Pith. sign in

REVIEW 4 major objections 6 minor 83 references

VGGT's latent space sits on a product of four zero-mean hyperspheres, and flow matching that respects this geometry lets a frozen 3D foundation model generate plausible novel views from sparse, unposed images.

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

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

T0 review · deepseek-v4-flash

2026-08-01 13:22 UTC pith:ATLAQLY4

load-bearing objection A promising new idea with strong geometry results, but the paper as written mismatches its own training objective and leaves the hypersphere premise unverified. the 4 major comments →

arxiv 2607.19120 v1 pith:ATLAQLY4 submitted 2026-07-21 cs.CV

Latent Riemannian Flow Matching for Geometry-Grounded 3D Foundation Models

Lisa Weijler , Irene Ballester , Guofeng Mei , Tolga Birdal , Pedro Hermosilla This is my paper
classification cs.CV
keywords Riemannian flow matchinghypersphere latent spacegeometric foundation modelsnovel view synthesis3D scene generationlatent space modelingVGGTproduct manifold
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.

This paper tries to establish that a frozen geometric foundation model, a transformer that turns unposed images into 3D geometry, can be made generative without retraining or compressing it, by running flow matching directly inside its latent space. The central claim is that those latent tokens do not live in Euclidean space: after normalization they form a product manifold of four zero-mean hyperspheres, one per encoder block, and Euclidean flow paths cut through low-density off-manifold regions, leading to mode collapse. The authors therefore build a conditional Riemannian flow matching model on that product manifold, conditioned on context views and a target pose, whose generated tokens are decoded by the frozen heads into point maps, depth, and RGB. If right, this shows that the geometry of a latent space, not just its dimension, is the right interface for generating 3D scenes from strong learned priors. The experiments support the claim by improving sparse-view geometry and fused 3D scans over recent baselines while staying competitive on appearance.

Core claim

The paper claims that VGGT's post-LayerNorm encoder features are not Euclidean vectors but elements of a product manifold M = (S^{C-2})^4, with C = 2048, i.e. four zero-mean hyperspheres of radius sqrt(C) embedded in R^8192. It then trains a velocity network on this product manifold using conditional Riemannian flow matching, with geodesic interpolants and parallel-transported target velocities, conditioning on any number of unposed context views and a single relative target pose. Generated tokens are passed to VGGT's frozen DPT heads for geometry and to a separately trained RGB head for appearance. The paper's key evidence is that Euclidean flow matching in the same latent space degrades al

What carries the argument

The central object is the product manifold M = (S^{C-2})^4, four zero-mean hyperspheres of radius sqrt(C), obtained from the LayerNorm normalization applied to the four encoder block outputs of VGGT. Each token in R^8192 is decomposed blockwise, and tangent vectors must be orthogonal to both the all-ones direction (zero mean) and the point itself. The method uses the spherical exponential and logarithm maps, tangent projection, and parallel transport to define geodesic probability paths and regress a conditional velocity field, integrated with a Riemannian Euler solver that keeps samples on M. The same maps act blockwise, so the flow is a product of independent spherical flows tied together

Load-bearing premise

The load-bearing premise is that every post-LayerNorm VGGT token is exactly a point on a product of four zero-mean hyperspheres, and that any point on that product manifold, not just the learned distribution on it, is decodable by VGGT's frozen heads into valid geometry and appearance; if the normalization includes learnable affine parameters, or if natural tokens occupy only a thin submanifold, staying on the spheres is not by itself enough.

What would settle it

Take a large sample of real VGGT tokens and measure per-block norms and means: if norms systematically deviate from sqrt(C) or means from zero, especially under affine LayerNorm parameters, the product-manifold description is only approximate. A second check: uniformly sample many points on the product manifold and feed them to the frozen decoders; if most yield meaningless depth and RGB, the manifold is not by itself the region the heads can decode, and the Riemannian flow's success must come from learning a sub-distribution rather than from manifold geometry alone.

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

If this is right

  • A frozen feed-forward 3D reconstruction model can be turned into a generative model of scenes without changing its weights or building an explicit 3D representation such as Gaussians or meshes.
  • The generative model inherits the foundation model's learned 3D priors, producing coherent geometry in low-overlap regions where deterministic methods leave gaps or misalign, and generalizing to out-of-distribution scenes.
  • Euclidean flow matching in this latent space is a wrong modeling choice; respecting latent manifold structure becomes a quantitative improvement, not a cosmetic one.
  • Because the flow is conditioned only on context views and a target pose, the approach is order-invariant, works from a single unposed view, and drops the trajectory and adjacency constraints of video-diffusion generation.
  • Sampling the learned flow yields a distribution over plausible scenes, opening a route to uncertainty estimation in unobserved regions.

Where Pith is reading between the lines

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

  • If the product-of-spheres structure is generic to LayerNorm-based transformer latents, the same recipe could make other frozen encoders generative; this prediction is not tested in the paper.
  • The sphere identification is exact only if VGGT's LayerNorm lacks learnable affine parameters or if those parameters are folded into downstream weights; otherwise the true decoder-supported region is a shifted or scaled image of the sphere. Even with exact spheres, natural latents occupy a low-dimensional submanifold, so staying on the product manifold alone does not guarantee staying on the data
  • A testable extension: measure whether generated tokens, when decoded by the frozen heads, become less consistent with real VGGT tokens as the number of context views decreases; if so, stronger conditioning or a manifold-aware discriminator could close the residual gap.

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

4 major / 6 minor

Summary. The paper proposes to turn the frozen VGGT geometric foundation model into a generative 3D scene model by performing conditional Riemannian flow matching directly on VGGT's encoder latent space. The central claim is that VGGT's post-LayerNorm tokens lie on a product of four zero-mean hyperspheres S^{C-2} of radius sqrt(C), and that Euclidean flow matching on this space fails because probability paths leave the low-dimensional manifold. The authors introduce a conditional Riemannian flow matching objective (Sec. 3.3), parameterize the velocity field with a ~50M parameter transformer conditioned on context tokens and a target camera pose, and decode the generated tokens with VGGT's frozen DPT heads plus a separately trained RGB head. They evaluate on RealEstate10K, ScanNet++, and ETH3D against Gen3R and DepthSplat, reporting improvements in depth and 3D scan metrics, competitiveness in RGB metrics, and an ablation showing Riemannian flow matching outperforms Euclidean flow matching. An appendix adds implementation details, a careful region-split evaluation protocol, and ablations on prediction target and ODE step count.

Significance. If the geometric identification is correct, the paper makes a valuable contribution: it shows that latent-space flow matching on a geometry foundation model can combine strong learned 3D priors with generative diversity, without an explicit 3D representation or a video-diffusion adapter. The paper is careful in several respects: it provides a parallel-transport derivation of the RCFM target, a detailed common-grid evaluation protocol to make baseline comparisons fair, a principled visible/generated region split, and a stability analysis of the ODE solver. These are genuine strengths. However, the two load-bearing pillars — the exact product-of-hyperspheres identification of VGGT's latent space and the training objective actually used for the final model — are not consistently established. The paper's central theoretical claim is therefore currently unverified, and the empirical conclusions lack statistical support.

major comments (4)
  1. [Sec. 3.2, Eq. (4)] The identification of VGGT's post-LayerNorm tokens with a product of zero-mean hyperspheres assumes that the LayerNorm operation has no affine parameters, or that those parameters are inverted before decoding. Standard LayerNorm computes h = gamma * y + beta, where y is zero-mean and unit-variance; with learned gamma, beta, h is neither zero-mean nor of constant norm, so it lies on a translated, axis-scaled ellipsoid, not on S^{C-2}. The paper neither states that VGGT disables affine in its LayerNorms nor describes an inverse-affine transformation before decoding. The sentence "operating directly on S^{C-2} incurs no loss of information ... allowing us to fully exploit the learned affine parameters of LayerNorm" is internally inconsistent: normalizing to zero mean and unit norm discards the affine parameters. This is load-bearing because the justification for using Riemannian flow matchi
  2. [Sec. 3.3, Eq. (6), vs. Appendix D.1, Table D.1] The main derivation defines the training objective as a squared-norm regression of the network velocity v_w to the conditional RCFM target velocity u_t (Eq. 6). Appendix D.1 states that the final model adopts x-prediction with "v-loss reweighting", a prediction target that is never defined in Section 3.3 or in Appendix A. The reported results therefore correspond to an objective that is not derived anywhere in the paper. This is a central inconsistency: the paper's theory describes one algorithm, and the experiments evaluate another. Please derive the x-prediction objective for the Riemannian setting (or show that Eq. (6) is used with a reparameterized network output), and state clearly which objective is used to produce Tables 1-5.
  3. [Tables 1-5] No error bars, confidence intervals, or multi-seed results are reported for any experiment. Since the method and the main baselines are generative (Gen3R and Ours) or stochastic in data sampling, the reported differences — especially small ones such as the depth RMSE gap in Table D.1 (0.202 vs. 0.205) or the RGB FID differences in Tables 2-3 — may be within sampling noise. Provide means and standard deviations over at least three independent evaluation draws or seeds, or otherwise justify that the differences are statistically significant.
  4. [Abstract, Sec. 3.2, Sec. 5] The claim that Riemannian flow matching "keeps generated tokens on the valid data manifold required by the frozen decoding heads" overstates what the sphere constraint guarantees. Even if the product of spheres were the exact support of the normalized features, natural-scene latents occupy a proper submanifold of that product; staying on the sphere does not by itself ensure that generated tokens lie near the training distribution that the DPT heads expect. The flow-matching training can learn the data distribution on the sphere, but the paper's justification should not conflate the ambient constraint set with the data manifold. Please soften or qualify this claim.
minor comments (6)
  1. [Sec. 3.2, last paragraph] "Eq. (3) and (6)" should be "Eq. (3) and (5)"; Eq. (6) is introduced only in Sec. 3.3.
  2. [Sec. 3.2, Eq. (4)-(5)] Eq. (4) defines S^{C-2} with radius sqrt(C), while Eq. (5) and the projection formula assume a unit-radius sphere. The text says "we rescale each block to unit radius" but the equations should be made consistent, or the rescaling should be made explicit in the definitions.
  3. [Sec. B.2, DepthSplat paragraph] Typo: "sampled in sampled in inverse depth" should be "sampled in inverse depth".
  4. [Sec. 4.1, Table 2] Inconsistent capitalization: "Gen3r" appears in the text while tables and references use "Gen3R".
  5. [Table 5] The Euclidean flow baseline is not described: same architecture, same training steps, same conditioning? Without these details the ablation is difficult to interpret.
  6. [Appendix D.1] The statement that "under the Riemannian formulation, the choice of target becomes largely inconsequential" is contradicted by the fact that the final model adopts x-prediction; if the choice is inconsequential, report both and justify the selection with the reported metrics.

Circularity Check

0 steps flagged

No construction-level circularity; the product-sphere premise is an architectural claim, not an output of the model. The LayerNorm-affine ambiguity is a correctness concern, not a circular reduction.

full rationale

The paper's derivation chain is: (1) characterize VGGT's post-LayerNorm encoder features as a product of four zero-mean hyperspheres; (2) train a conditional Riemannian flow matching model on that product manifold using the external RCFM objective of Chen & Lipman (Eq. 3 / Eq. 6); (3) decode generated tokens with frozen VGGT heads and evaluate against external baselines. None of these steps reduces to its own input by construction. The sphere identification is presented as an architectural fact about VGGT ('Before x is processed by the DPT heads, each h_i passes through a separate LayerNorm... projecting it onto the hypersphere S^{C-2}'), not fitted to the target metric or renamed from the baseline. The claim that Euclidean flow matching fails is supported by an explicit ablation (Table 5), not by definitional equivalence. The RFM objective and geodesic maps are taken from Chen & Lipman, an external and independently established framework. Self-citations appear only as related work (RFMPose, PoseD-Flow, NRDF, DSG, etc.) and do not carry the paper's load-bearing argument. One genuine concern, but not a circularity, is that standard LayerNorm includes learnable affine scale/shift, so the post-LayerNorm representation may be an ellipsoid rather than the sphere; the paper's phrase 'allowing us to fully exploit the learned affine parameters' is ambiguous and the sphere premise is unverified. This is a correctness risk about whether the manifold claim matches VGGT's actual architecture, not an instance of the paper predicting what it fitted or citing itself to force the conclusion. Accordingly, the paper is not significantly circular.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim rests primarily on one domain assumption: that VGGT's latent space is exactly a product of four zero-mean hyperspheres and that every point on that product is decodable. Free parameters are limited to training/inference hyperparameters (time-shift s, RGB loss weights, step count), none of which encode the target result. No new physical or model entities are introduced; the only structure is the existing VGGT latent, re-described as a manifold.

free parameters (3)
  • Time-schedule shift s = 6.0
    Sec 4 Training details. Chosen by hand to concentrate supervision near the noise end; no sensitivity analysis of s is reported, and it affects the training objective of the central method.
  • RGB head loss weights = lambda_pix=1.0, lambda_lpips=0.3, lambda_ssim=0.1
    Appendix B.1. Selected for the auxiliary RGB decoder; affects appearance metrics but not the geometry claim.
  • Inference step count = 20
    Appendix D.2. Chosen as a balance between sharpness and speed from a step-count ablation; used for all reported numbers.
axioms (5)
  • domain assumption The post-LayerNorm VGGT encoder tokens h_i lie exactly on the zero-mean hypersphere S^{C-2} of radius sqrt(C), and the full latent space is their product (K=4).
    Stated in Sec 3.2, eq. (4). Depends on the exact LayerNorm configuration of VGGT; if LayerNorm has learnable affine scale/shift, the output is not zero-mean with fixed norm, so the sphere identification is only approximate.
  • domain assumption Generated tokens on the product manifold M are decodable by the frozen DPT heads into valid geometry and appearance.
    Core premise of latent-space generation; tested only indirectly through downstream metrics, not verified directly against the true latent data distribution.
  • standard math Sampling x0 ~ Unif(M) and regressing the geodesic conditional field defines a valid training objective for conditional generation.
    Sec 3.1/3.3 from Chen & Lipman [7]; standard Riemannian conditional flow matching.
  • standard math The parallel-transported velocity equals (1/(1-t)) Log_{x_t}(x1) along constant-speed geodesics.
    Appendix A.1; standard sphere geometry, used in the implementation.
  • ad hoc to paper The time-shift schedule with s=6 is a valid fixed training choice and does not qualitatively change the conclusions.
    Sec 4 Training details; no ablation of s is provided, so the sensitivity of the central results to this choice is unknown.

pith-pipeline@v1.3.0-alltime-deepseek · 25363 in / 14662 out tokens · 138515 ms · 2026-08-01T13:22:47.790834+00:00 · methodology

0 comments
read the original abstract

Geometric foundation models, such as the Visual Geometry Grounded Transformer (VGGT), provide strong 3D priors from unposed images. However, such models operate purely in a feed-forward, deterministic regime, \ie~they cannot generate plausible geometry beyond what the input views directly support. Generative models for 3D scenes, on the other hand, must rely on strong geometric priors to produce coherent outputs from sparse inputs. We bridge these two paradigms by performing flow matching directly in VGGT's latent space, leveraging its learned 3D priors without committing to any explicit downstream representation such as Gaussians, meshes, or video-VAE latents. This requires respecting the latent geometry: VGGT tokens occupy a product of high-dimensional hyperspheres on which standard Euclidean flow matching fails. We address this with a Riemannian Flow Matching framework defined on a product manifold of four hyperspheres, aligned with VGGT's multi-scale encoder, which keeps generated tokens on the valid data manifold required by the frozen decoding heads. On RealEstate10K, ScanNet++ and ETH3D, our method achieves strong performance against recent scene generation baselines in both per-view appearance and aggregated 3D geometry, establishing latent-space flow matching on geometric foundation models as a viable paradigm for 3D generation. The project page can be found $\href{https://lisaweijler.github.io/geometry-grounded-rfm/}{\text{here}}$.

Figures

Figures reproduced from arXiv: 2607.19120 by Guofeng Mei, Irene Ballester, Lisa Weijler, Pedro Hermosilla, Tolga Birdal.

Figure 1
Figure 1. Figure 1: Method overview. Given a sparse set of unposed RGB observations, a frozen VGGT encoder extracts per-view latent tokens. A learned Riemannian Flow Matching model vw generates latent codes conditioned on the tokens of the input views and a target camera pose, operating directly on the product manifold of four zero-mean hyperspheres. Frozen DPT decoding heads reconstruct point maps (PM) from the generated lat… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative novel view generation on ScanNet++ and RE10K. For each dataset we show the context view(s) followed by RGB and depth predictions for Ours, Gen3R, and DepthSplat. Results are shown for the one-sided setup with a single context view (left) and the two-sided setup with two context views (right). DepthSplat, lacking generative capability, leaves empty regions and artifacts where the target extends … view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results for 3D scan generation on ScanNet++ and ETH3D. Given the context RGB frames (left), we show the fused point clouds for the ground truth, VGGT (upper bound), our method, and Gen3R. Our approach yields reconstructions that are geometrically closer to the ground truth and more consistent across frames, while Gen3R suffers from distortions and frame misalignment. removing the trajectory and… 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

83 extracted references · 11 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip- nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5855–5864 (2021) 3

    2021

  2. [2]

    In: The Twelfth International Conference on Learning Representations (2024) 3

    Bose, J., Akhound-Sadegh, T., Huguet, G., FATRAS, K., Rector-Brooks, J., Liu, C.H., Nica, A.C., Korablyov, M., Bronstein, M.M., Tong, A.: Se (3)-stochastic flow matching for protein backbone generation. In: The Twelfth International Conference on Learning Representations (2024) 3

    2024

  3. [3]

    Cambridge University Press (2023) 18

    Boumal, N.: An introduction to optimization on smooth manifolds. Cambridge University Press (2023) 18

    2023

  4. [4]

    arXiv preprint arXiv:2506.07198 (2025) 4

    Bu, T., Wang, C., Ma, H., Zheng, H., Lu, X., Wu, T.: Ggball: Graph generative model on poincar\’e ball. arXiv preprint arXiv:2506.07198 (2025) 4

    arXiv 2025

  5. [5]

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

    Charatan, D., Li, S.L., Tagliasacchi, A., Sitzmann, V .: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19457–19467 (2024) 1, 3

    2024

  6. [6]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generaliz- able radiance field reconstruction from multi-view stereo. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 14124–14133 (2021) 3

    2021

  7. [7]

    In: The Twelfth International Conference on Learning Representations (2024) 2, 4, 17

    Chen, R.T.Q., Lipman, Y .: Flow matching on general geometries. In: The Twelfth International Conference on Learning Representations (2024) 2, 4, 17

    2024

  8. [8]

    arXiv preprint arXiv:2506.10981 (2025) 2, 3

    Chen, W., Bi, J., Huang, Y ., Zheng, W., Duan, Y .: Scenecompleter: Dense 3d scene completion for generative novel view synthesis. arXiv preprint arXiv:2506.10981 (2025) 2, 3

    Pith/arXiv arXiv 2025

  9. [9]

    In: European conference on computer vision

    Chen, Y ., Xu, H., Zheng, C., Zhuang, B., Pollefeys, M., Geiger, A., Cham, T.J., Cai, J.: Mvs- plat: Efficient 3d gaussian splatting from sparse multi-view images. In: European conference on computer vision. pp. 370–386. Springer (2024) 1, 3

    2024

  10. [10]

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

    Deng, K., Liu, A., Zhu, J.Y ., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12882–12891 (2022) 3

    2022

  11. [11]

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

    Elata, N., Kawar, B., Ostrovsky-Berman, Y ., Farber, M., Sokolovsky, R.: Novel view synthesis with pixel-space diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26756–26766 (2025) 1, 3

    2025

  12. [12]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Fischer, T., Bul `o, S.R., Yang, Y .H., Keetha, N., Porzi, L., M ¨uller, N., Schwarz, K., Luiten, J., Pollefeys, M., Kontschieder, P.: Flowr: Flowing from sparse to dense 3d reconstructions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 27702– 27712 (2025) 3

    2025

  13. [13]

    arXiv preprint arXiv:2510.14586 (2025) 3

    Frolova, D., Daulbaev, T., Sevryugov, E., Nikolenko, S.A., Ivankov, D.N., Oseledets, I., Pak, M.A.: Matcha: Multi-stage riemannian flow matching for accurate and physically valid molec- ular docking. arXiv preprint arXiv:2510.14586 (2025) 3

    arXiv 2025

  14. [14]

    Advances in Neural Information Processing Systems (2024) 3

    Gao*, R., Holynski*, A., Henzler, P., Brussee, A., Martin-Brualla, R., Srinivasan, P.P., Barron, J.T., Poole*, B.: Cat3d: Create anything in 3d with multi-view diffusion models. Advances in Neural Information Processing Systems (2024) 3

    2024

  15. [15]

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

    He, Y ., Tiwari, G., Birdal, T., Lenssen, J.E., Pons-Moll, G.: Nrdf: Neural riemannian distance fields for learning articulated pose priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1661–1671 (2024) 3

    2024

  16. [16]

    Advances in neural information processing systems30(2017) 7

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems30(2017) 7

    2017

  17. [17]

    In: ACM SIGGRAPH 2024 conference papers

    Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S.: 2d gaussian splatting for geometrically accurate radiance fields. In: ACM SIGGRAPH 2024 conference papers. pp. 1–11 (2024) 3

    2024

  18. [18]

    arXiv preprint arXiv:2601.04090 (2026) 2, 3, 6, 7, 19 12

    Huang, J., Yang, Y ., Yang, B., Ma, L., Ma, Y ., Liao, Y .: Gen3r: 3d scene generation meets feed-forward reconstruction. arXiv preprint arXiv:2601.04090 (2026) 2, 3, 6, 7, 19 12

    arXiv 2026

  19. [19]

    Advances in neural information processing systems37, 33007–33036 (2024) 3

    Huguet, G., Vuckovic, J., Fatras, K., Thibodeau-Laufer, E., Lemos, P., Islam, R., Liu, C., Rector-Brooks, J., Akhound-Sadegh, T., Bronstein, M., et al.: Sequence-augmented se (3)- flow matching for conditional protein generation. Advances in neural information processing systems37, 33007–33036 (2024) 3

    2024

  20. [20]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5885–5894 (2021) 3

    2021

  21. [21]

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

    Jang, W., Agapito, L.: Nvist: In the wild new view synthesis from a single image with trans- formers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 10181–10193 (June 2024) 1

    2024

  22. [22]

    Jang, W., Tremblay, J., Agapito, L.: Dt-nvs: Diffusion transformers for novel view synthesis (2025) 1, 3

    2025

  23. [23]

    arXiv preprint arXiv:2603.22275 (2026) 3

    Jang, W., Jeon, S., Han, J., Choi, J., Kwon, M., Kim, S., Xie, S., Liu, S.: Repurposing ge- ometric foundation models for multi-view diffusion. arXiv preprint arXiv:2603.22275 (2026) 3

    arXiv 2026

  24. [24]

    In: The Thirteenth Interna- tional Conference on Learning Representations (2025),https://openreview.net/forum? id=QQBPWtvtcn1, 3

    Jin, H., Jiang, H., Tan, H., Zhang, K., Bi, S., Zhang, T., Luan, F., Snavely, N., Xu, Z.: Lvsm: A large view synthesis model with minimal 3d inductive bias. In: The Thirteenth Interna- tional Conference on Learning Representations (2025),https://openreview.net/forum? id=QQBPWtvtcn1, 3

    2025

  25. [25]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimk ¨uhler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023) 1, 3

    2023

  26. [26]

    In: arXiv preprint arXiv:2602.21341 (2026) 1, 3

    Kim, E., Ryu, H., Mitchel, T.W., Sitzmann, V .: Scaling view synthesis transformers. In: arXiv preprint arXiv:2602.21341 (2026) 1, 3

    arXiv 2026

  27. [27]

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

    Kim, M., Seo, S., Han, B.: Infonerf: Ray entropy minimization for few-shot neural volume rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12912–12921 (2022) 3

    2022

  28. [28]

    arXiv preprint arXiv:2410.17270 (2024) 3

    Kim, N., Kim, S., Kim, M., Park, J., Ahn, S.: Mofflow: Flow matching for structure prediction of metal-organic frameworks. arXiv preprint arXiv:2410.17270 (2024) 3

    Pith/arXiv arXiv 2024

  29. [29]

    arXiv preprint arXiv:1412.6980 (2014) 7

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 7

    Pith/arXiv arXiv 2014

  30. [30]

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

    Kong, H., Yang, X., Wang, X.: Generative sparse-view gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26745–26755 (2025) 3

    2025

  31. [31]

    arXiv preprint arXiv:2602.10099 (2026) 2, 3, 4, 5

    Kumar, A., Patel, V .M.: Learning on the manifold: Unlocking standard diffusion transformers with representation encoders. arXiv preprint arXiv:2602.10099 (2026) 2, 3, 4, 5

    Pith/arXiv arXiv 2026

  32. [32]

    In: Proceed- ings of the European Conference on Computer Vision (2024) 3

    Leroy, V ., Cabon, Y ., Revaud, J.: Grounding image matching in 3d with mast3r. In: Proceed- ings of the European Conference on Computer Vision (2024) 3

    2024

  33. [33]

    arXiv preprint arXiv:2507.10496 (2025) 6

    Li, R., Yi, B., Liu, J., Gao, H., Ma, Y ., Kanazawa, A.: Cameras as relative positional encoding. arXiv preprint arXiv:2507.10496 (2025) 6

    arXiv 2025

  34. [34]

    arXiv preprint arXiv:2511.13720 (2025) 4

    Li, T., He, K.: Back to basics: Let denoising generative models denoise. arXiv preprint arXiv:2511.13720 (2025) 4

    Pith/arXiv arXiv 2025

  35. [35]

    In: 8th Annual Conference on Robot Learning (2024) 3

    Lim, B., Kim, J., Kim, J., Lee, Y ., Park, F.C.: Equigraspflow: Se (3)-equivariant 6-dof grasp pose generative flows. In: 8th Annual Conference on Robot Learning (2024) 3

    2024

  36. [36]

    arXiv preprint arXiv:2511.10647 (2025) 3

    Lin, H., Chen, S., Liew, J., Chen, D.Y ., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025) 3

    Pith/arXiv arXiv 2025

  37. [37]

    In: The Eleventh International Conference on Learning Representations (2023) 2, 3, 5

    Lipman, Y ., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: The Eleventh International Conference on Learning Representations (2023) 2, 3, 5

    2023

  38. [38]

    Liu, F., Sun, W., Wang, H., Wang, Y ., Sun, H., Ye, J., Zhang, J., Duan, Y .: Reconx: Reconstruct any scene from sparse views with video diffusion model (2024),https://arxiv.org/abs/ 2408.167672, 3 13

    Pith/arXiv arXiv 2024

  39. [39]

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

    Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., V ondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9298–9309 (October 2023) 1, 3

    2023

  40. [40]

    arXiv preprint arXiv:2506.07670 (2025) 3

    Lu, X., Fu, J., Zhang, J., Song, Z., Jia, C., Ma, S.: Prosplat: Improved feed-forward 3d gaussian splatting for wide-baseline sparse views. arXiv preprint arXiv:2506.07670 (2025) 3

    Pith/arXiv arXiv 2025

  41. [41]

    Communications of the ACM 65(1), 99–106 (2021) 1, 3

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99–106 (2021) 1, 3

    2021

  42. [42]

    In: International Conference on Machine Learning

    Miller, B.K., Chen, R.T., Sriram, A., Wood, B.M.: Flowmm: Generating materials with rie- mannian flow matching. In: International Conference on Machine Learning. pp. 35664–35686. PMLR (2024) 3

    2024

  43. [43]

    In: Intelligent Systems for Molecular Biology (ISMB) (2025) 3

    Morehead, A., Cheng, J.: Flowdock: Geometric flow matching for generative protein-ligand docking and affinity prediction. In: Intelligent Systems for Molecular Biology (ISMB) (2025) 3

    2025

  44. [44]

    ACM transactions on graphics (TOG)41(4), 1–15 (2022) 3

    M ¨uller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multires- olution hash encoding. ACM transactions on graphics (TOG)41(4), 1–15 (2022) 3

    2022

  45. [45]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Nadar, J., Foti, S., Birdal, T.: Posed-flow: Versatile and guided flow matching model of human pose. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 21165–21175 (2026) 3

    2026

  46. [46]

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

    Nair, N.G., Kaza, S., Luo, X., Patel, V .M., Lombardi, S., Park, J.: Scaling transformer-based novel view synthesis with models token disentanglement and synthetic data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 28567–28576 (October 2025) 1, 3

    2025

  47. [47]

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

    Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5480–5490 (2022) 3

    2022

  48. [48]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025) 3

    Ouyang, W., Ye, Q., Wang, J., Xu, Z., Chen, J.: Rfmpose: Generative category-level object pose estimation via riemannian flow matching. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025) 3

    2025

  49. [49]

    In: Proceed- ings of the IEEE/CVF international conference on computer vision

    Ranftl, R., Bochkovskiy, A., Koltun, V .: Vision transformers for dense prediction. In: Proceed- ings of the IEEE/CVF international conference on computer vision. pp. 12179–12188 (2021) 19

    2021

  50. [50]

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

    Ren, X., Shen, T., Huang, J., Ling, H., Lu, Y ., Nimier-David, M., M ¨uller, T., Keller, A., Fi- dler, S., Gao, J.: Gen3c: 3d-informed world-consistent video generation with precise camera control. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 6121–6132 (June 2025) 2, 3

    2025

  51. [51]

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

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthe- sis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684–10695 (2022) 3

    2022

  52. [52]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 3

    Sch ¨onberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 3

    2016

  53. [53]

    In: European Conference on Computer Vision (ECCV) (2016) 3

    Sch ¨onberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for un- structured multi-view stereo. In: European Conference on Computer Vision (ECCV) (2016) 3

    2016

  54. [54]

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

    Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3260– 3269 (2017) 6

    2017

  55. [55]

    Advances in Neural Information Processing Systems (NeurIPS) (2024) 2, 3

    Seo, J., Fukuda, K., Shibuya, T., Narihira, T., Murata, N., Hu, S., Lai, C.H., Kim, S., Mitsu- fuji, Y .: Genwarp: Single image to novel views with semantic-preserving generative warping. Advances in Neural Information Processing Systems (NeurIPS) (2024) 2, 3

    2024

  56. [56]

    Neurocomputing568, 127063 (2024) 6 14

    Su, J., Ahmed, M., Lu, Y ., Pan, S., Bo, W., Liu, Y .: Roformer: Enhanced transformer with rotary position embedding. Neurocomputing568, 127063 (2024) 6 14

    2024

  57. [57]

    arXiv preprint arXiv:2603.12655 (2026) 3

    Sun, X., Wang, S., Zhang, F., Liu, L., Jia, C., Song, Z., Huang, Z., Luo, Y .: Vggt-world: Trans- forming vggt into an autoregressive geometry world model. arXiv preprint arXiv:2603.12655 (2026) 3

    arXiv 2026

  58. [58]

    IEEE Transactions on pattern analysis and machine intelligence13(4), 376–380 (1991) 6, 22

    Umeyama, S.: Least-squares estimation of transformation parameters between two point pat- terns. IEEE Transactions on pattern analysis and machine intelligence13(4), 376–380 (1991) 6, 22

    1991

  59. [59]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Verninas, H., Korkmaz, C., Zafeiriou, S., Birdal, T., Foti, S.: Parallelised differentiable straightest geodesics for 3d meshes. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 14637–14647 (2026) 3

    2026

  60. [60]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025) 2, 3, 4, 7, 19

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geom- etry grounded transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025) 2, 3, 4, 7, 19

    2025

  61. [61]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024) 2, 3

    Wang, S., Leroy, V ., Cabon, Y ., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024) 2, 3

    2024

  62. [62]

    IEEE transactions on image processing13(4), 600–612 (2004) 7

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing13(4), 600–612 (2004) 7

    2004

  63. [63]

    In: ACM SIGGRAPH 2024 Con- ference Papers (2024)

    Wang, Z., Yuan, Z., Wang, X., Li, Y ., Chen, T., Xia, M., Luo, P., Shan, Y .: Motionctrl: A unified and flexible motion controller for video generation. In: ACM SIGGRAPH 2024 Con- ference Papers (2024). https://doi.org/10.1145/3641519.3657518 2, 3

    arXiv 2024

  64. [64]

    arXiv preprint arXiv:2508.14717 (2025) 3

    Wei, J., Leutenegger, S., Schaefer, S.: Gsfix3d: Diffusion-guided repair of novel views in gaussian splatting. arXiv preprint arXiv:2508.14717 (2025) 3

    Pith/arXiv arXiv 2025

  65. [65]

    In: European conference on computer vi- sion

    Wewer, C., Raj, K., Ilg, E., Schiele, B., Lenssen, J.E.: latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. In: European conference on computer vi- sion. pp. 456–473. Springer (2024) 3, 7

    2024

  66. [66]

    International Conference on Learning Representations (ICLR) (2026) 2, 3

    Wu, H., Wu, D., He, T., Guo, J., Ye, Y ., Duan, Y ., Bian, J.: Geometry forcing: Marrying video diffusion and 3d representation for consistent world modeling. International Conference on Learning Representations (ICLR) (2026) 2, 3

    2026

  67. [67]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    Wu, R., Mildenhall, B., Henzler, P., Park, K., Gao, R., Watson, D., Srinivasan, P.P., Verbin, D., Barron, J.T., Poole, B., Hoły ´nski, A.: Reconfusion: 3d reconstruction with diffusion priors. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024). https://doi.org/10.1109/CVPR52733.2024.02036 2, 3

    arXiv 2024

  68. [68]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Wu, S., Xu, C., Huang, B., Geiger, A., Chen, A.: Genfusion: Closing the loop between recon- struction and generation via videos. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 6078–6088 (2025) 3

    2025

  69. [69]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xu, H., Peng, S., Wang, F., Blum, H., Barath, D., Geiger, A., Pollefeys, M.: Depthsplat: Connecting gaussian splatting and depth. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16453–16463 (2025) 3, 6, 7, 20

    2025

  70. [70]

    generation: Taming optimization dilemma in latent diffusion models

    Yao, J., Yang, B., Wang, X.: Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2025) 4

    2025

  71. [71]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yeshwanth, C., Liu, Y .C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d in- door scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12–22 (2023) 6

    2023

  72. [72]

    arXiv preprint arXiv:2310.05297 (2023) 3

    Yim, J., Campbell, A., Foong, A.Y ., Gastegger, M., Jim ´enez-Luna, J., Lewis, S., Satorras, V .G., Veeling, B.S., Barzilay, R., Jaakkola, T., et al.: Fast protein backbone generation with se (3) flow matching. arXiv preprint arXiv:2310.05297 (2023) 3

    Pith/arXiv arXiv 2023

  73. [73]

    arXiv preprint arXiv:2508.09667 (2025) 3

    Yin, X., Zhang, Q., Chang, J., Feng, Y ., Fan, Q., Yang, X., Pun, C.M., Zhang, H., Cun, X.: Gsfixer: Improving 3d gaussian splatting with reference-guided video diffusion priors. arXiv preprint arXiv:2508.09667 (2025) 3

    Pith/arXiv arXiv 2025

  74. [74]

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

    Yu, A., Ye, V ., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4578–4587 (2021) 1, 3 15

    2021

  75. [75]

    IEEE Transactions on Pattern Analysis & Machine Intelligence (2025)

    Yu, W., Xing, J., Yuan, L., Hu, W., Li, X., Huang, Z., Gao, X., Wong, T.T., Shan, Y ., Tian, Y .: ViewCrafter: Taming Video Diffusion Models for High-fidelity Novel View Synthesis . IEEE Transactions on Pattern Analysis & Machine Intelligence (2025). https://doi.org/10.1109/TPAMI.2025.3613256 2, 3

    arXiv 2025

  76. [76]

    In: European Conference on Computer Vision

    Zhang, K., Bi, S., Tan, H., Xiangli, Y ., Zhao, N., Sunkavalli, K., Xu, Z.: Gs-lrm: Large reconstruction model for 3d gaussian splatting. In: European Conference on Computer Vision. pp. 1–19. Springer (2024) 1, 3

    2024

  77. [77]

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

    Zhang, Q., Zhai, S., Martin, M.A.B., Miao, K., Toshev, A., Susskind, J., Gu, J.: World- consistent video diffusion with explicit 3d modeling. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 21685–21695 (June 2025) 1, 3

    2025

  78. [78]

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

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018) 7

    2018

  79. [79]

    In: The Fourteenth International Conference on Learning Representations (2026) 2, 3, 4

    Zheng, B., Ma, N., Tong, S., Xie, S.: Diffusion transformers with representation autoencoders. In: The Fourteenth International Conference on Learning Representations (2026) 2, 3, 4

    2026

  80. [80]

    In: Advances in Neural Information Processing Systems (2024) 3

    Zhou, J., Zhang, W., Liu, Y .S.: Diffgs: Functional gaussian splatting diffusion. In: Advances in Neural Information Processing Systems (2024) 3

    2024

Showing first 80 references.