Pith. sign in

REVIEW 3 major objections 4 minor 49 references

SGFlow trains a model to learn the entire probability-flow ODE map with a stopgrad-based loss whose unique stationary point is the true flow map.

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 16:45 UTC pith:EYWWUTNV

load-bearing objection SGFlow has a genuinely new training objective and a clean Gaussian counterexample, but the central stationary-point theorem has a real proof gap and the training time-pair sampler is unspecified. the 3 major comments →

arxiv 2607.26398 v1 pith:EYWWUTNV submitted 2026-07-29 cs.LG

Flow Map Learning via Nongradient Vector Flow

Mark Goldstein , Anshuk Uppal , Raghav Singhal , Aahlad Puli , Rajesh Ranganath This is my paper
classification cs.LG
keywords flow map learningconsistency modelsprobability flow ODEstopgradnon-conservative dynamicsstationary pointgenerative modelingCIFAR-10
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.

SGFlow is a training objective for generative models that learns the whole probability-flow ODE map — the direct noise-to-data mapping along the ODE trajectory — without model inverses, nested model calls, or adversarial training. The paper's central claim is that, under a time-pair sampler that puts positive probability on the diagonal t=u, the true flow map is the unique stationary point of the stopgrad-induced training dynamics, so the optimization cannot settle at a map that integrates an incorrect velocity. The proof works by showing the stopgrad loss and the original transport-equation loss have identical stationary points, and that any stationary point must match the true conditional velocity on the diagonal. On CIFAR-10, SGFlow obtains the best FID among the compared methods at 10 sampling steps and stays competitive at 1, 50, and 100 steps, while being the only compared method with a proven stationary-point guarantee for its stopgrad dynamics.

Core claim

The paper claims the true flow map f(t,u,x) — the integral of the probability-flow ODE — is the unique stationary point of SGFlow's stopgrad loss Lsg, a functional built only from Jacobian-vector products and the sample-level velocity Ẋt, never the unknown conditional velocity v=E[Ẋt|Xt]. Theorem 1: for a time-pair sampler q(t,u) supported on t≤u with positive mass on t=u, a function is stationary for Lsg exactly when it is stationary for the original loss L that squares the transport-PDE residual ∂t f + (∂x f)v = 0. Lemma 1 shows any stationary point of L matches the true velocity at t=u, forcing it to solve the PDE and hence be the true flow map. So the true flow map is the unique stationa

What carries the argument

The central object is the transport PDE ∂t f(t,u,x) + (∂x f(t,u,x)) v(t,x) = 0, which characterises the flow map f that integrates the probability-flow ODE from t to u. The model is parameterised as fθ(t,u,x) = x + (u−t) f̃θ(t,u,x), which automatically enforces fθ(u,u,x)=x and makes the diagonal time-derivative equal to −f̃θ(t,t,x). SGFlow squares the residual of the transport PDE, replacing the unknown v with stopgrad[f̃θ(t,t,x)]; the stopgrad matters because, without it, the true flow map is provably not a stationary point (the Gaussian counterexample shows a spurious attractor). The proof machinery is the first-variation analysis of functionals with stopgrad, which shows Lsg and L share s

Load-bearing premise

The theorem requires the time-pair sampler q(t,u) to put positive probability on the diagonal t=u; the paper never states the q used in training, and with the usual continuous-time sampling the diagonal has probability zero, so the proven guarantee does not literally cover the implemented method.

What would settle it

Train SGFlow on the closed-form 1D Gaussian example using a purely continuous time-pair distribution (e.g., t and u drawn uniformly with t ≤ u, so P(t=u)=0) and measure the expected parameter update at the true flow map. A nonzero update would show the stationary-point guarantee fails outside the theorem's diagonal-mass condition; a zero update would indicate the condition is not needed in practice.

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

If this is right

  • Flow-map generative models can be trained from scratch with a regression-style, simulation-free loss — no teacher velocity, no model inverse, no nested model calls — and sampling can use any number of steps along the true ODE trajectory.
  • At 10 sampling steps on CIFAR-10, SGFlow achieves the best FID among flow matching, Meanflow, Lagrangian map matching, and SGFlow, and is competitive at 1, 50, and 100 steps, suggesting the stationary-point guarantee does not come at an empirical cost.
  • The stopgrad is a correctness requirement rather than a computational convenience: the paper shows that removing it makes the true flow map a non-stationary point of the Meanflow functional, with the bias concentrated in the variance parameter in the Gaussian example.
  • Any flow map that integrates the wrong velocity is not a stationary point of the training dynamics, because the diagonal t=u term continuously corrects the velocity estimate; this is the mechanism that prevents distillation of an incorrect velocity.
  • The method can be interpreted as training a normalizing-flow-style map without an invertible architecture: it learns only the forward map and gives up explicit access to the inverse, which is acceptable for generative sampling and allows likelihood evaluation by substituting −∂t fθ for the velocity.

Where Pith is reading between the lines

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

  • The theorem's hypothesis of positive probability on the diagonal t=u is not verified for the actual training schedule — the paper never specifies q(t,u) for the CIFAR runs, and standard continuous time-pair sampling gives zero diagonal mass. If the implementation did use a continuous q, the proven stationary-point guarantee does not literally apply, and the method's success rests on an unproven ex
  • The same stationary-point-equivalence technique could be applied to other stopgrad-based flow-map losses (such as the self-distillation losses that currently lack guarantees), giving a general recipe: find a term on the diagonal that pins the velocity, then show off-diagonal updates coincide with those of a loss whose optimum is known.
  • A natural testable extension is a hybrid time sampler that anneals the probability of drawing t=u from a positive value to zero; the paper's analysis predicts the true flow map remains stationary as long as the diagonal mass is positive, and one can measure whether the 10-step FID degrades as the mass goes to zero.
  • Since SGFlow learns only the forward map, pairing it with a separately learned approximate inverse (or using the likelihood substitution the authors mention) would turn it into a density-estimation method, which is a direction the paper notes but does not pursue.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. SGFlow is a method for learning the two-time probability-flow ODE map f(t,u,x) = x + (u−t) f̃(t,u,x). The paper derives a squared-residual loss L in Eq. (7), rewrites the conditional velocity term, and then proposes Lsg in Eq. (11) by replacing the unknown conditional velocity v = E[Ẋ|X] with stopgrad(f̃(t,t,x)). The central claim is Theorem 1: under a time-pair sampler with positive mass on the diagonal, L and Lsg have the same stationary points, so the true flow map is the unique stationary point of the stopgrad dynamics. The paper also proves that without stopgrad, a Meanflow-style loss has nonzero variation at the true flow map in a Gaussian example, and reports CIFAR-10 FID comparisons at 1, 10, 50, and 100 sampling steps.

Significance. The contribution is potentially valuable: if Theorem 1 is correct, SGFlow is the first stopgrad-based flow-map method with a stationary-point guarantee that avoids model inverses and nested differentiation. The Gaussian counterexample in Section 4/Appendix B.2 is a clean, closed-form demonstration of why the stopgrad is needed, and the experimental comparison is focused and informative. However, the central theorem's proof is incomplete as written, and the theorem's hypotheses are not verified for the experimental protocol. The headline claim that SGFlow is the only method in the comparison with a proven stationary-point guarantee is therefore not yet supported.

major comments (3)
  1. [Appendix A.5, Theorem 1 proof, Case 2a] The sentence 'RHS-L being zero means that the velocity matches' is not justified. For the chosen h, RHS-L = 2 E[1[t=u] (Ẋ − f̃*(t,t))·(−h)] = 2 E[1[t=u] (f̃* − v)·h] after taking the conditional expectation. Thus RHS-L vanishes whenever h is orthogonal to the diagonal mismatch, not only when f̃* = v. Consequently the claim that RHS-R has the same first variation for L and Lsg does not follow; Lsg's second term uses f̃*(t,t), and if this differs from v the two variations differ. The contrapositive in Case 2 is therefore unproven. The gap is fixable (e.g., split on whether the diagonal velocity matches; if not, construct a diagonal-supported direction), but as written Theorem 1 does not establish the unique-stationary-point guarantee, which is the paper's central theoretical claim.
  2. [Section 3 / Theorem 1 and Section 5] Theorem 1 requires q(t,u) to have positive probability on t=u. The paper never specifies q(t,u) in the experiments. If the implementation uses a continuous time-pair distribution, the theorem's premise is false and the 'SG theory' entry in Table 2 does not apply to the evaluated method. Either the experiments must use a sampler with an atom at t=u, or the theorem must be proven under the actual sampling distribution. Without this, the claimed stationary-point guarantee is not connected to the reported results.
  3. [Appendix A.4 / A.5] Lemma 1 proves velocity matching by constructing arbitrary C^1 directions (soft indicators) and requiring the first variation to vanish for all such directions. This is valid only if F is the full C^1 function space and stationarity is defined with respect to that space. The paper's language 'with respect to F' and the use of the claim for the trained method in Tables 1 and 2 do not make explicit whether the stationary-point result is for the unconstrained function class or for the actual parametric model family. For a finite-dimensional UNet family, stationarity along parameter-tangent directions does not imply stationarity along arbitrary C^1 directions, so the theorem does not, as stated, transfer to the trained model. The authors should state the intended reading and, if the claim is about the parametric family, supply a tangent-space argument.
minor comments (4)
  1. [Section 3, Eq. (9)] The equality L|_{t=u} = E||f̃ − Ẋ||² is false as written; the correct value is E||f̃ − v||², up to the conditional variance E||Ẋ − v||² which is constant in f̃. The minimizer statement is unaffected, but the displayed equality is misleading and should be corrected.
  2. [Section 5.2, Table 2] FID values are reported as best over training checkpoints, with no indication of the number of seeds or variability. Since the paper's conclusions are comparative, report mean±std over at least three seeds, or explicitly state that only one seed was used.
  3. [Appendix B.1] The non-conservative example assumes that parameter values ∂_t f=0, ∂_x f=θ1, and f̃=θ2 can be realized simultaneously at a point. This is a local existence illustration; the text should say so explicitly to avoid giving the impression that it holds for all models.
  4. [Section A.5, theorem statement] The notation q(X0)q(X1) is used for the data/noise expectation while q(t,u) is used for time pairs. Consider using distinct notation (e.g., q_data, q_noise, q_time) to avoid confusing the two.

Circularity Check

0 steps flagged

No circularity: SGFlow's stopgrad objective is a design-and-verify construction; Theorem 1 is unproven but is not self-referential.

full rationale

No circular step is present. The core claim is Theorem 1 (Appendix A.5), which attempts to equate stationary points of L and Lsg. The proof is not a definitional identity: it invokes Lemma 1, which itself attempts to prove velocity matching at stationary points of L by constructing a variation. Even though the proof is open to serious correctness objections—e.g., "RHS-L being zero means that the velocity matches" (A.5 Case 2a) and "This must mean that the velocity is not matched" (A.5 Case 2b-ii) are not logically valid, since vanishing of one directional derivative does not imply pointwise velocity equality—those are rigor gaps, not circular reductions. Lsg is not defined in terms of the claimed result; the stopgrad term simply replaces v with the model's own diagonal output, and the theorem is meant to show the fixed point is the true flow map. No fitted constants are renamed as predictions; the Gaussian example is closed form. Self-citations (Singhal et al. 2023/2024; Ma et al. 2024) are background/architecture only, not load-bearing. Empirical claims are against CIFAR-10 FID. The unspecified time-pair sampler q(t,u) is another completeness issue, not circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The derivation relies on standard stochastic interpolant machinery, smoothness/boundedness of flow maps, a well-specified model family, and an atom at t=u in the time-pair distribution. The last condition is not described in the experimental setup, and the well-specifiedness of the neural family is not verified.

free parameters (1)
  • Time-pair distribution q(t,u)
    Hand-chosen distribution; the theorem requires an atom at t=u but no concrete q is reported in Section 5.2, making the guarantee conditional on an unspecified choice.
axioms (4)
  • domain assumption Velocity v(t,x)=E[X_t_dot|X_t=x] gives the probability flow ODE dx=v dt with densities satisfying the continuity equation (1).
    Section 2 builds the whole method on the stochastic-interpolant transport equation from Lipman et al./Albergo et al.; taken as given input from prior literature.
  • domain assumption The probability flow ODE has a unique integral map f satisfying (4)-(5), and this f is in the model family F_tilde.
    Used to define the target stationary point; Lemma 1 and Theorem 1 require the true map to be represented. The paper states uniqueness but does not prove it; finite neural networks generally do not contain the true map exactly.
  • ad hoc to paper q(t,u) has support on t<=u and positive probability on t=u.
    Theorem 1's statement; the experimental section never specifies q(t,u), so this premise is unverified in the implementation.
  • domain assumption Boundedness, C^1 regularity, and finite-expectation conditions needed for dominated convergence (Lemma 1 proof).
    Appendix A.4 assumes f_tilde* bounded and C^1, v in C^1, and finite expectations; not checked for neural networks.

pith-pipeline@v1.3.0-daily-deepseek · 20822 in / 17731 out tokens · 156844 ms · 2026-08-01T16:45:25.450858+00:00 · methodology

0 comments
read the original abstract

Diffusion and flow-based models benefit from simple regression losses, but inference incurs significant overhead because sampling requires integration. Consistency models address this by directly learning the flow maps along the ODE trajectory, opening a design space between one-step and many-step approaches. However, existing methods face computational challenges such as requiring model inverses or backpropagation through iterated model calls, and do not always prove that the desired ODE flow map is a solution to the loss. We introduce SGFlow, an approach for learning flow maps that bypasses explicit invertibility constraints and expensive differentiation through model iteration. SGFlow trains a model to compute both the ODE solutions and the implied velocity from scratch by following non-conservative dynamics with a stationary point at the desired flow map. On the CIFAR image benchmark, no single method attains the best FID at every step count: SGFlow attains the best FID at 10 sampling steps and remains competitive with flow matching, Meanflow, and Lagrangian map matching at other step counts, while being the only one with a proven stationary-point guarantee for its stopgrad-based dynamics.

Figures

Figures reproduced from arXiv: 2607.26398 by Aahlad Puli, Anshuk Uppal, Mark Goldstein, Raghav Singhal, Rajesh Ranganath.

Figure 1
Figure 1. Figure 1: Update vector field of the Meanflow functional over the well-specified Gaussian family, with data (µ ∗ , λ∗ ) = (0.5, 1.0) (star, ⋆). We use ∇L to denote the update field, though it is not a gradient field in the stopgrad case. Arrows are unit-normalized; color encodes ∥∇L∥ on log scale. Left: with stopgrad. The truth is the unique stationary point; arrows converge to ⋆ from all directions, and the update … 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

49 extracted references · 19 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    A variational perspective on diffusion-based generative models and score matching , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    2012 , publisher=

    Ordinary differential equations and dynamical systems , author=. 2012 , publisher=

    2012

  3. [3]

    The Journal of Machine Learning Research , volume=

    A kernel two-sample test , author=. The Journal of Machine Learning Research , volume=. 2012 , publisher=

    2012

  4. [4]

    arXiv preprint arXiv:2503.07565 , year=

    Inductive Moment Matching , author=. arXiv preprint arXiv:2503.07565 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2310.02279 , year=

    Consistency trajectory models: Learning probability flow ode trajectory of diffusion , author=. arXiv preprint arXiv:2310.02279 , year=

    Pith/arXiv arXiv

  6. [6]

    Advances in neural information processing systems , volume=

    Generative adversarial nets , author=. Advances in neural information processing systems , volume=

  7. [7]

    arXiv preprint arXiv:2202.00512 , year=

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

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2310.14189 , year=

    Improved techniques for training consistency models , author=. arXiv preprint arXiv:2310.14189 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2303.01469 , year=

    Consistency models , author=. arXiv preprint arXiv:2303.01469 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2107.00630 , year=

    Variational diffusion models , author=. arXiv preprint arXiv:2107.00630 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2410.11081 , year=

    Simplifying, stabilizing and scaling continuous-time consistency models , author=. arXiv preprint arXiv:2410.11081 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:1710.10196 , year=

    Progressive growing of gans for improved quality, stability, and variation , author=. arXiv preprint arXiv:1710.10196 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2403.03206 , year=

    Scaling rectified flow transformers for high-resolution image synthesis , author=. arXiv preprint arXiv:2403.03206 , year=

    Pith/arXiv arXiv

  14. [14]

    Machine Learning: Science and Technology , volume=

    Probability flow solution of the fokker--planck equation , author=. Machine Learning: Science and Technology , volume=. 2023 , publisher=

    2023

  15. [15]

    13th international conference, ICONIP , volume=

    Maximum mean discrepancy , author=. 13th international conference, ICONIP , volume=

  16. [16]

    The Twelfth International Conference on Learning Representations , year=

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation , author=. The Twelfth International Conference on Learning Representations , year=

  17. [17]

    Advances in neural information processing systems , volume=

    Elucidating the design space of diffusion-based generative models , author=. Advances in neural information processing systems , volume=

  18. [18]

    Journal of Machine Learning Research , volume=

    Normalizing flows for probabilistic modeling and inference , author=. Journal of Machine Learning Research , volume=

  19. [19]

    Nature , volume=

    Accurate structure prediction of biomolecular interactions with AlphaFold 3 , author=. Nature , volume=. 2024 , publisher=

    2024

  20. [20]

    arXiv preprint arXiv:2209.15571 , year=

    Building normalizing flows with stochastic interpolants , author=. arXiv preprint arXiv:2209.15571 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2303.08797 , year=

    Stochastic interpolants: A unifying framework for flows and diffusions , author=. arXiv preprint arXiv:2303.08797 , year=

    Pith/arXiv arXiv

  22. [22]

    International conference on machine learning , pages=

    Deep unsupervised learning using nonequilibrium thermodynamics , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  23. [23]

    arXiv preprint arXiv:2011.13456 , year=

    Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

    Pith/arXiv arXiv 2011

  24. [24]

    European Conference on Computer Vision , pages=

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  25. [25]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    Pith/arXiv arXiv

  26. [26]

    Advances in neural information processing systems , volume=

    Diffusion models beat gans on image synthesis , author=. Advances in neural information processing systems , volume=

  27. [27]

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

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  28. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Analyzing and improving the training dynamics of diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  29. [29]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  30. [30]

    arXiv preprint arXiv:2302.07261 , year=

    Where to diffuse, how to diffuse, and how to get back: Automated learning for multivariate diffusions , author=. arXiv preprint arXiv:2302.07261 , year=

    Pith/arXiv arXiv

  31. [31]

    International Conference on Machine Learning , pages=

    What’s the score? Automated Denoising Score Matching for Nonlinear Diffusions , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  32. [32]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    A complete recipe for diffusion generative models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  33. [33]

    Advances in Neural Information Processing Systems , volume=

    Neural flow diffusion models: Learnable forward process for improved diffusion modelling , author=. Advances in Neural Information Processing Systems , volume=

  34. [34]

    Communications in Mathematical Sciences , volume=

    Density estimation by dual ascent of the log-likelihood , author=. Communications in Mathematical Sciences , volume=

  35. [35]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  36. [36]

    arXiv preprint arXiv:2505.13447 , year=

    Mean flows for one-step generative modeling , author=. arXiv preprint arXiv:2505.13447 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2506.14603 , year=

    Align Your Flow: Scaling Continuous-Time Flow Map Distillation , author=. arXiv preprint arXiv:2506.14603 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2505.18825 , year=

    How to build a consistency model: Learning flow maps via self-distillation , author=. arXiv preprint arXiv:2505.18825 , year=

    arXiv

  39. [39]

    Journal of Statistical Mechanics: Theory and Experiment , volume=

    Learning to sample better , author=. Journal of Statistical Mechanics: Theory and Experiment , volume=. 2024 , publisher=

    2024

  40. [40]

    International Conference on Machine Learning , pages=

    The mechanics of n-player differentiable games , author=. International Conference on Machine Learning , pages=. 2018 , organization=

    2018

  41. [41]

    Journal of Machine Learning Research , volume=

    Differentiable game mechanics , author=. Journal of Machine Learning Research , volume=

  42. [42]

    Communications on Pure and Applied Mathematics , volume=

    A family of nonparametric density estimation algorithms , author=. Communications on Pure and Applied Mathematics , volume=. 2013 , publisher=

    2013

  43. [43]

    International conference on machine learning , pages=

    Variational inference with normalizing flows , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  44. [44]

    arXiv preprint arXiv:2406.07507 , year=

    Flow Map Matching , author=. arXiv preprint arXiv:2406.07507 , year=

    Pith/arXiv arXiv

  45. [45]

    Advances in neural information processing systems , volume=

    Maximum likelihood training of score-based diffusion models , author=. Advances in neural information processing systems , volume=

  46. [46]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:2209.03003 , year=

    Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. arXiv preprint arXiv:2209.03003 , year=

    Pith/arXiv arXiv

  48. [48]

    Population Matching Discrepancy and Applications in Deep Learning , volume =

    Chen, Jianfei and LI, Chongxuan and Ru, Yizhong and Zhu, Jun , booktitle =. Population Matching Discrepancy and Applications in Deep Learning , volume =

  49. [49]

    ICML 2024 Workshop on Structured Probabilistic Inference

    Bidirectional Consistency Models , author=. ICML 2024 Workshop on Structured Probabilistic Inference. 2024 , url=

    2024