedobjectstexture
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

EmbodiedGen claims to be the first full toolkit that turns images or text prompts into ready-to-use 3D worlds for embodied intelligence research. Its six modules — Image-to-3D, Text-to-3D, Texture Generation, Articulated Object Generation, Scene Generation, and Layout Generation — produce assets in URDF with real-world scale, physical properties, watertight geometry, and dual 3DGS/mesh representations. The point is to replace hand-built graphics assets with cheap, generated assets that can be dropped into physics simulators for robot training and evaluation. If the claim holds, robot learning data could be produced at far lower cost and with more diversity than manual asset creation allows.

What carries the argument

The machinery that carries the claim is the URDF container plus a physics-expert agent built on large language models. The agent looks at a rendered frontal view and the text prompt, estimates real-world height, and derives the remaining dimensions, mass, friction coefficient, and semantic category, turning a bare graphics object into a simulation-ready one. Around that, the pipeline adds automated quality inspection — aesthetic scoring, foreground segmentation checking, geometric rationality checking — with retry on failure, and a texture back-projection algorithm that removes highlights, super-resolves each view, and fuses multi-view colors through per-pixel confidence weights into a 2K UV map. Articulated objects are generated from a dual-state image pair, where one image shows the rest pose and the other shows the articulated pose, and a graph reasoner infers part connectivity to predict the joints.

What would settle it

Run a set of real household objects with known dimensions and masses through the image-to-3D pipeline and compare the URDF height, mass, and friction coefficient against the measured values; if height errors routinely exceed about 20 percent, the claim of accurate physical properties fails for real-to-sim use. A second check is to simulate a simple stacking or dropping task and compare the long-term stability of generated assets with that of the real objects.

Watch

Extended reading notes

Core claim

EmbodiedGen's core claim is that a single open-source pipeline can generate interactive 3D worlds for embodied AI directly from images or text, producing assets that are ready for physics simulation rather than being mere graphics. The integration layer wraps existing generators — Trellis for image-to-3D, Kolors for text-to-image, DIPO for articulated objects, Diffusion360 and Pano2Room for scenes — with physical-property restoration, real-scale estimation, automated quality inspection with retry, and conversion to URDF. Every asset ends up as both a watertight mesh and a 3D Gaussian splatting representation, with textures that undergo highlight removal and super-resolution before back-projection. The paper states that the resulting assets reach state-of-the-art visual fidelity and physical realism, and that the toolkit is the first of its kind for embodied intelligence research.

Load-bearing premise

The load-bearing premise is that the language-model physics expert can recover real-world scale, mass, and friction from a rendered frontal view and a text prompt accurately enough for physical simulation; this is asserted in the pipeline but not checked against measured values.

Editorial extensions

If this is right

  • Assets produced by EmbodiedGen are in URDF with watertight geometry, so they can be imported into physics simulators without manual cleanup, lowering the cost of building robot training environments.
  • Image-to-3D with physics restoration enables real-to-sim digital twins from a single photograph, letting simulators reproduce specific real objects.
  • Text-to-3D with automated quality filtering supports cheap, large-scale augmentation of asset libraries for training and evaluation.
  • Articulated-object generation from dual-state images provides manipulable cabinets, drawers, and appliances for interaction and manipulation tasks.
  • The dual 3DGS/mesh representation allows the same asset to be rendered photorealistically while remaining suitable for collision and dynamics simulation.

Reading between the lines

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

  • If the scale and physics estimates prove accurate enough, the main bottleneck of embodied-AI data generation could shift from manual modeling to prompt and task design, because an entire interactive scene can be assembled from a single photograph or one sentence.
  • The reported quality-inspection precision of 68.7% and recall of 76.7% imply that fully automated large-scale asset generation will still pass a share of unusable assets, so users may need an additional screening stage before training on them.
  • A controlled benchmark comparing manipulation success in a simulator stocked with EmbodiedGen assets versus hand-built assets would be the natural next test; the paper shows qualitative demonstrations but does not report such a comparison.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. EmbodiedGen is presented as a modular toolkit for generating interactive 3D worlds for embodied intelligence. It comprises Image-to-3D, Text-to-3D, Articulated Object Generation, Texture Generation, Scene Generation, and, as claimed in the abstract, Layout Generation. The pipeline combines existing generators (Trellis, Kolors, DIPO, Pano2Room) with LLM-based physics restoration, automated quality inspection, and URDF conversion, and it also introduces RoboSplatter for 3DGS-based rendering in physical simulators. The central claims are that generated assets achieve state-of-the-art visual fidelity and physical realism, have watertight geometry, carry accurate real-world scale and physical properties, and can be directly imported into simulators such as MuJoCo, Isaac Lab, OpenAI Gym, and SAPIEN.

Significance. If the central claims were fully substantiated, EmbodiedGen would be a practically useful contribution: it addresses a real bottleneck in embodied AI by connecting generative 3D asset pipelines to physics simulators, and it ships an open-source implementation with dual mesh/3DGS representations. The paper also gives credit to existing open models and makes a reasonable design choice in decoupling text-to-3D into text-to-image and image-to-3D for controllable asset production. However, the evidence provided is disproportionately thin relative to the strength of the claims: the only quantitative evaluation is an internal quality-inspection study on 150 cups, and the physics-restoration and watertightness claims are not measured against any external benchmark. The significance of the contribution therefore currently depends on unverified assumptions rather than demonstrated results.

major comments (4)
  1. [Abstract; §3.1; §3.5] The claim that assets have 'accurate physical properties and real-world scale' is not supported by the presented evidence. In §3.1, the physics expert estimates real-world height from a rendered frontal view and text-prompt constraints, then scales width and length proportionally; it also estimates mass and friction from four orthographic views. In §3.5, scene scale is estimated from Qwen-predicted building height. None of these predictions is compared against real measurements, known mesh dimensions, or a physics benchmark, and no volume/density computation is described that could make mass estimation well-posed. A rendered image alone does not determine metric scale, mass, or the pair-dependent friction coefficient, so these outputs are effectively LLM guesses. To retain the central physical-realism claim, the authors should report per-asset errors for scale, mass, and friction against ground-truth data (e.g., real object scans and measured friction pairs) or substantially weaken the claim.
  2. [§3.2, 'Evaluation of Automated Quality Inspection'] The paper's only quantitative evaluation is the 150-cup study, which reports precision 68.7% and recall 76.7% for the automated quality inspection module. These results are modest and are an internal screening check, not a measure of visual fidelity or physical realism. There are no quantitative comparisons of geometry quality, watertightness, texture fidelity, or physical plausibility against any baseline method, and no statistical uncertainty or per-failure-category breakdown is given. The qualitative comparisons in Figures 8, 13, and 17 are useful but do not substantiate the 'state-of-the-art' claim. The authors should add quantitative asset-quality metrics or user studies, or restrict the claim to what the evidence supports.
  3. [§1, contributions; §3.1, Figure 3] The contribution list states that generated assets have 'watertight geometry', but no watertightness test or repair procedure is described or measured. Trellis is not guaranteed to produce closed manifold meshes, and the MeshGeoChecker is a GPT-4o-based visual plausibility check rather than a topological verification. The paper should report the fraction of generated assets that pass a well-defined watertightness test (e.g., closed-manifold and Euler characteristic checks) and describe any repair step, if one exists. Without this, watertightness remains an unverified assertion.
  4. [§3.1; §3.2; §3.5] The evaluation pipeline is partly self-referential because the same LLM family (GPT-4o/Qwen) is used both to generate physical property estimates and to run the quality checkers (ImageSegChecker and MeshGeoChecker in §3.1, PanoSelector in §3.5). This means that systematic errors shared by these models may be missed by the automated inspection. The manual annotation of 150 cups is independent, but it covers only the screening module and not scale, mass, friction, or watertightness. An external anchor, such as comparison with real scan data or a curated set of measured physical properties, is needed before the physical-realism claim can be accepted.
minor comments (6)
  1. [Abstract; §3.5] The abstract lists Layout Generation as one of six key modules, but the body of the paper does not contain a Section 3.6 and gives no method description for layout generation beyond its appearance in Figure 2. The authors should add the missing section or remove the module from the list.
  2. [§3.4, Eq. (2)] The notation in Eq. (2) is unclear: f_b(·) ∈ R^{C×N_b} is said to denote extracted feature vectors at corresponding coordinates, but the text does not define how the reference and search point sets r_b and s_b are selected or how the two feature sets are aligned. Please clarify the dimensions and the matching procedure.
  3. [Algorithm 1] The reshaping on line 2, I_grid ∈ R^{(N×H_0)×W_0×3} ← I ∈ R^{N×H_0×W_0×3}, is ambiguous; the intended view concatenation order should be stated explicitly. Also, the buffers T and C are initialized to zero but their dimensions are not given, and the scatter-add operations on lines 25-26 need a precise definition.
  4. [References] Several duplicate references appear: [22] and [23] both refer to Zero-1-to-3, [30] and [31] both refer to DreamFusion, and [56] and [57] both refer to CLAY. These should be merged or clearly distinguished to avoid citation errors.
  5. [Abstract; §1] There are typographical issues: 'EmbodiedGengenerates' in the Abstract is missing a space, and in §1 the sentence 'existing 3D generation toolkit often fall short' should read 'existing 3D generation toolkits often fall short'.
  6. [§3.3, Figure 10] The PM-X dataset is described as containing 600 articulated objects annotated with rendered images and physical properties, but no statistics or examples of the physical property annotations are provided. A small table or example annotations would help the reader assess the dataset's usefulness.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline derives from external generative models and independent manual evaluation; the physical-property claims are unvalidated but not constructed from their own outputs.

full rationale

EmbodiedGen is a system paper that assembles external generative models (Trellis, Diffusion360, Pano2Room, DIPO, etc.) with added LLM-based modules. The central claim of "accurate physical properties and real-world scale" rests on the "physics expert agent" of Sec. 3.1, which uses GPT-4o/Qwen to estimate height, mass, and friction from rendered views and text prompts. This is an unvalidated estimate, not a circular construction: no equation or definition forces the output to match the input, and no fitted parameter is renamed as a prediction. The only quantitative evaluation (Sec. 3.2) is an independent human annotation of 150 cups against the automated quality-inspection module, reporting precision 68.7% and recall 76.7%; this does not reduce to the system's own outputs. The paper does cite prior work by its authors (DIPO [48]) as the articulated-object module, but this is a normal component reuse and is not invoked as a uniqueness theorem or as evidence for the physical-realism claim. The absence of external validation for scale/mass/friction is a correctness/evidence concern, not a circularity concern. No equation-level circularity, self-definitional step, or fitted-input-called-prediction was found.

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

The toolkit introduces no new physical entities. It reuses existing generative models and adds software modules (GeoLifter, quality checkers, RoboSplatter). The main embedded assumptions are about the reliability of LLM-based estimation and the watertightness of generated meshes.

free parameters (3)
  • Back-projection angle threshold = 70 degrees
    Algorithm 1 uses theta = 70 degrees to exclude faces viewed at large angles; chosen by hand, no sensitivity analysis.
  • Spatial loss weight = lambda_spatial = 0.02
    Loss 3 in Section 3.4 sets lambda_ldm = 1 and lambda_spatial = 0.02; no ablation reported.
  • LLM-predicted per-asset scale/mass/friction = not disclosed per asset
    The physics expert agent (GPT-4o/Qwen) estimates height, mass, friction per asset; these values are used as ground truth for the URDF but never compared to real measurements.
assumptions (3)
  • domain assumption Trellis output meshes are watertight or can be made watertight by the pipeline
    The paper claims watertight geometry (Section 5) but does not describe a meshing or repair step to guarantee it.
  • ad hoc to paper LLM visual estimation of physical properties is accurate enough for simulation
    Section 3.1 relies on GPT-4o/Qwen to predict scale and physics from images without external validation.
  • domain assumption The automatic quality checkers' pass/fail criteria align with true simulation-usability
    Section 3.2 measures precision/recall against manual labels, but manual usability is not the same as physical correctness in a simulator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EmbodiedGen: Towards a Generative 3D World Engine for Embodied Intelligence." pith.science (2026). https://pith.science/paper/ZBQ2YYCA

@misc{pith2026250610600,
  author       = {Pith},
  title        = {Pith review of: EmbodiedGen: Towards a Generative 3D World Engine for Embodied Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZBQ2YYCA}},
  note         = {Machine review of arXiv:2506.10600}
}
read the original abstract

Constructing a physically realistic and accurately scaled simulated 3D world is crucial for the training and evaluation of embodied intelligence tasks. The diversity, realism, low cost accessibility and affordability of 3D data assets are critical for achieving generalization and scalability in embodied AI. However, most current embodied intelligence tasks still rely heavily on traditional 3D computer graphics assets manually created and annotated, which suffer from high production costs and limited realism. These limitations significantly hinder the scalability of data driven approaches. We present EmbodiedGen, a foundational platform for interactive 3D world generation. It enables the scalable generation of high-quality, controllable and photorealistic 3D assets with accurate physical properties and real-world scale in the Unified Robotics Description Format (URDF) at low cost. These assets can be directly imported into various physics simulation engines for fine-grained physical control, supporting downstream tasks in training and evaluation. EmbodiedGen is an easy-to-use, full-featured toolkit composed of six key modules: Image-to-3D, Text-to-3D, Texture Generation, Articulated Object Generation, Scene Generation and Layout Generation. EmbodiedGen generates diverse and interactive 3D worlds composed of generative 3D assets, leveraging generative AI to address the challenges of generalization and evaluation to the needs of embodied intelligence related research. Code is available at https://horizonrobotics.github.io/robot_lab/embodied_gen/index.html.

Figures

Figures reproduced from arXiv: 2506.10600 by the authors.

Figure 1
Figure 1. EmbodiedGen, a toolkit for embodied intelligence interactive 3D world generation. EmbodiedGen enables controllable gen￾eration of rigid and articulated assets with accurate real-world scale and physical properties, along with stylistically diverse background generation and visually rich texture generation and editing. These assets can be seamlessly integrated into various simulators such as OpenAI Gym[4], Isaac Lab[… view at source ↗
Figure 2
Figure 2. The framework of EmbodiedGen. It enables the creation of a digital twin within a simulation environment from a single image. Alternatively, given a task description, EmbodiedGen autonomously generates the scene layout, synthesizes detailed 3D object assets, and arranges them in semantically and physically plausible configurations. This facilitates the effortless construction of an interactive 3D world, supporting a … view at source ↗
Figure 3
Figure 3. Overview of EmbodiedGen Image-to-3D Pipeline. From a single image, the system generates mesh and 3DGS assets, conducts automatic quality inspectioin (aesthetics, segmentation, geometry), and re-generate failed outputs by auto-adjusted settings. A physics expert module restores real-world scale and physical semantics, and the assets are saved in URDF format. generates 3D objects from text descriptions for low cost, h… view at source ↗
Figures from the paper (20 more)
Figure 4
Figure 4. Figure 4: AestheticChecker is used to evaluate the texture quality of generated assets. Assets displaying richer texture details receiv￾ing higher scores. NO. The segmentation has truncated the bottom part of the chair's legs. NO. The legs of the chair are significantly truncate…
Figure 5
Figure 5. Figure 5: Examples of segmentation failure cases automatically [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Examples of geometric rationality inspection by [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: From left to right: the original image, the result of our [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of Text-to-3D result. The left [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: EmbodiedGen Text-to-3D module for large-scale 3D as￾set generation. A prompt generator decomposes user requirements into prompts targeting different asset styles. The pipeline proceeds through text-to-image and image-to-3D stages, each equipped with automatic quality i…
Figure 10
Figure 10. Figure 10: Visual examples of articulated objects constructed by [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 11
Figure 11. Figure 11: Overview of EmbodiedGen Texture Generation Module. Given a mesh and a text prompt, the module generates six-view consistent textures with controllable styles via text, reference image, or both. Geometry-aware conditions (normals, positions, masks) are extracted and in…
Figure 12
Figure 12. Figure 12: Blue dots: reference points; red dots: projected corre [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]
Figure 13
Figure 13. Figure 13: EmbodiedGen texture generation module effectively adheres to text descriptions, generating high-quality textures with strong spatial and geometric consistency. It also demonstrates robust control over text generation on textures, accurately rendering common Chinese an…
Figure 14
Figure 14. Figure 14: Overview of EmbodiedGen 3D Scene Generation. A panorama is generated from a text prompt or input image, guided by a style prompt. After quality assessment via vlm-based selector, a refined mesh and 3DGS[19] are generated by panorama projection, inpainting, and repair.…
Figure 15
Figure 15. Figure 15: Qualitative Comparison With and Without Style [PITH_FULL_IMAGE:figures/full_fig_p010_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison with and without super [PITH_FULL_IMAGE:figures/full_fig_p011_16.png]
Figure 17
Figure 17. Figure 17: Qualitative Comparison with WorldGen[51] Our method produces more detailed textures and more complete ge￾ometry than WorldGen, under both text and image input settings. representations. Additionally, the coordinate system is re￾centered to the floor plane of the scene…
Figure 18
Figure 18. Figure 18: EmbodiedGen Image-to-3D: large-scale and diverse 3D object asset generation [PITH_FULL_IMAGE:figures/full_fig_p012_18.png]
Figure 19
Figure 19. Figure 19: EmbodiedGen texture generation module enables rich and flexible visual texture editing. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_19.png]
Figure 20
Figure 20. Figure 20: EmbodiedGen Image-to-3D: Digital twin creation and simulation in RoboSplatter and MuJoCo[42]. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_20.png]
Figure 21
Figure 21. Figure 21: EmbodiedGen Image-to-3D: Real-to-sim closed-loop simulation evaluation of a grasping model in Isaac Lab environment [26] [PITH_FULL_IMAGE:figures/full_fig_p014_21.png]
Figure 22
Figure 22. Figure 22: Interaction 3D World Generation with EmbodiedGen. EmbodiedGen enables easy construction of diverse interactive 3D worlds for simulating and evaluating dual-arm shoe-grasping tasks in RoboTwin[28] [PITH_FULL_IMAGE:figures/full_fig_p014_22.png]
Figure 23
Figure 23. Figure 23: EmbodiedGen Text-to-3D: Real-to-sim object transfer and quadruped navigation with obstacle avoidance in OpenAI Gym[4]. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_23.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    SimFoundry automates zero-shot real-to-sim scene generation from video, producing digital twins and cousins that enable policy training with 0.911 mean Pearson correlation to real-world results and 17-40% success gain...

Reference graph

Works this paper leans on

58 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    arXiv

  2. [2]

    Rmbg-1.4: Background removal model.https: //huggingface.co/briaai/RMBG-1.4, 2023

    BRIA AI. Rmbg-1.4: Background removal model.https: //huggingface.co/briaai/RMBG-1.4, 2023. Ac- cessed: 2025-05-19. 4

    work page 2023

  3. [3]

    Meta 3d gen, 2024

    Raphael Bensadoun, Tom Monnier, Yanir Kleiman, Filippos Kokkinos, Yawar Siddiqui, Mahendra Kariya, Omri Harosh, Roman Shapovalov, Benjamin Graham, Emilien Garreau, Animesh Karnewar, Ang Cao, Idan Azuri, Iurii Makarov, Eric-Tuan Le, Antoine Toisoul, David Novotny, Oran Gafni, Natalia Neverova, and Andrea Vedaldi. Meta 3d gen, 2024. 3

    work page 2024

  4. [4]

    Openai gym, 2016

    Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym, 2016. 1, 3, 11, 14

    work page 2016

  5. [5]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction, 2024

    David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction, 2024. 2

    work page 2024

  6. [6]

    Springer Nature Switzer- land, 2024

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai.MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-view Images, page 370–386. Springer Nature Switzer- land, 2024. 2

    work page 2024

  7. [7]

    Urdformer: A pipeline for constructing articulated simulation environments from real-world images,

    Zoey Chen, Aaron Walsman, Marius Memmel, Kaichun Mo, Alex Fang, Karthikeya Vemuri, Alan Wu, Dieter Fox, and Abhishek Gupta. Urdformer: A pipeline for constructing articulated simulation environments from real-world images,

  8. [8]

    Luciddreamer: Domain-free genera- tion of 3d gaussian splatting scenes, 2023

    Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free genera- tion of 3d gaussian splatting scenes, 2023. 3

    work page 2023

Show all 58 references
  1. [9]

    Auto- mated creation of digital cousins for robust policy learning,

    Tianyuan Dai, Josiah Wong, Yunfan Jiang, Chen Wang, Cem Gokmen, Ruohan Zhang, Jiajun Wu, and Li Fei-Fei. Auto- mated creation of digital cousins for robust policy learning,

  2. [10]

    Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    Jinze Bai et al. Qwen technical report.arXiv preprint arXiv:2309.16609, 2023. 4, 10

    2023 arXiv

  3. [11]

    Lrm: Large reconstruction model for single image to 3d, 2024

    Yicong Hong et al. Lrm: Large reconstruction model for single image to 3d, 2024. 2

    2024

  4. [12]

    Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation, 2025

    Zibo Zhao et al. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation, 2025. 2, 8

    2025

  5. [13]

    Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models, 2023

    Mengyang Feng, Jinlin Liu, Miaomiao Cui, and Xuansong Xie. Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models, 2023. 10

    2023

  6. [14]

    Act the part: Learning interaction strategies for articulated object part discovery

    Samir Yitzhak Gadre, Kiana Ehsani, and Shuran Song. Act the part: Learning interaction strategies for articulated object part discovery. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 15752–15761,

  7. [15]

    rembg, 2025

    Daniel Gatis. rembg, 2025. A tool to remove images back- ground. 4

    2025

  8. [16]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 2

    2020

  9. [17]

    Mv-adapter: Multi-view consistent image generation made easy, 2024

    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, 2024. 2

    2024

  10. [18]

    Gen2sim: Scaling up robot learning in simulation with gen- erative models, 2023

    Pushkal Katara, Zhou Xian, and Katerina Fragkiadaki. Gen2sim: Scaling up robot learning in simulation with gen- erative models, 2023. 3

    2023

  11. [19]

    3d gaussian splatting for real-time radiance field rendering, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023. 3, 4, 5, 10, 11

    2023

  12. [20]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything.arXiv:2304.02643, 2023. 4

    2023 arXiv

  13. [21]

    Chang, Manolis Savva, and Ali Mahdavi-Amiri

    Jiayi Liu, Denys Iliash, Angel X. Chang, Manolis Savva, and Ali Mahdavi-Amiri. Singapo: Single image controlled gen- eration of articulated parts in objects, 2025. 3

    2025

  14. [23]

    Zero-1-to-3: Zero-shot one image to 3d object, 2023

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

    2023

  15. [24]

    Text-guided texturing by synchronized multi-view diffusion

    Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 8

    2023 arXiv

  16. [25]

    Meshy.ai: Ai-powered 3d mesh generation, 2025

    Meshy.ai. Meshy.ai: Ai-powered 3d mesh generation, 2025. Accessed: 2025-05-09. 8

    2025

  17. [26]

    Orbit: A unified simulation framework for interactive robot learning environments.IEEE Robotics and Automation Letters, 8(6): 3740–3747, 2023

    Mayank Mittal, Calvin Yu, Qinxi Yu, Jingzhou Liu, Nikita Rudin, David Hoeller, Jia Lin Yuan, Ritvik Singh, Yun- rong Guo, Hammad Mazhar, Ajay Mandlekar, Buck Babich, Gavriel State, Marco Hutter, and Animesh Garg. Orbit: A unified simulation framework for interactive robot lear...

    2023

  18. [27]

    Where2act: From pixels to actions for articulated 3d objects

    Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhi- nav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813–6823, 2021. 7

    2021

  19. [28]

    Robotwin: Dual-arm robot benchmark with generative digi- tal twins (early version), 2025

    Yao Mu, Tianxing Chen, Shijia Peng, Zanxin Chen, Zeyu Gao, Yude Zou, Lunkai Lin, Zhiqiang Xie, and Ping Luo. Robotwin: Dual-arm robot benchmark with generative digi- tal twins (early version), 2025. 3, 14

    2025

  20. [29]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. InInternational conference on machine learning, pages 8162–8171. PMLR,

  21. [31]

    Barron, and Ben Milden- hall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion, 2022. 2

    2022

  22. [32]

    Pano2room: Novel view synthesis from a single indoor panorama

    Guo Pu, Yiming Zhao, and Zhouhui Lian. Pano2room: Novel view synthesis from a single indoor panorama. InSIG- 15 GRAPH Asia 2024 Conference Papers, page 1–11. ACM,

    2024

  23. [33]

    Understanding 3d ob- ject interaction from a single image

    Shengyi Qian and David F Fouhey. Understanding 3d ob- ject interaction from a single image. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 21753–21763, 2023. 7

    2023

  24. [34]

    Language models are unsuper- vised multitask learners

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsuper- vised multitask learners. 2019. 2

    2019

  25. [35]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2

    2021

  26. [36]

    Texture: Text-guided texturing of 3d shapes, 2023

    Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes, 2023. 8

    2023

  27. [37]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 8

    2022

  28. [38]

    Aesthetic subsets in laion 2170337258 samples, 2025

    Christoph Schuhmann. Aesthetic subsets in laion 2170337258 samples, 2025. Retrieved May 16, 2025. 4

    2025

  29. [39]

    Mvdream: Multi-view diffusion for 3d gen- eration, 2024

    Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d gen- eration, 2024. 2

    2024

  30. [40]

    Kolors: Effective training of diffusion model for photorealistic text-to-image synthesis.arXiv preprint,

    Kolors Team. Kolors: Effective training of diffusion model for photorealistic text-to-image synthesis.arXiv preprint,

  31. [41]

    Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation, 2025.https : / / huggingface

    Tencent Hunyuan3D Team. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation, 2025.https : / / huggingface . co / tencent/Hunyuan3D-2/tree/main/hunyuan3d- delight-v2-0. 3, 5

    2025

  32. [42]

    Mujoco: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026–5033. IEEE, 2012. 1, 3, 4, 11, 13

    2012

  33. [43]

    Perf: Panoramic neural radiance field from a single panorama, 2023

    Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama, 2023. 3

    2023

  34. [44]

    Real-esrgan: Training real-world blind super-resolution with pure synthetic data

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. InInternational Conference on Com- puter Vision Workshops (ICCVW). 5, 10

  35. [45]

    Matchmaker: Automated asset generation for robotic assembly, 2025

    Yian Wang, Bingjie Tang, Chuang Gan, Dieter Fox, Kaichun Mo, Yashraj Narang, and Iretiayo Akinola. Matchmaker: Automated asset generation for robotic assembly, 2025. 3

    2025

  36. [46]

    Unique3d: High-quality and efficient 3d mesh generation from a single image, 2024

    Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image, 2024. 2

    2024

  37. [47]

    Lamp: Learn a motion pattern for few-shot video generation

    Ruiqi Wu, Liangyu Chen, Tong Yang, Chunle Guo, Chongyi Li, and Xiangyu Zhang. Lamp: Learn a motion pattern for few-shot video generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7089–7098, 2024. 2

    2024

  38. [48]

    Dipo: Dual-state images controlled articulated object generation powered by diverse data, 2025

    Ruiqi Wu, Xinjie Wang, Liu Liu, Chunle Guo, Jiaxiong Qiu, Chongyi Li, Lichao Huang, Zhizhong Su, and Ming-Ming Cheng. Dipo: Dual-state images controlled articulated object generation powered by diverse data, 2025. 7

    2025

  39. [49]

    Chang, Leonidas J

    Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive envi- ronment. InThe IEEE Conference on Computer Vision and ...

    2020

  40. [50]

    Structured 3d latents for scalable and versatile 3d gen- eration.arXiv preprint arXiv:2412.01506, 2024

    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 gen- eration.arXiv preprint arXiv:2412.01506, 2024. 3, 4, 6

    2024 arXiv

  41. [51]

    Worldgen: Generate any 3d scene in seconds

    Ziyang Xie. Worldgen: Generate any 3d scene in seconds. https://github.com/ZiYang- xie/WorldGen,

  42. [52]

    Grm: Large gaussian reconstruction model for ef- ficient 3d reconstruction and generation, 2024

    Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wet- zstein. Grm: Large gaussian reconstruction model for ef- ficient 3d reconstruction and generation, 2024. 2

    2024

  43. [53]

    Layer- pano3d: Layered 3d panorama for hyper-immersive scene generation, 2025

    Shuai Yang, Jing Tan, Mengchen Zhang, Tong Wu, Yixuan Li, Gordon Wetzstein, Ziwei Liu, and Dahua Lin. Layer- pano3d: Layered 3d panorama for hyper-immersive scene generation, 2025. 3

    2025

  44. [54]

    Paint3d: Paint anything 3d with lighting-less texture diffusion models

    Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252–4262, 2024. 8

    2024

  45. [55]

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

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

    2023

  46. [56]

    Clay: A controllable large-scale generative model for creating high-quality 3d assets, 2024

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets, 2024. 3

    2024

  47. [57]

    Clay: A controllable large-scale generative model for creating high-quality 3d assets, 2024

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets, 2024. 2

    2024

  48. [58]

    Dreammat: High-quality pbr material generation with geometry- and light-aware dif- fusion models, 2024

    Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. Dreammat: High-quality pbr material generation with geometry- and light-aware dif- fusion models, 2024. 2

    2024

  49. [59]

    Holodreamer: Holistic 3d panoramic world generation from text descriptions, 2024

    Haiyang Zhou, Xinhua Cheng, Wangbo Yu, Yonghong Tian, and Li Yuan. Holodreamer: Holistic 3d panoramic world generation from text descriptions, 2024. 3

    2024

  50. [60]

    Dreamscene360: Uncon- strained text-to-3d scene generation with panoramic gaussian splatting, 2024

    Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Uncon- strained text-to-3d scene generation with panoramic gaussian splatting, 2024. 3 16

    2024

Pith tools

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