Pith. sign in

REVIEW 3 major objections 5 minor 41 references

GraspMeanFlow: SE(3)-Equivariant MeanFlow for Few-Step 6-DoF Grasp Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read This paper claims that predicting the interval-averaged SE(3) velocity, defined via the time-ordered exponential, lets one network evaluation match the grasp quality an iterative flow reaches in five steps, with equivariance provably preser

desk verdict Careful and useful SE(3) MeanFlow extension, but the headline few-step success gain rests on an unsupported endpoint extrapolation in the exp-SO(3) sampler. read the letter →

arxiv 2608.03295 v1 pith:UMBLAJYQ submitted 2026-08-04 cs.RO

classification cs.RO
keywords SE(3)equivariance6-DoFgraspgenerationMeanFlowfew-stepsamplingaveragevelocitymatchingsemigroupconsistencyACRONYM
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

This paper claims that the bottleneck of flow-based 6-DoF grasp generation — tens of network evaluations per grasp to integrate a pose trajectory — can be removed by learning the average velocity over a finite interval rather than the instantaneous velocity, with the SE(3) average defined through the time-ordered exponential so that one Lie-algebra element reproduces exactly the rigid-body displacement accumulated over the interval. On the ACRONYM benchmark, a single function evaluation of the framework reaches the Earth-mover distance (EMD) that the iterative SE(3) flow baseline EquiGraspFlow needs five steps to approach (0.418 versus 0.416), and a second instantiation of the same framework improves simulated grasp success by up to 24.3 points at five steps. The paper also proves that the generated grasp distribution still transforms exactly with the object under few-step sampling, because the equivariance argument is carried by composition of equivariant flow maps — stated for pushforward measures, not diffeomorphisms. A sympathetic reader would care because real-time manipulation regenerates and re-ranks hundreds of grasp candidates per scene change, where per-evaluation sampling cost is the binding constraint.

What carries the argument

The central object is the interval-averaged SE(3) velocity field $(\omega^{\mathrm{avg}}, v^{\mathrm{avg}})(s,t,\mathcal{P},T_t)$, defined so that $\exp((t-s)[\omega^{\mathrm{avg}}])=\mathrm{Texp}(\int_s^t[\omega_\tau]\,d\tau)$ exactly reproduces the rotation accumulated over $[s,t]$ (translation is the plain integral): one Lie-algebra element whose step executes an entire interval jump, with sampling as a direct update instead of numerical integration. Around it sit three supporting mechanisms: the semigroup identity $D(s,t)=D(m,t)D(s,m)$, giving a Jacobian-free consistency loss (GMF-SG); the differential MeanFlow identity $J(\omega^{s\to t})\frac{d}{dt}\omega^{s\to t}=\omega_t$ with the SO

What would settle it

Sample poses near the prior ($t$ close to 1), form the interval field's implied endpoint $\hat{R}_0=\exp(-t[\bar{\omega}])R_t$ against the true data rotation $R_0$, and measure the rotation error. If the error is of the order of a random rotation, the exp-SO(3) sampler's few-step success cannot rest on accurate interval averages and the single-evaluation EMD parity is not explained by the claimed mechanism; if the error is small, the premise holds. A companion check: add explicit whole-interval endpoint supervision to the training objective and see whether the NFE-5 success numbers move.

Watch

Extended reading notes

Core claim

The paper's claim is that the iterative sampling cost of equivariant flow models for 6-DoF grasp generation is not inherent. It defines the average velocity on SE(3) through the time-ordered exponential, so that one learned Lie-algebra element $(\omega^{\mathrm{avg}}, v^{\mathrm{avg}})(s,t,\mathcal{P},T_t)$ reproduces exactly the rigid-body displacement over $[s,t]$, and sampling becomes one direct step, $R_s=\exp(-(t-s)[\omega^{\mathrm{avg}}])R_t$, $x_s=x_t-(t-s)v^{\mathrm{avg}}$. The load-bearing result is that equivariance survives few-step sampling: an equivariant average-velocity flow map commutes with the SE(3) action, so an invariant prior pushed through any composition of such maps s

Load-bearing premise

The few-step gains — in particular the exp-SO(3) sampler's 24.3-point success advantage — require the network to predict accurately where a full-interval jump lands when starting from a pose that is still almost a random prior sample, yet the training objective supervises only the diagonal ($s=t$) and interval self-consistency, never that whole-interval endpoint prediction; if the extrapolation is wrong, the headline results are not guaranteed.

Editorial extensions

If this is right

  • Equivariance survives one-step and few-step sampling: because the invariance argument is stated for pushforward measures, the generated grasp distribution continues to transform exactly with the object even when the transport map is not a diffeomorphism.
  • Real-time manipulation becomes feasible at the candidate counts pipelines want: one field evaluation costs 31.5 ms on a single A100 and reaches the EMD the iterative baseline achieves in five steps, so a hundred-candidate set costs roughly one short integration.
  • The interval-averaged field admits an endpoint prediction at every step, which an instantaneous field does not; this is the mechanism behind the exp-SO(3) sampler and why the advantage concentrates at NFE 5–10, with the fully integrated baseline regaining the lead at NFE 20.
  • The few-step gains come from the training objective and the sampler it enables, not from capacity: both instantiations share the baseline network plus 512 added parameters (<0.1% of the model), under the same training budget.
  • Classifier-free guidance costs no extra evaluations: the guided average velocity is modeled directly as an affine combination of two equivariant fields, so the equivariance guarantee applies unchanged at the same step budget.

Reading between the lines

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

  • Because the average velocity is defined purely through the time-ordered exponential and the semigroup identity, the construction should transfer to other SE(3)-valued generation tasks — object-pose hypotheses, camera relocalization, assembly motions — wherever endpoint accuracy matters more than trajectory fidelity; the paper notes a concurrent application to protein backbones in the body frame.
  • The paper's stated symmetry limitation points to a concrete repair: a target that takes the logarithm to the nearest representative in the object's symmetry group should recover the Pencil category, where rotations about the pencil's long axis that are physically equivalent are currently counted as genuine distance.
  • The negative result on success-filtered ReFlow suggests the full-interval endpoint prediction is the least-regulated quantity in the objective; a direct test would add an explicit whole-interval endpoint-supervision term and check whether the NFE-5 success gains move.
  • Since the field already implies a data endpoint for every t, one evaluation at t=1 is in principle a full one-step generator; the reported NFE-1 success (~18.7%) shows that jump from the prior is the hard part, so an annealed or learned starting schedule that begins closer to the data could extend the few-step regime.
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

3 major / 5 minor

Summary. The paper proposes GraspMeanFlow, an SE(3)-equivariant MeanFlow framework for few-step 6-DoF grasp generation. The average velocity over a finite interval is defined on SE(3) through the time-ordered exponential, so that a single Lie-algebra element reproduces the rigid-body displacement over that interval. The authors prove that equivariant average-velocity flow maps preserve invariance of the generated conditional distribution under few-step sampling, and they design a two-time equivariant conditioning mechanism that adds only 512 parameters to the EquiGraspFlow backbone. Training combines a flow-matching boundary term with either a semigroup consistency loss (GMF-SG) or a differential/JVP identity (GMF-JVP), preceded by an alpha-flow warm-up. Experiments on ACRONYM report that one evaluation of GMF-SG reaches an EMD comparable to five-step EquiGraspFlow, that GMF-JVP improves grasp success by up to 24.3 points at NFE 5 with an exp-SO(3) sampler, and that both variants generate equivariant grasp distributions. The theoretical derivations are detailed and numerically validated in the appendices, but the headline few-step success claim rests on a sampler heuristic whose endpoint prediction is not covered by the training objective or the equivariance propositions.

Significance. If the results hold, GraspMeanFlow is a valuable contribution: it extends MeanFlow to SE(3) in a principled way, retains equivariance in the few-step regime, and demonstrates practical gains in grasp generation with very few function evaluations. The paper ships a large amount of supporting evidence: closed-form derivations of the SE(3) MeanFlow identity, a semigroup formulation that avoids Jacobian-vector products, numerical verification of the differential identity (Figure 5) and of the alpha->0 limit (Figure 6), and machine-precision equivariance checks (Figure 7). The ablation study and the explicit negative result on ReFlow post-training add credibility. However, the experimental claim of up to 24.3-point success improvement is produced by the exp-SO(3) sampler, whose endpoint extrapolation is not supervised by the training losses and is not guaranteed by the theory. The significance of the paper therefore hinges on whether that heuristic can be justified, replaced, or re-framed.

major comments (3)
  1. [§4 / Appendix L (Algorithm 4)] The headline few-step success gain (24.3 points at NFE 5) is produced by the exp-SO(3) sampler, not by the Euler sampler that directly realizes the trained average-velocity step. In Algorithm 4, line 5 evaluates the field on [t-Δt, t], and line 6 forms the endpoint prediction as exp(-t[ω̄])R, i.e., it multiplies the near-interval average by the full remaining time t. This treats the average velocity over [t-Δt, t] as if it were the average over [0, t]. The training losses (Eqs. 17, 19, 20) enforce interval self-consistency and a diagonal boundary; they do not enforce equality between a short-interval average and the full-interval average. Proposition 9 gives exactness for the field's own [s,t] displacement, not for this extrapolated endpoint. Moreover, Appendix L's prose says the sampler 'queries the field over [0,t]', but the pseudocode queries [t-Δt,t], a discrepancy suggesting the end
  2. [§4 / Appendix J] All main quantitative results (Table 1, Table 4, Figure 2) come from single training runs, as stated in Appendix J: 'Each reported configuration is a single training run.' The error bars in Figure 3 are standard errors over rotated test instances, not over training seeds. The objective ablation averages three seeds, but the headline comparisons do not. Given that several NFE-5 differences are small (e.g., Mug GMF-JVP 88.1 vs. BRIDGER 90.6; Pencil GMF-SG 77.9 vs. EquiGraspFlow 81.3), and hyperparameters were selected under a reduced protocol before the reported runs, the cross-method rankings at NFE 5/10 are not yet supported by evidence of seed stability. Please report seed variance for at least the headline comparisons (EMD at NFE 1 and success at NFE 5).
  3. [§4 / Appendix P] GMF-SG and GMF-JVP differ in two factors: the consistency term and the prior–data coupling (independent vs. per-object optimal transport). Appendix P explicitly states that the four-way comparison separating these effects was outside the compute budget and that 'the gap between the two configurations should not be attributed to either factor alone.' The main text nevertheless describes the two as complementary instantiations and uses the success difference as evidence for the objective. This is acknowledged in the appendix, but the main-text framing overstates what the comparison identifies. Please clarify in the main text that the GMF-SG vs. GMF-JVP gap is not attributable to the consistency term alone.
minor comments (5)
  1. [References] The reference 'Geng, Z.; et al. 2025b. MeanFlow: One-step generative modeling via average velocity' is incomplete; it should include the full author list, venue, and publication details.
  2. [Abstract / §4] The abstract states that a single function evaluation 'reaches the EMD that an iterative SE(3) flow model needs five steps to approach.' Table 4 shows 0.418 vs. 0.416, i.e., within 0.002. This is a fair statement, but 'approach' is vague; a precise statement such as 'within 0.002 EMD' would be clearer.
  3. [Appendix L] Algorithm 4 uses tmin and a final jump at line 16-17, but tmin is only mentioned in the text after the algorithm. Define tmin in the pseudocode or immediately before it.
  4. [Figure 7] The x-axis labels 'rot trans SE(3) Euler exp-SO(3)' are cryptic; please expand them to 'rotation-only, translation-only, full SE(3), Euler sampler, exp-SO(3) sampler' for readability.
  5. [§5] The sentence 'an interval-averaged field admits an endpoint prediction at every step' is exactly the disputed claim in the major comment; please qualify it by noting that the endpoint prediction used in Algorithm 4 is an extrapolation from a near-interval average, not the field's own full-interval prediction.

Circularity Check

0 steps flagged · score 2.0 of 10

No by-construction circularity: the SE(3) average-velocity field is defined by Eq. (7), the training losses enforce derived consistency identities plus a flow-matching anchor, and equivariance is proved from the architecture. The only self-citation is descriptive and non-load-bearing; the exp-SO(3) sampler discrepancy is a correctness/extrapolation concern, not a circular step.

full rationale

The derivation chain is not circular. The SE(3) average velocity is defined in Eq. (7) through the time-ordered exponential as the constant Lie-algebra velocity reproducing the interval displacement; the differential MeanFlow identity Eq. (10) is derived from that definition, and the semigroup identity Eq. (15) is the associativity of group multiplication. The training targets in Eqs. (13), (17), and (20) regress the network against these mathematical identities together with the flow-matching boundary Eq. (19). Proposition 9's zero-loss exactness is a theorem whose hypotheses are stated independently of the empirical results, and Proposition 10 proves equivariance from the vector-neuron architecture rather than assuming it. The comparisons to EquiGraspFlow, SE(3)-DiffusionFields, and BRIDGER on ACRONYM are external and not constructed from the method's own outputs. There is one self-citation—Bai et al. (2026), the concurrent protein SE(3)-MeanFlow paper—but it is used only to note the parallel formulation and is not load-bearing for any grasp-generation claim. The internal discrepancy in Appendix L is a real missing-support issue: the prose says the endpoint-style sampler 'queries the field over [0,t]', while Algorithm 4 line 5 evaluates the field on [t−Δt,t] and line 6 multiplies that near-interval average by the full remaining time t to form exp(−t[ω̄])R. This is an extrapolation, not an equivalence-by-construction, so it does not make the derived claim circular; it is a heuristic whose accuracy is not guaranteed by the training losses or the equivariance propositions. The paper also candidly discloses related limitations (Pencil symmetry, open endpoint-rate scheduling, the negative success-filtered ReFlow result, and the coupling confound in Appendix P), which further indicates that the reported gains are empirical rather than forced by definitions. Score 2 reflects only the minor, non-load-bearing self-citation; the central derivation is self-contained.

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

The method introduces no new physical entities. The only additions are mathematical constructions (time-ordered-exponential average velocity, semigroup/α-Flow losses) and training hyperparameters, listed above.

free parameters (5)
  • Consistency weight λ_sg = 1.0 (GMF-SG), 1.7 (GMF-JVP)
    Chosen so the boundary and consistency terms contribute comparably; tuned on a reduced single-category protocol (Appendix J).
  • Exp-SO(3) sampler rate c = 10
    Constant rate for the exponential rotation scheduler, capped at 1; taken from ReQFlow (Yue et al. 2025) and not tuned.
  • Huber radius on rotation residual = 100
    Tried {20,50,100}; kept the loosest, which 'almost never fires'; differences at most 0.007 EMD.
  • α-Flow warm-up schedule = 1 → 0.2 over 18k steps
    Annealed sigmoid; the warm-up is reported as optional (Appendix P), so this is a training recipe choice.
  • Guidance weight β = 2.0
    Classifier-free guidance weight inherited from EquiGraspFlow, not re-tuned.
assumptions (5)
  • domain assumption The conditional path between prior and data uses geodesic interpolation on SO(3) with constant angular velocity ω_t = log(R_1 R_0^T)^∨ and linear interpolation in R^3 (Eqs. (2)-(3)).
    This defines the training targets and the closed-form average velocity; if the path model is wrong for the grasp distribution, the objective targets are mis-specified.
  • domain assumption The prior p_1 is SE(3)-invariant: uniform on SO(3), Gaussian centered at the point-cloud mean with σ=1 (Section 3, Conditional prior).
    Needed for Proposition 2 (few-step invariance); relies on exact mean-centering in preprocessing.
  • domain assumption The backbone (VN-DGCNN encoder and vector-neuron head) is SE(3)-equivariant, and the two-time injection via a shared learned direction preserves equivariance (Lemmas 1-2, Appendix G/H).
    Underlies Proposition 10; depends on the architectural properties inherited from EquiGraspFlow.
  • standard math Uniqueness and smoothness of solutions to the ODEs on SO(3), the left Jacobian formula, the BCH expansion, and the Wilcox representation of d(exp) (Appendix C), along with the semigroup composition D(s,t)=D(m,t)D(s,m) (Proposition 6).
    Background facts used in the derivations of Eqs. (10), (45), (50).
  • ad hoc to paper Approximate minimization of the boundary-plus-consistency loss yields a field whose global minimizer is the exact average-velocity field (Proposition 9), and the α-Flow warm-up with data velocity on the near segment is a necessary practical device.
    The uniqueness result assumes the semigroup identity holds exactly; in practice the loss is minimized in expectation, and the α-Flow warm-up is introduced solely to stabilize early training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GraspMeanFlow: SE(3)-Equivariant MeanFlow for Few-Step 6-DoF Grasp Generation." pith.science (2026). https://pith.science/paper/UMBLAJYQ

@misc{pith2026260803295,
  author       = {Pith},
  title        = {Pith review of: GraspMeanFlow: SE(3)-Equivariant MeanFlow for Few-Step 6-DoF Grasp Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UMBLAJYQ}},
  note         = {Machine review of arXiv:2608.03295}
}
read the original abstract

Recent data-driven methods for synthesizing 6-DoF grasp poses use generative models to learn complex grasp pose distributions and generate diverse candidate poses. In particular, SE(3)-equivariant flow-based models generate grasp poses that transform consistently with object rotations and translations. However, these methods sample by iterative numerical integration, requiring tens of function evaluations per grasp and limiting their use in real-time manipulation. We propose GraspMeanFlow, an SE(3)-equivariant MeanFlow framework for few-step 6-DoF grasp generation. Our method learns the average velocity over a finite time interval, defined through the time-ordered exponential so that it reproduces exactly the rigid-body displacement accumulated over that interval. We prove that a point-cloud-conditioned distribution transported by an equivariant average-velocity flow map remains invariant, so equivariance is retained under few-step sampling, and we condition the field on a pair of times by lifting both to equivariant vectors, leaving the backbone otherwise unchanged. For stable training, we pair a flow-matching boundary term with either of two consistency terms: the differential MeanFlow identity, whose target requires a Jacobian-vector product, or an equivalent semigroup loss that avoids it. Experiments on ACRONYM show that a single function evaluation of GraspMeanFlow reaches the EMD that an iterative SE(3) flow model needs five steps to approach, that a second instantiation of the same framework improves grasp success by up to 24.3 points in the few-step regime, and that both generate grasp distributions transforming exactly with the object.

Figures

Figures reproduced from arXiv: 2608.03295 by the authors.

Figure 1
Figure 1. Overview of the proposed GraspMeanFlow framework. Given an object point cloud P, a current grasp pose Tt initialized from the object-centered prior at t = 1, and a time pair (s, t) with s < t, the SE(3)-equivariant average-velocity network predicts the interval-averaged an￾gular and linear velocities (ω avg θ , v avg θ ). An Euler or exp￾SO(3) sampler then uses these predictions to transport Tt to the earlier-time p… view at source ↗
Figure 2
Figure 2. EMD versus NFE, averaged over the four cate [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Grasps generated at NFE 5 for one test instance per [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Relative residual of (10) over 256 random smooth trajectories, computed with the left Jacobian (red) and with the Jacobian dropped (J=I, gray). Thirteen orders of magnitude separate the exact identity from its linearization. Proposition 5. If L diff = 0, i.e., J(ω s→t …
Figure 6
Figure 6. Figure 6: Numerical verification of Proposition 7, Eq. (46): gap between the scaled [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Equivariance of the trained model in double precision. Each dot is one random [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 24 canonical work pages

  1. [1]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    6-dof graspnet: Variational grasp generation for object manipulation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  2. [2]

    Urain, Julen and Funk, Niklas and Peters, Jan and Chalvatzaki, Georgia , booktitle =

  3. [3]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Vector neurons: A general framework for so (3)-equivariant networks , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  4. [4]

    MeanFlow: One-step generative modeling via average velocity , author =

  5. [5]

    Riemannian MeanFlow for One-Step Generation on Manifolds

    Riemannian MeanFlow for One-Step Generation on Manifolds , author=. arXiv preprint arXiv:2603.10718 , year=

  6. [6]

    arXiv preprint arXiv:2602.07744 , year=

    Riemannian MeanFlow , author=. arXiv preprint arXiv:2602.07744 , year=

  7. [7]

    Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , year =

    ACRONYM: A Large-Scale Grasp Dataset Based on Simulation , author =. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , year =

  8. [8]

    Cheng, Chaoran and Li, Jiahan and Fan, Jiajun and Liu, Ge , journal=

Show all 41 references
  1. [9]

    , booktitle =

    Lim, Byeongdo and Kim, Jongmin and Kim, Jihwan and Lee, Yonghyeon and Park, Frank C. , booktitle =

  2. [10]

    Proceedings of Robotics: Science and Systems (RSS) , address =

    Don't Start From Scratch: Behavioral Refinement via Interpolant-based Policy Diffusion , author =. Proceedings of Robotics: Science and Systems (RSS) , address =. 2024 , doi =

  3. [11]

    2021 , url =

    Makoviychuk, Viktor and Wawrzyniak, Lukasz and Guo, Yunrong and Lu, Michelle and Storey, Kier and Macklin, Miles and Hoeller, David and Rudin, Nikita and Allshire, Arthur and Handa, Ankur and State, Gavriel , booktitle =. 2021 , url =

  4. [12]

    International Conference on Learning Representations , year =

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author =. International Conference on Learning Representations , year =

  5. [13]

    Advances in Neural Information Processing Systems , volume=

    Equivariant manifold flows , author=. Advances in Neural Information Processing Systems , volume=

  6. [14]

    The International Journal of Robotics Research , volume =

    ten Pas, Andreas and Gualtieri, Marcus and Saenko, Kate and Platt, Robert , title =. The International Journal of Robotics Research , volume =. 2017 , doi =

  7. [15]

    2019 International Conference on Robotics and Automation (ICRA) , pages =

    Liang, Hongzhuo and Ma, Xiaojian and Li, Shuang and G. 2019 International Conference on Robotics and Automation (ICRA) , pages =. 2019 , doi =

  8. [16]

    Proceedings of the Conference on Robot Learning , editor =

    Qin, Yuzhe and Chen, Rui and Zhu, Hao and Song, Meng and Xu, Jing and Su, Hao , title =. Proceedings of the Conference on Robot Learning , editor =. 2020 , url =

  9. [17]

    2021 IEEE International Conference on Robotics and Automation (ICRA) , pages =

    Sundermeyer, Martin and Mousavian, Arsalan and Triebel, Rudolph and Fox, Dieter , title =. 2021 IEEE International Conference on Robotics and Automation (ICRA) , pages =. 2021 , doi =

  10. [18]

    Proceedings of the 2020 Conference on Robot Learning , editor =

    Breyer, Michel and Chung, Jen Jen and Ott, Lionel and Siegwart, Roland and Nieto, Juan , title =. Proceedings of the 2020 Conference on Robot Learning , editor =. 2021 , url =

  11. [19]

    IEEE Robotics and Automation Letters , volume =

    Weng, Zehang and Lu, Haofei and Lundell, Jens and Kragic, Danica , title =. IEEE Robotics and Automation Letters , volume =. 2024 , doi =

  12. [20]

    arXiv preprint arXiv:2310.02391 , year=

    Se (3)-stochastic flow matching for protein backbone generation , author=. arXiv preprint arXiv:2310.02391 , year=

  13. [21]

    arXiv preprint arXiv:2302.00482 , year=

    Improving and generalizing flow-based generative models with minibatch optimal transport , author=. arXiv preprint arXiv:2302.00482 , year=

  14. [22]

    arXiv preprint arXiv:2502.14637 , year=

    Reqflow: Rectified quaternion flow for efficient and high-quality protein backbone generation , author=. arXiv preprint arXiv:2502.14637 , year=

  15. [23]

    arXiv preprint arXiv:2512.02012 , year=

    Improved Mean Flows: On the Challenges of Fastforward Generative Models , author=. arXiv preprint arXiv:2512.02012 , year=

  16. [24]

    2009 , publisher=

    Optimal transport: old and new , author=. 2009 , publisher=

  17. [25]

    arXiv preprint arXiv:2304.14772 , year=

    Multisample flow matching: Straightening flows with minibatch couplings , author=. arXiv preprint arXiv:2304.14772 , year=

  18. [26]

    Bai, Yikun and Lu, Binghang and Liu, Yikai and Akbari, Elaheh and Kolouri, Soheil and Wang, Linxuan and He, Ping and Wang, Shuchan and Zhang, Ruqi and Lin, Guang , journal=

  19. [27]

    International Conference on Learning Representations , year =

    Flow Matching for Generative Modeling , author =. International Conference on Learning Representations , year =

  20. [28]

    Flow Matching for Generative Modeling , author =

  21. [29]

    International Conference on Learning Representations (ICLR) , year=

    Flow Matching on General Geometries , author=. International Conference on Learning Representations (ICLR) , year=

  22. [30]

    Nature , volume=

    De novo design of protein structure and function with RFdiffusion , author=. Nature , volume=. 2023 , publisher=

  23. [31]

    arXiv preprint arXiv:2302.02277 , year=

    SE (3) diffusion model with application to protein backbone generation , author=. arXiv preprint arXiv:2302.02277 , year=

  24. [32]

    arXiv preprint arXiv:2311.13443 , year=

    Guided flows for generative modeling and decision making , author=. arXiv preprint arXiv:2311.13443 , year=

  25. [33]

    and Jahr, Philipp and Sun, Qiao and Urain, Julen and Koert, Dorothea and Peters, Jan , journal =

    Carvalho, Joao and Le, An T. and Jahr, Philipp and Sun, Qiao and Urain, Julen and Koert, Dorothea and Peters, Jan , journal =. Grasp Diffusion Network: Learning Grasp Generators from Partial Point Clouds with Diffusion Models in

  26. [34]

    Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , year =

    GraspGen: A Diffusion-based Framework for 6-DOF Grasping with On-Generator Training , author =. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , year =

  27. [35]

    Nature biotechnology , volume=

    Computational design of trimeric influenza-neutralizing proteins targeting the hemagglutinin receptor binding site , author=. Nature biotechnology , volume=. 2017 , publisher=

  28. [36]

    Consistency models , author=

  29. [37]

    arXiv preprint arXiv:2202.00512 , year=

    Progressive distillation for fast sampling of diffusion models , author=. arXiv preprint arXiv:2202.00512 , year=

  30. [38]

    Advances in neural information processing systems , volume=

    Riemannian score-based generative modelling , author=. Advances in neural information processing systems , volume=

  31. [39]

    Advances in neural information processing systems , volume=

    Riemannian continuous normalizing flows , author=. Advances in neural information processing systems , volume=

  32. [40]

    International Conference on Learning Representations , volume=

    Decentralized Riemannian conjugate gradient method on the Stiefel manifold , author=. International Conference on Learning Representations , volume=

  33. [41]

    M. J. Kearns , title =

Pith tools

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