Pith. sign in

REVIEW 3 major objections 6 minor 62 references

One model completes 3D shapes from images, masks, cameras, and partial point clouds, outperforming task-specific pipelines on both clean and occluded benchmarks.

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 ↗

A single diffusion model, fine-tuned from Hunyuan3D, completes 3D shapes from any mix of images, masks, cameras, and point clouds, reaching state-of-the-art or competitive accuracy on Toys4K and OmniObject3D.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid, honest engineering paper with a plausible unified conditioning stack; verdict conditional on code release and real-occlusion validation. the 3 major comments →

arxiv 2607.20660 v1 pith:NXZW7O73 submitted 2026-07-22 cs.CV

Axolotl3D: a Unified Framework for Faithful 3D Shape Completion

Anita Hu , Maria Shugrina This is my paper
classification cs.CV
keywords 3D shape completionmulti-modal generationocclusion-awarediffusion transformerpoint cloud conditioningcamera conditioning3D editingamodal completion
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.

The reading

The paper argues that single-view generation, multi-view reconstruction, occlusion completion, and geometry-aware editing are all forms of constrained 3D completion from incomplete observations. It proposes Axolotl3D, a single diffusion-based model that takes posed images, per-view visibility masks, camera parameters, and a partial point cloud, and outputs a complete mesh. The point cloud acts as a geometric anchor for fidelity, camera embeddings align multiple views in a shared coordinate frame, and mask-biased attention keeps occluded image regions from corrupting the result. Trained on 407k synthetic meshes with simulated occluders, point dropout, and edit masks, the model beats prior single- and multi-view completion methods on Toys4K and OmniObject3D, with the largest gains in single-view and occluded settings. A sympathetic reader would care because a single unified model could replace separate tools across common 3D capture, reconstruction, and editing workflows.

Core claim

Axolotl3D shows that one model, jointly conditioned on images, visibility masks, camera parameters, and a partial point cloud, can complete 3D shapes faithfully across single-view generation, sparse multi-view reconstruction, occlusion completion, and geometry-aware editing. The point cloud is the geometric anchor: it keeps generated geometry faithful to known regions, while camera Plücker embeddings and mask-biased cross-attention ensure multi-view consistency and prevent occluded image tokens from interfering. A unified training strategy synthesizes diverse conditioning regimes—random brush/box occluders, large-area point dropout, and simulated edit regions—from large-scale mesh data, enab

What carries the argument

The central mechanism is multi-modal condition-token fusion through mask-biased cross-attention in a flow-based diffusion transformer. Images are encoded with a self-supervised visual encoder and enriched with Plücker ray embeddings from camera parameters; the partial point cloud is encoded into a fixed-length latent by a point-cloud encoder; and per-view visibility masks are patchified into an attention bias that sets occluded visual tokens to negative infinity. A gated feed-forward network projects each modality's tokens to a shared channel dimension, and learnable modality embeddings keep visual and geometric tokens distinct. The training pipeline is equally load-bearing: it synthesizes s

Load-bearing premise

The training-time synthetic data pipeline—random brush/box occluders, large-area point dropout, and simulated edit masks—adequately represents the partial, noisy, and misaligned observations encountered in real capture.

What would settle it

Feed the model a real SLAM-style capture with predicted monocular depth and estimated camera poses, without manual alignment, and compare F-score against a high-quality scan. If performance drops substantially below the simulated-test numbers while a baseline trained on real partial data does not, or if adding per-point Gaussian noise of 0.1 in normalized units plus 5-degree pose jitter collapses geometric metrics, the claim that the synthetic regime transfers to real inputs is refuted.

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

If this is right

  • A single model can replace separate amodal-completion, multi-view reconstruction, and editing modules, simplifying real-world capture and content-creation pipelines.
  • The point-cloud anchor makes outputs faithful to known geometry, which is directly useful for physics simulation, robotics, and digital-twins where complete watertight meshes are needed.
  • Explicit camera conditioning makes the model resilient to moderate camera pose and focal-length perturbations, as shown under simulated noise.
  • The synthetic conditioning-regime training recipe can be reused to build unified multi-modal 3D models for other input types or tasks.
  • Lower output variance across runs means more consistent and predictable results for production use.

Where Pith is reading between the lines

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

  • The main transfer risk is the synthetic-to-real gap: if real capture noise (depth discontinuities, misalignment, irregular surfaces) exceeds what the random occluders and point dropout simulate, benchmark gains may not carry over to actual SLAM or monocular-depth inputs.
  • The editing formulation is a natural testbed for inpainting-based 3D editing; a direct test would measure how well untouched regions are preserved by comparing pre-edit and post-edit geometry outside the edit box.
  • Because training fixes camera distance and intrinsics, extending to variable intrinsics and closer viewpoints is a concrete next step that would broaden applicability to handheld captures.
  • The framework is shape-only; swapping or augmenting the latent decoder with a textured representation could extend it to texture completion, which the authors list as future work.
Share X Bluesky LinkedIn Reddit HN

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 / 6 minor

Summary. The paper introduces Axolotl3D, a diffusion-based 3D shape completion model built by adapting Hunyuan3D 2.1's DiT. It takes up to six posed images, visibility masks, and a partial point cloud; images are encoded with DINOv2 and Plücker camera embeddings, points with a pretrained VecSetX encoder, and the resulting tokens feed mask-biased cross-attention. Training synthesizes single-view, sparse multi-view, occluded, large-dropout, and editing conditioning regimes from 407k meshes. On Toys4K and OmniObject3D, the model is evaluated with F-score/vIoU/CD under clean and occluded conditions, with ablations, 2048-point robustness, and predicted-depth robustness. The paper claims state-of-the-art performance while acknowledging that the multi-view results are comparable to concurrent ShapeR; applications include editing and image-to-3D.

Significance. If the framework and results hold, the contribution is useful: a single model unifying posed multi-view reconstruction, single-view completion under occlusion, and geometry-aware editing, with explicit point-cloud anchoring. The experimental design is generally careful: ICP alignment for all methods, point-budget control, a 2048-point ablation, predicted-depth and camera-perturbation tests, and genuine ablations showing that the point condition is the dominant component. The main value is system-level unification rather than a new mathematical derivation; the evidence for 'faithful' behavior in real capture is the weakest link.

major comments (3)
  1. [Table 1, §4.3] The claim that Axolotl3D is state-of-the-art in all evaluated settings is not supported by the reported metrics. In the multi-view OmniObject3D rows, ShapeR obtains higher vIoU (0.4412 vs 0.4127 without occlusion; 0.4123 vs 0.3967 with occlusion) and lower CD (0.1329 vs 0.1389 without; 0.1443 vs 0.1467 with), while Axolotl3D wins on F-score and variance. The abstract's 'state-of-the-art performance' should be qualified, e.g., 'comparable to ShapeR on multi-view OmniObject3D and superior on single-view and Toys4K'; §4.3 acknowledges this only in the body.
  2. [§3.2, §C.1, §E] The central 'faithful completion under occlusion' evidence is obtained with synthetic occluders from the same family as the training augmentation (brush/box masks in §3.2; box/cylinder/sphere/cone primitives in §C.1), so Table 1 largely measures fit to that synthetic occlusion prior. The only real-world evidence is qualitative (§4.6, Figs. 8/11, §C.3), and §E lists sensitivities to noisy points, fixed camera distances/intrinsics, and thin structures. The claim that the method is occlusion-robust for real capture is therefore not yet quantitatively established; temper the claim or add a small real-capture/sensor-noise evaluation.
  3. [§4.2] AmodalGen3D [61] is the closest prior combining multi-view reasoning, occlusion handling, and partial point conditioning, yet it is omitted because its code is unavailable. While code unavailability is a practical constraint, the paper still claims state-of-the-art performance; without this baseline the benchmark cannot resolve whether the unified framework advances over the strongest specialized/nearest system. State explicitly that the claim is relative to available methods and to the listed baselines, or include numbers from the cited report if trustworthy.
minor comments (6)
  1. [§4.1] Typo: 'initalization' should be 'initialization'.
  2. [Tables 1–3] The captions use 'Tb.' inconsistently; standardize to 'Table'.
  3. [§C.1] The sentence 'For each mesh, we sample 3 occluders from the set of basic shapes with equal probability.' appears twice in consecutive paragraphs; remove the duplicate.
  4. [Fig. 15] Caption contains a stray space: 'T oys4k' should be 'Toys4K'.
  5. [Table 2] The caption should state why camera-perturbation entries are N/A for SAM3D and Hy3D-Omni; the reason currently appears only in the body text.
  6. [Eq. (6)] The notation for F_m and F_c is unclear; define the elementwise product of masks and the normalization explicitly.

Circularity Check

0 steps flagged

No significant circularity: Axolotl3D's claims are empirical and evaluated on held-out benchmarks.

full rationale

The paper's central claim is an empirical system result: a diffusion transformer is fine-tuned to denoise shape latents conditioned on images, masks, cameras, and point clouds, and then evaluated on held-out Toys4K/OmniObject3D meshes. No load-bearing step reduces to its own inputs by construction. The point-cloud condition is an input signal, not a fitted target; the diffusion training loss is defined on ground-truth shape latents, not on the evaluation metrics. The reported F-score/vIoU/CD numbers compare generated meshes against held-out ground-truth meshes, and no training constant is fitted to those benchmarks. Self-citations ([7], [13], [30]) appear only in the simulation application and do not support the main completion claim. The closest concern is that the main-table point conditions are back-projected from ground-truth depth and that occluders are synthetic geometric primitives; however, this is an evaluation-protocol and synthetic-to-real domain-gap issue, not a circular reduction, and the same ground-truth-derived conditions are given to all point-conditioned baselines. The paper's own §E candidly lists sensitivity to noisy points, fixed camera distances/intrinsics, and thin-structure failures, which further indicates the claimed gains are empirical and contingent rather than definitionally forced.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on the pretrained Hunyuan3D prior, on VecSetX and DINOv2 as frozen perceptual/geometric encoders, on a synthetic augmentation distribution standing in for real partial observations, and on geometric metrics as proxies for fidelity. None of these are independently established in this paper; they are inherited or assumed.

free parameters (4)
  • Synthetic occlusion probabilities (p_occl = 0.25 single-view, 1.0 multi-view; p_mask = 0.5) = 0.25 / 1.0 / 0.5 (hand-set)
    These augmentation mixing probabilities define the training distribution for the claimed occlusion robustness and are chosen by hand (§3.2); changing them would change benchmark results.
  • Conditioning point count = 8192 (2048 in ablation; 2048 cap for Hy3D-Omni)
    Evaluation and training use fixed numbers of conditioning points; performance depends on this capacity (§4.1, Table 3).
  • Inference CFG scale and steps = 5 and 50
    Classifier-free-guidance scale and sampling steps set at inference; they trade fidelity vs. diversity and are not swept/reported (§4.1).
  • Fine-tuned model weights (Hunyuan3D DiT + adapters) = not released
    The central benchmark numbers are produced by 500k optimization steps on 407k meshes; these fitted weights are the actual artifact behind the claims and are not shipped.
axioms (5)
  • domain assumption Synthetic occluders (box, cylinder, sphere, cone) and brush/bbox masks are representative of real-world occlusions.
    The claimed occlusion robustness is trained and evaluated entirely on generated occluders (§3.2, §C.1); if real occlusions differ, the SOTA claim may not transfer. The paper's own limitations mention noise and camera robustness.
  • domain assumption VecSetX, pretrained on ShapeNet-v2, encodes partial point clouds with fidelity sufficient to anchor faithful completion.
    Geometry encoding is lifted from an external pretrained model; its fidelity is assumed rather than demonstrated in this paper (§3.1).
  • domain assumption DINOv2 visual features plus Plücker camera embeddings are sufficient to align multiple views in a shared 3D coordinate frame.
    The model relies on this for multi-view consistency (§3.1, Eq. 1); ablations show cameras matter mainly in single-view, so the sufficiency assumption is not fully validated.
  • domain assumption Fine-tuning Hunyuan3D 2.1 transfers its large-scale single-image 3D prior to multi-modal completion.
    The method inherits the pretrained shape prior; if the prior is incompatible with completion inputs, the framework fails. This is the backbone choice (§3.1).
  • domain assumption F-score, vIoU, and Chamfer Distance after ICP alignment are valid proxies for faithful 3D completion.
    All quantitative conclusions rest on these proxies; ICP alignment may hide coordinate-frame errors (§4.2, supplement D.1).

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Axolotl3D: a Unified Framework for Faithful 3D Shape Completion." pith.science (2026). https://pith.science/paper/NXZW7O73

@misc{pith2026260720660,
  author       = {Pith},
  title        = {Pith review of: Axolotl3D: a Unified Framework for Faithful 3D Shape Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NXZW7O73}},
  note         = {Machine review of arXiv:2607.20660}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent 3D generative models produce high-quality geometry from a single image using large-scale priors and diffusion architectures. However, they assume complete visibility and single-view inputs, limiting applicability in multi-view, occluded, or editing scenarios. Although prior works address these challenges individually, they lack a unified framework for controllable 3D completion under diverse conditioning signals. We present Axolotl3D, a multi-modal and occlusion-aware 3D generation model that jointly conditions on images, visibility masks, camera parameters, and a partial point cloud. The point cloud serves as a geometric anchor promoting faithful shape completion, while camera parameters ensure consistent multi-view alignment in a shared 3D coordinate system. A unified training strategy synthesizes diverse conditioning regimes from large-scale 3D data, enabling robust cross-modal reasoning. Experiments on Toys4K and OmniObject3D demonstrate state-of-the-art performance under both clean and occluded settings, as well as strong results in real-world reconstruction and geometry-consistent editing.

Figures

Figures reproduced from arXiv: 2607.20660 by Anita Hu, Maria Shugrina.

Figure 1
Figure 1. Figure 1: Axolotl3D enables controllable, occlusion-aware 3D generation across diverse applications with consistent and faithful shape completion. Abstract. Recent 3D generative models produce high-quality geometry from a single image using large-scale priors and diffusion architectures. However, they assume complete visibility and single-view inputs, limiting applicability in multi-view, occluded, or editing scenar… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Axolotl3D. Given multi-modal inputs—posed images, visibil￾ity masks, and partial points—our method encodes each modality and fuses them into multi-modal condition tokens. We fine-tune Hunyuan3D-DiT on these tokens to pro￾duce completed shape latents, which are decoded by the ShapeVAE. 3.1 Model Architecture Axolotl3D is built upon recent state-of-the-art image-to-3D diffusion model, Hunyuan3D 2… view at source ↗
Figure 3
Figure 3. Figure 3: Data Augmentation: Examples of our on-the-fly data augmentation. a dense surface point cloud P k surface (see Supplement). We also sample 150 cam￾eras C k on a sphere and render corresponding images I k following [52]. During training, we sample and augment this data to produce input partial point cloud P and multi-view {Ii} N i=1, {Mi} N i=1, {Ci} N i=1, N = 6, according to the follow￾ing scenarios chosen… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Comparisons: Single- and multi-view completion results against baselines. Occluders are shown in grey in the input view. Zoom in for details [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Predicted Depth Comparisons: Unlike SAM3D, our method prioritizes adherence to input points, yielding more faithful completion, but degrades under severe depth prediction errors—a performance trade off. Zoom in for details. Effect of Masked Attention: To assess the contribution of visibility-aware attention, we remove the attention bias from image masks applied to cross￾attention layers. Removing the maski… view at source ↗
Figure 6
Figure 6. Figure 6: Completion Stress Test: Model performance on challenging single- and multi-view examples with increasing amounts of occluders (3, 5, 7). Zoom in for details [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation Examples: Comparing completion results for the full model and ablated variants. Occluders are shown in grey in the input view. Zoom in for details. To provide conditioning points, we sample the surface of the original shape and discard points that project into the masked region of the edited view. Fig.8a shows effective 3D edits that remain faithful to the unedited parts of the shape. Image to 3D:… view at source ↗
Figure 8
Figure 8. Figure 8: Applications: Axolotl3D enables precise object edits and accurate single-view generation with completion of missing regions. Zoom in for details. aligned and normalized, before being passed to our method (see Supplement), resulting in faithful 3D generations from a single view (Fig.8b). Simulation: Accurate and faithful completion is also important for physical applications, such as physics simulation of s… view at source ↗
Figure 9
Figure 9. Figure 9: Training Data Filtering: Example of objects that are removed [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: 3D-Consistent Masks: Example of masks with generated occluders. address this, we leverage the G-Objaverse [62] dataset, which provides human￾provided category labels for the original Objaverse dataset. It defines 10 gen￾eral categories, including Human-Shape (41,557), Animals (28,882), Daily-Used (220,222), Furnitures (19,284), Buildings & Outdoor (116,545), Transportations (20,075), Plants (7,195), Food … view at source ↗
Figure 11
Figure 11. Figure 11: Image-to-3D Pipeline: Fully automated with SAM 2 [37] segmentation, MapAnything [22] points and cameras, and Orient Anything v2 [47] point alignment. where F_m denotes the fraction of foreground pixels that are occluded, F_c is the fraction of image pixels occupied by the object, and \protect \text {R} represents the image resolution. For unoccluded scenarios, the view with the largest object coverage, i.… view at source ↗
Figure 12
Figure 12. Figure 12: Object edit comparison against baselines with 2048 input points. – Surface quality and fine details: While our approach produces high￾quality reconstructions overall, extremely thin structures can sometimes be disconnected and fine surface details may be smoothed or thicker than in re￾ality. Leveraging more powerful base models or image encoders could enable better reconstruction of finer geometry. – Robu… view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative Comparisons on OmniObject3D: Single- and multi-view completion results against baselines with occlusions [PITH_FULL_IMAGE:figures/full_fig_p030_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative Comparisons on OmniObject3D: Single- and multi-view completion results against baselines without occlusions [PITH_FULL_IMAGE:figures/full_fig_p031_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative Comparisons on Toys4k: Single- and multi-view completion results against baselines without occlusions [PITH_FULL_IMAGE:figures/full_fig_p032_15.png] view at source ↗

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

62 extracted references · 2 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Barda, A., Gadelha, M., Kim, V.G., Aigerman, N., Bermano, A.H., Groueix, T.: Instant3dit: Multiview inpainting for fast editing of 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16273–16282 (June 2025)

    2025

  2. [2]

    In: Sensor fusion IV: control paradigms and data structures

    Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586–606. Spie (1992)

    1992

  3. [3]

    arXiv preprint arXiv:1512.03012 (2015)

    Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)

    Pith/arXiv arXiv 2015

  4. [4]

    In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum?id=z0QLeooEEf

    Chang, J., Ye, C., Wu, Y., Chen, Y., Zhang, Y., Luo, Z., Li, C., Zhi, Y., Han, X.: Reconviagen: Towards accurate multi-view 3d object reconstruction via generation. In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum?id=z0QLeooEEf

    2026

  5. [5]

    arXiv preprint arXiv:2511.16624 (2025)

    Chen, X., Chu, F.J., Gleize, P., Liang, K.J., Sax, A., Tang, H., Wang, W., Guo, M., Hardin, T., Li, X., et al.: Sam 3d: 3dfy anything in images. arXiv preprint arXiv:2511.16624 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Vicente, T.F.Y., Dideriksen, T., Arora, H., Guillaumin, M., Malik, J.: Abo: Dataset and benchmarks for real-world 3d object understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 21126–21136 (June 2022)

    2022

  7. [7]

    arXiv preprint arXiv:2510.22975 (2025)

    Dagli, R., Xiang, D., Modi, V., Loop, C., Tsang, C.F., Chen, A.H., Hu, A., State, G., Levin, D.I., Shugrina, M.: Vomp: Predicting volumetric mechanical property fields. arXiv preprint arXiv:2510.22975 (2025)

    arXiv 2025

  8. [8]

    Advances in Neural Information Processing Systems36, 35799–35813 (2023)

    Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of 10m+ 3d objects. Advances in Neural Information Processing Systems36, 35799–35813 (2023)

    2023

  9. [9]

    IEEE Transactions on Pattern Analysis and Machine Intelligence40(3), 611–625 (2018)

    Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence40(3), 611–625 (2018). https://doi.org/ 10.1109/TPAMI.2017.2658577

    arXiv 2018

  10. [10]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)

    Erkoç, Z., Gümeli, C., Wang, C., Nießner, M., Dai, A., Wonka, P., Lee, H.Y., Zhuang, P.: Preditor3d: Fast and precise 3d shape editing. In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR). pp. 640–649 (June 2025)

    2025

  11. [11]

    arXiv preprint arXiv:2508.18271 (2025)

    Feng, H., Liu, J., Tang, J., Wu, G., Chen, B., Lai, J., Wang, G.: Objfiller-3d: Consistent multi-view 3d inpainting via video diffusion models. arXiv preprint arXiv:2508.18271 (2025)

    Pith/arXiv arXiv 2025

  12. [12]

    International Journal of Computer Vision129(12), 3313–3337 (2021)

    Fu, H., Jia, R., Gao, L., Gong, M., Zhao, B., Maybank, S., Tao, D.: 3d-future: 3d furniture shape with texture. International Journal of Computer Vision129(12), 3313–3337 (2021). https://doi.org/10.1007/s11263-021-01534-z

    work page doi:10.1007/s11263-021-01534-z 2021

  13. [13]

    arXiv preprint arXiv:2602.10173 (2026)

    Fuji Tsang, C., Hu, A., Perel, O., Kolve, C., Shugrina, M.: Artisangs: Interactive tools for gaussian splat selection with ai and human in the loop. arXiv preprint arXiv:2602.10173 (2026)

    arXiv 2026

  14. [14]

    arXiv preprint arXiv:2407.21783 (2024) Axolotl3D 17

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Let- man, A., Mathur, A., Schelten, A., Vaughan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024) Axolotl3D 17

    Pith/arXiv arXiv 2024

  15. [15]

    In: Proceedings of the IEEE International Conference on Computer Vision (ICCV)

    He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 1026–1034 (2015). https://doi.org/10.1109/ICCV.2015.123

    doi:10.1109/iccv.2015.123 2015

  16. [16]

    In: Advances in Neural Information Processing Systems

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems. vol. 33, pp. 6840– 6851 (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf

    2020

  17. [17]

    In: NeurIPS 2021 Work- shop on Deep Generative Models and Downstream Applications (2021), https: //openreview.net/forum?id=qw8AKxfYbI

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Work- shop on Deep Generative Models and Downstream Applications (2021), https: //openreview.net/forum?id=qw8AKxfYbI

    2021

  18. [18]

    arXiv preprint arXiv:2602.04349 (2026)

    Hsiao, T.F., Ruan, B.K., Liu, Y.L., Shuai, H.H.: Vecset-edit: Unleashing pre- trained lrm for mesh editing from single image. arXiv preprint arXiv:2602.04349 (2026)

    Pith/arXiv arXiv 2026

  19. [19]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)

    Huang, Z., Guo, Y.C., An, X., Yang, Y., Li, Y., Zou, Z.X., Liang, D., Liu, X., Cao, Y.P., Sheng, L.: Midi: Multi-instance diffusion for single image to 3d scene genera- tion. In: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR). pp. 23646–23657 (June 2025)

    2025

  20. [20]

    arXiv preprint arXiv:2506.15442 (2025)

    Hunyuan3D, T., Yang, S., Yang, M., Feng, Y., Huang, X., Zhang, S., He, Z., Luo, D., Liu, H., Zhao, Y., et al.: Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material. arXiv preprint arXiv:2506.15442 (2025)

    Pith/arXiv arXiv 2025

  21. [21]

    arXiv preprint arXiv:2509.21245 (2025)

    Hunyuan3D, T., Zhang, B., Guo, C., Liu, H., Yan, H., Shi, H., Huang, J., Yu, J., Li, K., Wang, P., et al.: Hunyuan3d-omni: A unified framework for controllable generation of 3d assets. arXiv preprint arXiv:2509.21245 (2025)

    arXiv 2025

  22. [22]

    In: International Con- ference on 3D Vision (3DV)

    Keetha, N., Müller, N., Schönberger, J., Porzi, L., Zhang, Y., Fischer, T., Knapitsch, A., Zauss, D., Weber, E., Antunes, N., Luiten, J., Lopez-Antequera, M., Bulò, S.R., Richardt, C., Ramanan, D., Scherer, S., Kontschieder, P.: MapA- nything: Universal feed-forward metric 3D reconstruction. In: International Con- ference on 3D Vision (3DV). IEEE (2026)

    2026

  23. [23]

    ACM Transactions On Graphics (TOG)42(4) (2023)

    Kerbl, B., Kopanas, G., Leimkuehler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions On Graphics (TOG)42(4) (2023). https://doi.org/10.1145/3592433, https://doi.org/10.1145/3592433

    doi:10.1145/3592433 2023

  24. [24]

    ACM Transactions On Graphics (TOG)42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions On Graphics (TOG)42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

    2023

  25. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B., Batra, D., Clegg, A., Undersander, E., Chang, A.X., Savva, M.: Habitat synthetic scenes dataset (hssd- 200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16384–16393 ...

    2024

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Kong,X.,Liu,S.,Lyu,X.,Taher,M.,Qi,X.,Davison,A.J.:Eschernet:Agenerative model for scalable view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9503–9513 (June 2024)

    2024

  27. [27]

    In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum?id=yirunib8l8

    Lin, H., Chen, S., Liew, J.H., Chen, D.Y., Li, Z., Zhao, Y., Peng, S., Guo, H., Zhou, X., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. In: The Fourteenth International Conference on Learning Representations (2026), https://openreview.net/forum?id=yirunib8l8

    2026

  28. [28]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi- view generation and 3d diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10072–10083 (June 2024) 18 A. Hu and M. Shugrina

    2024

  29. [29]

    In: International Conference on Learning Representations (ICLR) (2019), https://openreview.net/ forum?id=Bkg6RiCqY7

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (ICLR) (2019), https://openreview.net/ forum?id=Bkg6RiCqY7

    2019

  30. [30]

    In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Labs

    Modi, V., Fuji Tsang, C., Shugrina, M., Daviet, G.: Tackling gaussian splats, physics simulation and visualization with nvidia kaolin and warp libraries. In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Labs. SIGGRAPH Labs ’25, Association for Computing Machinery,NewYork,NY,USA(2025).https://doi.org/...

    arXiv 2025

  31. [31]

    ACM Transactions on Graphics (TOG) 43(4), 1–11 (2024)

    Modi, V., Sharp, N., Perel, O., Sueda, S., Levin, D.I.: Simplicits: Mesh-free, geometry-agnostic elastic simulation. ACM Transactions on Graphics (TOG) 43(4), 1–11 (2024)

    2024

  32. [32]

    ACM Transactions on Graphics (TOG)43(6), 1–19 (2024)

    Moenne-Loccoz, N., Mirzaei, A., Perel, O., de Lutio, R., Martinez Esturo, J., State, G., Fidler, S., Sharp, N., Gojcic, Z.: 3d gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG)43(6), 1–19 (2024)

    2024

  33. [33]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    Pith/arXiv arXiv 2023

  34. [34]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Ozguroglu, E., Liu, R., Surís, D., Chen, D., Dave, A., Tokmakov, P., Vondrick, C.: pix2gestalt: Amodal segmentation by synthesizing wholes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3931–3940 (June 2024)

    2024

  35. [35]

    In: Advances in Neural Information Processing Systems

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learn- ing on point sets in a metric space. In: Advances in Neural Information Processing Systems. vol. 30 (2017), https://proceedings.neurips.cc/paper_files/paper/2017/ file/d8bf84be3800d12f74d8b05e9b89836f-Paper.pdf

    2017

  36. [36]

    arXiv preprint arXiv:2506.21544 (2025)

    Qu, Y., Dai, S., Li, X., Wang, Y., Shen, Y., Cao, L., Ji, R.: Deocc-1-to-3: 3d de-occlusion from a single image via self-supervised multi-view diffusion. arXiv preprint arXiv:2506.21544 (2025)

    Pith/arXiv arXiv 2025

  37. [37]

    In: The Thirteenth International Conference on Learning Representations (2025), https://openreview.net/forum?id=Ha6RTeWMd0

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollar, P., Feichtenhofer, C.: SAM 2: Segment anything in images and videos. In: The Thirteenth International Conference on Learning Representations (2025), https://openreview.n...

    2025

  38. [38]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684– 10695 (June 2022)

    2022

  39. [39]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

    2016

  40. [40]

    arXiv preprint arXiv:2601.11514 (2026)

    Siddiqui, Y., Frost, D., Aroudj, S., Avetisyan, A., Howard-Jenkins, H., DeTone, D., Moulon, P., Wu, Q., Li, Z., Straub, J., Newcombe, R., Engel, J.: Shaper: Robust conditional 3d shape generation from casual captures. arXiv preprint arXiv:2601.11514 (2026)

    arXiv 2026

  41. [41]

    In: Advances in Neural Information Processing Systems

    Sitzmann, V., Rezchikov, S., Freeman, B., Tenenbaum, J., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. In: Advances in Neural Information Processing Systems. vol. 34, pp. 19313– 19325 (2021), https://proceedings.neurips.cc/paper_files/paper/2021/file/ a11ce019e96a4c60832eadd755a17a58-Paper.pdf Axolotl3D 19

    2021

  42. [42]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Stojanov, S., Thai, A., Rehg, J.M.: Using shape to categorize: Low-shot learn- ing with an explicit shape bias. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1798–1808 (June 2021)

    2021

  43. [43]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 2149–2159 (January 2022)

    2022

  44. [44]

    In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR)

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR). pp. 5294–5306 (June 2025)

    2025

  45. [45]

    arXiv preprint arXiv:2106.10689 (2023)

    Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689 (2023)

    Pith/arXiv arXiv 2023

  46. [46]

    In: The Fourteenth International Conference on Learning Representations (2026), https: //openreview.net/forum?id=DTQIjngDta

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.: $\pi^3$: Permutation-equivariant visual geometry learning. In: The Fourteenth International Conference on Learning Representations (2026), https: //openreview.net/forum?id=DTQIjngDta

    2026

  47. [47]

    In: The Thirty- ninth Annual Conference on Neural Information Processing Systems (2026), https: //openreview.net/forum?id=n3armuTFit

    Wang, Z., Zhang, Z., Xu, J., Wang, J., Pang, T., Du, C., Zhao, H., Zhao, Z.: Orient anything v2: Unifying orientation and rotation understanding. In: The Thirty- ninth Annual Conference on Neural Information Processing Systems (2026), https: //openreview.net/forum?id=n3armuTFit

    2026

  48. [48]

    Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    Wu, Q., Martinez Esturo, J., Mirzaei, A., Moenne-Loccoz, N., Gojcic, Z.: 3dgut: Enabling distorted cameras and secondary rays in gaussian splatting. Conference on Computer Vision and Pattern Recognition (CVPR) (2025)

    2025

  49. [49]

    In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision

    Wu, T., Zheng, C., Guan, F., Vedaldi, A., Cham, T.J.: Amodal3r: Amodal 3d reconstruction from occluded 2d images. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision. pp. 9181–9193 (2025)

    2025

  50. [50]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 803–814 (June 2023)

    2023

  51. [51]

    Tech report (2025)

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

    2025

  52. [52]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 21469–21480 (2025)

    2025

  53. [53]

    In: Proceedings of the European Conference on Computer Vision (ECCV)

    Xu, C., Li, A., Chen, L., Liu, Y., Shi, R., Su, H., Liu, M.: Sparp: Fast 3d ob- ject reconstruction and pose estimation from sparse views. In: Proceedings of the European Conference on Computer Vision (ECCV). p. 143–163 (2024)

    2024

  54. [54]

    arXiv preprint arXiv:2506.21076 (2025)

    Yan, H., Luo, K., Li, W., Liang, Y., Li, S., Huang, J., Guo, C., Tan, P.: Posemaster: Generating 3d characters in arbitrary poses from a single image. arXiv preprint arXiv:2506.21076 (2025)

    Pith/arXiv arXiv 2025

  55. [55]

    ACM Transactions On Graphics (TOG)44(4) (2025)

    Yao, K., Zhang, L., Yan, X., Zeng, Y., Zhang, Q., Xu, L., Yang, W., Gu, J., Yu, J.: Cast: Component-aligned 3d scene reconstruction from an rgb image. ACM Transactions On Graphics (TOG)44(4) (2025). https://doi.org/10.1145/3730841, https://doi.org/10.1145/3730841 20 A. Hu and M. Shugrina

    doi:10.1145/3730841 2025

  56. [56]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., Chen, G., Cui, S., Han, X.: Mvimgnet: A large-scale dataset of multi- view images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9150–9161 (June 2023)

    2023

  57. [57]

    https://github.com/1zb/VecSetX (2025)

    Zhang, B.: Vecsetx. https://github.com/1zb/VecSetX (2025)

    2025

  58. [58]

    ACM Transactions On Graphics (TOG)42(4), 1–16 (2023)

    Zhang, B., Tang, J., Niessner, M., Wonka, P.: 3dshape2vecset: A 3d shape repre- sentation for neural fields and generative diffusion models. ACM Transactions On Graphics (TOG)42(4), 1–16 (2023)

    2023

  59. [59]

    ACM Transactions On Graphics (TOG)43(4) (2024)

    Zhang, L., Wang, Z., Zhang, Q., Qiu, Q., Pang, A., Jiang, H., Yang, W., Xu, L., Yu, J.: Clay: A controllable large-scale generative model for creating high- quality 3d assets. ACM Transactions On Graphics (TOG)43(4) (2024). https: //doi.org/10.1145/3658146

    doi:10.1145/3658146 2024

  60. [60]

    arXiv preprint arXiv:2507.07410 (2025)

    Zhang,X.,Irshad,M.Z.,Yezzi,A.,Tsai,Y.C.,Kira,Z.:Eschernet++:Simultaneous amodal completion and scalable view synthesis through masked fine-tuning and enhanced feed-forward 3d reconstruction. arXiv preprint arXiv:2507.07410 (2025)

    Pith/arXiv arXiv 2025

  61. [61]

    arXiv preprint arXiv:2511.21945 (2025)

    Zhou, J., Tai, Y.W.: Amodalgen3d: Generative amodal 3d object reconstruction from sparse unposed views. arXiv preprint arXiv:2511.21945 (2025)

    Pith/arXiv arXiv 2025

  62. [62]

    In: Proceedings of the European Conference on Computer Vision (ECCV)

    Zuo, Q., Gu, X., Dong, Y., Zhao, Z., Yuan, W., Qiu, L., Bo, L., Dong, Z.: High- fidelity 3d textured shapes generation by sparse encoding and adversarial decoding. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 52– 69 (2024). https://doi.org/10.1007/978-3-031-72684-2_4 Supplementary Material Contents A Model Architecture ..........

    work page doi:10.1007/978-3-031-72684-2_4 2024

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.