Pith. sign in

REVIEW 4 major objections 5 minor 45 references

MV-RAG: Retrieval Augmented Multiview Diffusion

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Retrieved 2D photos make text-to-3D models faithful on rare objects.

desk verdict A promising RAG-for-3D idea and a useful OOD benchmark, but the 2D held-out objective likely doesn't teach 3D consistency, and the evaluation risks circularity. read the letter →

arxiv 2508.16577 v1 pith:OG3EDDZW submitted 2025-08-22 cs.CV cs.AI

classification cs.CVcs.AI
keywords retrieval-augmentedgenerationmultiviewdiffusiontext-to-3Dout-of-domainrareconcepts3Dconsistencyimageconditioningmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that text-to-3D models fail on rare or out-of-domain concepts because the underlying 2D prior lacks visual knowledge of these concepts, and 3D fine-tuning data cannot fully fill the gap. MV-RAG addresses this by retrieving real-world 2D images for the prompt from a large corpus and conditioning a multiview diffusion model on those images during generation. The core training idea is hybrid: on 3D data the model reconstructs target views from heavily augmented 'retrieval-like' views, and on 2D data it must predict a held-out image from K retrieved neighbors, forcing it to reason about 3D consistency from unposed, unstructured photos. The paper introduces OOD-Eval, a curated set of 196 challenging prompts, and reports that MV-RAG outperforms text-to-3D, image-to-3D, and personalization baselines on OOD consistency, photorealism, and prompt adherence while staying competitive on standard benchmarks.

What carries the argument

The central object is a retrieval-conditioned multiview diffusion model: a multiview diffusion backbone whose denoising network receives, in addition to text and camera poses, visual tokens encoded from K retrieved images via a resampler, integrated through decoupled cross-attention alongside the base text attention. Carry of the argument is split across two training modes. 3D mode renders ground-truth multiviews and uses heavily augmented extra views as simulated retrievals, teaching view-specific reconstruction under retrieval variance. 2D mode takes K retrieved real-world images and trains the model to predict the held-out K+1 image without camera poses, using 2D self-attention rather tha

What would settle it

Take an OOD prompt with retrieved images that are visually similar but show inconsistent viewpoints and lighting, generate MV-RAG's multiviews, and reconstruct a 3D mesh from them. If reconstruction produces high reprojection error or inconsistent geometry even though each generated view closely matches one retrieved image, then the held-out objective delivered 2D copying rather than 3D consistency.

Watch

Extended reading notes

Core claim

MV-RAG's central claim is that a multiview diffusion model can become robust to out-of-domain prompts by conditioning on retrieved 2D images, provided it is trained with a hybrid objective that simulates retrieval variance on 3D data and uses a held-out view prediction objective on 2D data. The model injects retrieved image features through a resampler into decoupled cross-attention layers, and a prior-guided attention mechanism dynamically weights the base model's prior against the retrieved signals according to how out-of-distribution the prompt appears. This yields consistent and accurate multiview outputs for rare concepts without per-subject fine-tuning, unlike retrieval of 3D assets, w

Load-bearing premise

The load-bearing premise is that training the model to predict a held-out retrieved image from its neighbors teaches genuine 3D consistency that transfers to novel OOD objects, rather than teaching it to copy or interpolate the retrieved images; if that premise fails, the OOD gains disappear and the model is just an image-conditioned generator with no real 3D reasoning.

Editorial extensions

If this is right

  • Rare and newly emerging objects become generatable in 3D with a single forward pass, without per-subject fine-tuning or optimization.
  • Because retrievals come from web-scale 2D corpora, the approach sidesteps the coverage limits of 3D retrieval databases.
  • The held-out view prediction objective can in principle extract 3D-aware supervision from any large 2D image collection, reducing dependence on scarce 3D training data.
  • The dynamic fusion mechanism gives a general recipe for combining a strong base prior with external visual evidence for unseen concepts.
  • OOD-Eval provides a standardized set of hard prompts, enabling direct comparison of future text-to-3D methods on out-of-domain concepts.

Reading between the lines

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

  • If the held-out view prediction genuinely induces implicit 3D reasoning, the same objective could improve other feed-forward multiview and novel-view-synthesis models trained on unposed photo sets.
  • The retrieval stage could be closed-loop at inference: use the initially generated multiviews to refine the retrieval set, potentially improving consistency on very rare concepts; the paper does not explore this.
  • The adaptive fusion weight, currently based on similarity between initial views and retrieved images, could be replaced by a learned 3D-aware out-of-distribution scorer; the paper's own limitation section hints at this.
  • A testable extension is to vary K and retrieval diversity systematically, measuring how much of the consistency gain comes from retrieved geometry versus the base prior's smoothing.
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 / 5 minor

Summary. The paper proposes MV-RAG, a retrieval-augmented multiview diffusion model for text-to-3D generation. Given a text prompt, the method retrieves 2D images from a large in-the-wild database and conditions a multiview diffusion model on these images through attention adapters. Training uses a hybrid strategy: (i) a 3D mode that renders ground-truth multiviews and simulates retrieval variance via augmentation, and (ii) a 2D mode that retrieves K+1 semantically similar images and trains the model to predict the held-out image from the other K. An adaptive fusion mechanism weights the base model prior against retrieved-image signals. The authors introduce OOD-Eval, a 196-prompt benchmark for OOD/rare concepts, and report improvements over text-to-3D, image-to-3D, and personalization baselines in 3D consistency, photorealism, and text adherence, with competitive in-domain performance. The provided manuscript, however, omits key sections (3.2–3.5 and 4), including the training objectives, architecture details, and full experimental results.

Significance. If the central claims hold, the work would be a valuable contribution: it addresses a known failure mode of multiview diffusion models on OOD/rare concepts by leveraging abundant 2D data through retrieval, without per-subject fine-tuning. The proposed OOD-Eval benchmark and the adaptive fusion idea are potentially useful to the community. The approach is timely and the problem is important. However, the current manuscript does not supply enough technical detail to verify the mechanism, and the ill-posedness of the 2D-mode objective and the potential circularity in evaluation are serious concerns that must be resolved before the claims can be accepted.

major comments (4)
  1. [Sec. 3.1, Fig. 3] The 2D Data Mode objective is not a view-prediction task in any 3D sense. The K+1 images are retrieved for semantic similarity to the text; nothing ensures they are different views of the same object instance. No camera poses are provided in this mode, and the model uses 2D self-attention. The held-out image is therefore not a held-out view: the model can satisfy the objective by generating an image consistent with the appearance/text distribution of the retrieved set without any geometric inference. This undermines the paper's claim that the 2D objective teaches 3D consistency. The limitations section (Sec. F) acknowledges retrieval quality but does not address this ill-posedness. Since the hybrid training scheme is the central novelty, this needs to be either fixed or substantially re-argued.
  2. [Sec. E, Sec. C.4] The evaluation protocol appears to use the same retrieved images both as conditioning for MV-RAG and as references for image-image similarity metrics (CLIP, DINOv2, IR). If so, high similarity scores may reflect copying or interpolating the conditioning images rather than improved 3D generalization. The text does not specify whether the evaluation references are disjoint from the conditioning set. This is load-bearing for the OOD claims. The authors must clarify the protocol and, ideally, evaluate on held-out references not used as conditioning, or otherwise demonstrate that the similarity gain is not an artifact of conditioning on the evaluation targets.
  3. [Sections 3.2–3.5 and 4] The provided manuscript omits the core technical content: the training objectives (loss formulations), the architecture of the Resampler and retrieval-attention modules, the adaptive fusion mechanism (how alpha is computed), the retrieval process described in Sec. 3.5, and all experimental results, tables, and ablations. Without these, the abstract's claims cannot be independently checked. This is not a minor presentation issue; it prevents a soundness assessment of the central contribution.
  4. [Sec. F, adaptive fusion] The adaptive fusion mechanism is described only at a high level. The limitation section shows that when the base model is assigned high alpha, 3D structural errors are inherited, but no analysis or ablation of alpha's sensitivity is present in the provided text. Since the claimed benefit is the adaptive balancing of prior and retrieval, the paper should provide evidence—e.g., ablations across alpha values or OOD scores—that the mechanism behaves as intended and does not silently regress to the base model or to copy behavior.
minor comments (5)
  1. [Sec. 3.1, Fig. 2] Notation is inconsistent: the text uses K and K+1, while Fig. 2's caption says 'k relevant images'. Define K consistently.
  2. [References] The reference list is incomplete: entries [41]–[61] are missing. In particular, [50], cited for the Instance Retrieval model, is absent. This needs to be fixed.
  3. [Sec. E] The text refers to 'Sec.3 in main paper' for retrieval details, but the main paper's Sec. 3 is truncated. Ensure cross-references point to the actual equations/sections.
  4. [Sec. C.4, Fig. 8] Fig. 8 illustrates 'limitations of CLIP text-image similarity', but the metrics section then motivates image-image similarity. Clarify whether the final OOD-Eval metric is image-image or text-image, and why the CLIP text-image result is presented.
  5. [Title page] The manuscript header includes 'Preprint. Under review.' which is not appropriate for a journal submission; remove.

Circularity Check

1 steps flagged · score 5.0 of 10

2D-mode 'held-out view' objective is a semantic image-completion task, not a 3D view-prediction task; the OOD claim is partially definitional.

  1. fitted input called prediction [Section 1 (Introduction), paragraph describing 2D supervision; restated in Section 3.1 '2D Data Mode Supervision']
    "To incorporate supervision from 2D data, we utilize K +1 images that are semantically similar to the input text from a given 2D text-image dataset. We then present K of these images as conditional inputs, and our model is trained to generate the held-out image. This novel objective pushes our model to infer 3D relationships and consistent appearances directly from sets of unstructured, real-world 2D views."

    The K+1 images are selected only by semantic similarity to the prompt, not as registered camera views of one object (no poses, no object-identity constraint). The 2D-mode loss is therefore a retrieval-cluster completion task: generate a held-out member from the other K members. A model can lower this loss by copying/interpolating the conditioning images' appearance; no 3D geometry is required. Yet the paper names this 'held-out view prediction' and concludes that it 'infers 3D relationships ... from 2D views' — importing the 3D content by definition rather than deriving it from the data. If the OOD-Eval image-image metrics compare generated views to the same retrieved/reference images, the reported OOD '3D consistency/photorealism' gains are essentially the training loss evaluated at infer

full rationale

The paper's central novelty is the hybrid training scheme. The 3D mode is self-contained: it renders ground-truth views from 3D objects, applies augmentations to simulate retrieval variance, and trains with explicit camera poses, so that part does not reduce to its inputs. The problematic step is the 2D mode, where 'held-out view prediction' is defined over K+1 text-semantically-similar images rather than over distinct views of a single 3D object. The training target is thus a held-out image from a semantic retrieval cluster, and the paper labels this as learning 3D consistency. This is a definitional/fitted-input issue: the objective can be satisfied without geometric inference, and any OOD evaluation that uses image-image similarity to the same retrieved/reference images would be measuring the training loss. However, the paper also provides independent evidence: standard-benchmark reconstruction against ground-truth Objaverse views, ablations, and a human study asking about cross-view consistency. Those parts are not circular. There are no load-bearing self-citations; MVDream and Objaverse are external. Weighing the partially circular 2D-mode claim against the independent 3D-mode and human-evaluation support, a moderate score of 5 is appropriate.

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

The approach depends on the availability of relevant images in the retrieval corpus and on the assumption that training on unposed 2D image sets transfers to 3D-consistent generation. The counts above capture the main tunable choices and background assumptions.

free parameters (4)
  • K (number of retrieved conditioning views)
    Not stated in the visible text; determines how many images condition each generated view and how many are used in the 2D held-out training mode.
  • alpha (α) adaptive fusion weight
    Controls the balance between the base MVDream prior and retrieval features; authors note artifacts when α is high (Appendix F).
  • Augmentation strength for simulated retrievals in 3D mode
    Heavy augmentations of ground-truth multiviews are used to mimic retrieval variance; the choice likely affects the 3D-to-2D domain gap.
  • Retrieval relevance threshold / score cutoff
    The method's dependence on retrieval quality is acknowledged; the cutoff for accepting retrieved images is a design choice.
assumptions (5)
  • domain assumption Pretrained MVDream multiview diffusion model is a suitable base for adding retrieval conditioning
    The method fine-tunes rather than builds from scratch; if the base model's prior is too weak, the adaptive fusion cannot compensate.
  • domain assumption LAION-400M (or similar) contains sufficiently many and relevant images for rare concepts, and BM25/CLIP retrieval finds them
    The whole conditioning mechanism relies on retrieval recall for OOD prompts.
  • ad hoc to paper The held-out view prediction objective from K unposed images teaches 3D consistency
    This is the core novelty; the paper provides no proof that this objective transfers to novel view generation beyond empirical results.
  • ad hoc to paper Augmenting ground-truth multiviews approximates the distribution of retrieved real-world images
    Used to bridge 3D and 2D training modes; if the augmentation does not match real retrieval noise, the 3D mode may not transfer.
  • domain assumption OOD-Eval prompts are genuinely out-of-distribution and the chosen image-image metrics are valid proxies for 3D consistency and fidelity
    The benchmark is built with GPT-4o and not validated against the base model's knowledge; CLIP text-image similarity is rejected in favor of image-image metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MV-RAG: Retrieval Augmented Multiview Diffusion." pith.science (2026). https://pith.science/paper/OG3EDDZW

@misc{pith2026250816577,
  author       = {Pith},
  title        = {Pith review of: MV-RAG: Retrieval Augmented Multiview Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OG3EDDZW}},
  note         = {Machine review of arXiv:2508.16577}
}
read the original abstract

Text-to-3D generation approaches have advanced significantly by leveraging pretrained 2D diffusion priors, producing high-quality and 3D-consistent outputs. However, they often fail to produce out-of-domain (OOD) or rare concepts, yielding inconsistent or inaccurate results. To this end, we propose MV-RAG, a novel text-to-3D pipeline that first retrieves relevant 2D images from a large in-the-wild 2D database and then conditions a multiview diffusion model on these images to synthesize consistent and accurate multiview outputs. Training such a retrieval-conditioned model is achieved via a novel hybrid strategy bridging structured multiview data and diverse 2D image collections. This involves training on multiview data using augmented conditioning views that simulate retrieval variance for view-specific reconstruction, alongside training on sets of retrieved real-world 2D images using a distinctive held-out view prediction objective: the model predicts the held-out view from the other views to infer 3D consistency from 2D data. To facilitate a rigorous OOD evaluation, we introduce a new collection of challenging OOD prompts. Experiments against state-of-the-art text-to-3D, image-to-3D, and personalization baselines show that our approach significantly improves 3D consistency, photorealism, and text adherence for OOD/rare concepts, while maintaining competitive performance on standard benchmarks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 32 canonical work pages

  1. [1]

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023

  2. [2]

    Barratt and R

    S. Barratt and R. Sharma. A note on the inception score. arXiv preprint arXiv:1801.01973, 2018

  3. [3]

    Blattmann, R

    A. Blattmann, R. Rombach, H. Ling, T. Dockhorn, S. W. Kim, S. Fidler, and B. Ommer. Retrieval- augmented diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), 2022

  4. [4]

    Borgeaud, A

    S. Borgeaud, A. Mensch, J. Hoffmann, T. Cai, E. Rutherford, K. Millican, G. van den Driessche, J.-B. Lespiau, B. Damoc, A. Doucet, M. Bärtschi, A. Méray, J. Roffi, A. Glaese, J. W. Noland, A. Cassirer, A. Clark, L. Guy, D. Budden, T. Hennigan, S. Osindero, L. Rimell, M. Tsimpoukelli, K. Simonyan, L. Sifre, S. Dieleman, and N. De Freitas. Improving languag...

  5. [5]

    C. Chen, X. Yang, F. Yang, C. Feng, Z. Fu, C.-S. Foo, G. Lin, and F. Liu. Sculpt3d: Multi-view consistent text-to-3d generation with sparse 3d prior. arXiv preprint arXiv:2403.09140, 2024

  6. [6]

    W. Chen, H. Hu, C. Saharia, and W. W. Cohen. Re-Imagen: Retrieval-Augmented Text-to-Image Generator. arXiv preprint arXiv:2209.14491, 2022

  7. [7]

    Deitke, R

    M. Deitke, R. Liu, M. Wallingford, H. Ngo, O. Michel, A. Kusupati, A. Fan, C. Laforte, V . V oleti, S. Y . Gadre, E. VanderBilt, A. Kembhavi, C. V ondrick, G. Gkioxari, K. Ehsani, L. Schmidt, and A. Farhadi. Objaverse-xl: A universe of 10m+ 3d objects, 2023

  8. [8]

    Deitke, D

    M. Deitke, D. Schwenk, J. Salvador, L. Weihs, O. Michel, E. VanderBilt, L. Schmidt, K. Ehsani, A. Kemb- havi, and A. Farhadi. Objaverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13142–13153, 2023

Show all 45 references
  1. [9]

    Douze, A

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.-E. Mazaré, M. Lomeli, L. Hosseini, and H. Jégou. The faiss library, 2025

  2. [10]

    R. Gal, Y . Alaluf, Y . Atzmon, O. Patashnik, A. H. Bermano, G. Chechik, and D. Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In International Conference on Learning Representations (ICLR), 2023

  3. [11]

    J. Gregory. Game Engine Architecture. A K Peters/CRC Press, third edition, 2018

  4. [12]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018

  5. [13]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 6840–6851, 2020

  6. [14]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models, 2020

  7. [15]

    S. Hong, W. Jang, I. H. Kim, I. Kim, and S. Kim. Variational score distillation for text-to-3d generation. arXiv preprint arXiv:2312.09334, 2023

  8. [16]

    Huang, Y .-C

    Z. Huang, Y .-C. Guo, H. Wang, R. Yi, L. Ma, Y .-P. Cao, and L. Sheng. MV-Adapter: Multi-View Consistent Image Generation Made Easy. arXiv preprint arXiv:2412.03632, 2024

  9. [17]

    Jaegle, F

    A. Jaegle, F. Gimeno, A. Brock, O. Vinyals, A. Zisserman, and J. Carreira. Perceiver: General perception with iterative attention. In Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139, pages 4651–4664. PMLR, 2021

  10. [18]

    Y . Kant, Z. Wu, M. Vasilkovsky, G. Qian, J. Ren, R. A. Guler, B. Ghanem, S. Tulyakov, I. Gilitschenski, and A. Siarohin. Spad : Spatially aware multiview diffusers, 2024

  11. [19]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering. In ACM SIGGRAPH 2023 Conference Proceedings, 2023

  12. [20]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Dollár, and R. Girshick. Segment anything.arXiv:2304.02643, 2023

  13. [21]

    Lasseter

    J. Lasseter. Principles of traditional animation applied to 3d computer animation. ACM SIGGRAPH Computer Graphics, 21(4):35–44, 1987. 12

  14. [22]

    J. P. Lewis, M. Jacobson, A. Witkin, and M. Cohen. Real-time rendering. ACM SIGGRAPH Course Notes, 2002

  15. [23]

    Lewis, E

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, and D. Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems (NeurIPS), volum...

  16. [24]

    P. Li, Y . Liu, X. Long, F. Zhang, C. Lin, M. Li, X. Qi, S. Zhang, W. Luo, P. Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024

  17. [25]

    Liang, X

    Y . Liang, X. Sun, Z. Lai, Z. Zhang, J. Wang, and J. Hu. LucidDreamer: Towards high-fidelity text-to-3d generation via interval score matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20795–20805, 2024

  18. [26]

    C.-H. Lin, J. Gao, L. Tang, T. Takikawa, X. Zeng, X. Huang, K. Kreis, S. Fidler, M.-Y . Liu, and T.-Y . Lin. Magic3D: High-resolution text-to-3d content creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 300–309, 2023

  19. [27]

    J. Lin, X. Ma, S.-C. Lin, J.-H. Yang, R. Pradeep, and R. Nogueira. Pyserini: A Python toolkit for reproducible information retrieval research with sparse and dense representations. In Proceedings of the 44th Annual International ACM SIGIR Conference on Research and Development...

  20. [28]

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740–755....

  21. [29]

    R. Liu, R. Wu, B. V . Hoorick, P. Tokmakov, S. Zakharov, and C. V ondrick. Zero-1-to-3: Zero-shot one image to 3d object. In International Conference on Computer Vision (ICCV), pages 9298–9309, 2023

  22. [30]

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023

  23. [31]

    Y . Liu, C. Lin, Z. Zeng, X. Long, L. Liu, T. Komura, and W. Wang. Syncdreamer: Generating multiview- consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023

  24. [32]

    Long, Y .-C

    X. Long, Y .-C. Guo, C. Lin, Y . Liu, Z. Dou, L. Liu, Y . Ma, S.-H. Zhang, M. Habermann, C. Theobalt, and W. Wang. Wonder3D: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2168...

  25. [33]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization, 2019

  26. [34]

    Mildenhall, P

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

  27. [35]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, R. Howes, P.-Y . Huang, H. Xu, V . Sharma, S.-W. Li, W. Galuba, M. Rabbat, M. Assran, N. Ballas, G. Synnaeve, I. Misra, H. Jegou, J. Mairal, P. Labatut,...

  28. [36]

    R. Parent. Computer Animation: Algorithms and Techniques. Morgan Kaufmann, third edition, 2012

  29. [37]

    Poole, A

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall. DreamFusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023

  30. [38]

    L. Qiu, G. Chen, X. Gu, Q. Zuo, M. Xu, Y . Wu, W. Yuan, Z. Dong, L. Bo, and X. Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024

  31. [39]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Lea...

  32. [40]

    A. Raj, S. Kaza, B. Poole, M. Niemeyer, A. Van Den Oord, S. Fidler, and A. Holynski. Dreambooth3d: Subject-driven text-to-3d generation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 4342–4352, 2023. 13

  33. [62]

    Xiang, Z

    J. Xiang, Z. Lv, S. Xu, Y . Deng, R. Wang, B. Zhang, D. Chen, X. Tong, and J. Yang. Structured 3d latents for scalable and versatile 3d generation, 2025

  34. [63]

    H. Ye, J. Zhang, S. Liu, X. Han, and W. Yang. IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models. arXiv preprint arXiv:2308.06721, 2023

  35. [64]

    T. Yi, J. Zhang, Z. Huang, Y . Liu, G. Chen, J. Zhang, S. Chen, J. Jia, Y . Chen, and G. Wang. Gaussian- dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

  36. [65]

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023

  37. [66]

    Peugeot 202

    S. Zhu, Y . Zhang, X. Tian, and X. Sun. Prompt reverse learning: Enhancing visual language models for rare image recognition. 15 (a) (b) Figure 9: (a). Utility. Our approach learns to utilize all relevant information in retrieved views. On the LHS, we show retrieved views. The...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.