acceptodds
Under review as a conference paper at ICLR 2027

Y-Shaped Generative Flows

Abstract

Modern continuous-time generative models typically induce V-shaped flows: each sample travels independently along a nearly straight trajectory from the prior to the data. Although effective, this independent movement overlooks the hierarchical structures that exist in real-world data. To address this, we introduce Y-shaped generative flows, a framework in which samples travel together along shared pathways before branching off to target-specific endpoints. Our formulation is theoretically justified, yet remains practical, requiring only minimal modifications to standard velocity-driven models. We implement this through a scalable, neural network-based training objective. Experiments on synthetic, image, and biological datasets demonstrate that our method recovers hierarchy-aware structures, improves distributional metrics over strong flow-based baselines, and reaches targets in fewer steps.

Primary area:
generative models
Keywords:
generative models, flow matching, continuous normalizing flows, neural odes, optimal transport, branched optimal transport, benamou-brenier, single-cell rna-seq

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1dbPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new flow-based generative framework based on Y-shaped flows. The idea is to make samples travel together along shared pathways before branching off to target-specific endpoints. The formulation seems to be justified, and the proposed method is claimed to be practical and scalable. Experiments on various datasets are provided to demonstrate the effectiveness of the proposed method.

Critical strengths
  • Overall the paper is easy to follow and the proposed method is well motivated
  • Sufficient background and related work are provided
Critical weaknesses

Main weaknesses:

  • Although the idea is interesting, practical implementation seems computationally costly since it is not simulation-free (as in flow matching) and a Neural ODE needs to be integrated accurately for each mini-batch during training
  • Detailed algorithms and runtime complexity are not provided to better assess the completeness and computationally efficiency of the method. Although wall-clock cost is mentioned briefly, there are no detailed comparisons with existing methods
  • The presentation needs to be improved to enhance clarity on the method proposed. The practical method proposed is essentially terminal distributional matching (with the Sinkhorn divergence loss) plus minimizing some expected cost functional (Dirichlet energy and sublinear transport cost). The paper should have highlighted this from the very beginning and then only justify its constructions via the branched transport perspective
  • It is not clear when one should use the proposed method over the existing flow based frameworks which are simpler and easier to train
  • There are no empirical comparisons with the OT-CFM of Tong et. al. (going beyond the independent couplings). It is not clear if the "CFM" in Table 1-2 refers to OT-CFM or the one with independent coupling (I-CFM)

Minor comments:

  • In line 74-77, it is written that "The pushforward density evolving under this dynamics necessarily satisfies continuity equation with the parameterized velocity field vθv_\theta:" but there is no vθv_\theta in Eq. (1)
Questions for authors
  • Why are we restricting to compact and convex domain Ω\Omega? This seems too restrictive and does not cover commonly used distributions such as Gaussian
  • How would the results compare to the ones obtained by minimizing (9) with λsink=1\lambda_{sink} = 1 and Y^=0\hat{Y} = 0? This is a basic ablation study that should be included (I could not find these in Appendix I: they are only ablating on the transport exponent and cohesion weight)
5 comments

Part 1/3

Dear Reviewer, thank you for your review.

Answers for weaknesses

W1. Training cost and practical scalability

Although the idea is interesting, practical implementation seems computationally costly since it is not simulation-free (as in flow matching) and a Neural ODE needs to be integrated accurately for each mini-batch during training

Y-Flows costs more per update than FM but can improve endpoint quality at matched training time. We differentiate through a short, fixed Euler rollout; training requires no adaptive solve to a prescribed accuracy.

TEDSim, 100-second active training budget. Methods share a 222,334-parameter network, batch size 256, float32, one RTX A6000, and 30D PCA coordinates fitted to training endpoints:

MethodW1W_{1}, 4 steps ↓W2W_{2}, 4 steps ↓W1W_{1}, 16 steps ↓W2W_{2}, 16 steps ↓Peak GPU memory (MiB)
FM3.866 ± 0.0953.948 ± 0.1063.934 ± 0.1524.060 ± 0.18123.29 ± 0.00
Y-Flows, training K=4K=43.263 ± 0.0163.326 ± 0.0263.278 ± 0.0163.339 ± 0.02555.91 ± 0.00
Y-Flows, training K=16K=163.336 ± 0.0133.403 ± 0.0183.333 ± 0.0123.398 ± 0.018154.84 ± 0.00

Values are mean ± sample SD over three seeds; KK denotes training rollout steps and columns denote Euler sampling steps. Exact empirical W1/W2W_{1}/W_{2} use fixed 256 source and 512 target cells. Active time includes complete synchronized updates and CPU overhead, excluding setup, warm-up, logging, saves, and evaluation. Checkpoints follow the first completed update at or beyond 100 seconds.

At K=4K=4, four-step W2W_{2} improves by 15.8%. Peak allocated memory is 2.40×/6.65× FM for K=4/16K=4/16, including resident data, model, and training tensors but excluding unused allocator cache and CUDA overhead.

W2. Algorithm and complexity

Detailed algorithms and runtime complexity are not provided to better assess the completeness and computationally efficiency of the method. Although wall-clock cost is mentioned briefly, there are no detailed comparisons with existing methods

Section 5's algorithm performs four operations per update:

  1. Sample independent source and target batches of size BB.
  2. Roll out the source batch for KK Euler steps, accumulating sublinear velocity action and Hutchinson estimates of the squared Jacobian Frobenius norm.
  3. Match endpoints with Sinkhorn divergence; YF-KD instead updates a critic and uses its dual loss.
  4. Backpropagate endpoint, action, and cohesion losses through the rollout and update the velocity network.

With PP probes, the rollout uses KK batch velocity evaluations and KPKP vector–Jacobian products, then differentiates through them without forming the full Jacobian. Dense Sinkhorn costs O(B2d+SB2)O(B^2d+SB^2) work and O(B2)O(B^2) pairwise storage for SS iterations in dimension dd. Retained activation memory grows with KK. YF-KD replaces Sinkhorn with critic evaluations, updates, and gradient penalties.

FM regresses velocity at sampled times without a rollout; exact minibatch OT-CFM also solves a batch assignment. Appendix I, Table 10's 2 ms per one-probe Jacobian call (20 ms over ten steps) measures only that component. W1 gives a matched-time TEDSim comparison.

W3. Presenting the practical objective directly

The presentation needs to be improved to enhance clarity on the method proposed. The practical method proposed is essentially terminal distributional matching (with the Sinkhorn divergence loss) plus minimizing some expected cost functional (Dirichlet energy and sublinear transport cost). The paper should have highlighted this from the very beginning and then only justify its constructions via the branched transport perspective

The practical objective is terminal distribution matching + sublinear velocity action + density-weighted Dirichlet regularization (Section 5). These terms respectively fit the target, favor temporally concentrated motion, and encourage nearby particles to share velocities.

Branched transport motivates shared motion, but the objectives differ: our action is linear in transported mass, whereas classical branched transport penalizes mass concavely (Appendix D). Cohesion provides a tractable bias toward coherent velocities in place of that mechanism.

Part 2/3

W4. When to prefer Y-Flows

It is not clear when one should use the proposed method over the existing flow based frameworks which are simpler and easier to train

We recommend Y-Flows when few-step sampling or coherent paths justify extra training cost; FM offers simpler, cheaper training. FFHQ latent FD favors YF-KD at 1–2 steps and OT-CFM at ≥4 (W5).

We measure local-mass cost (rewarding spatially bundled motion) on TEDSim and experimental LARRY: mouse hematopoietic differentiation profiled by single-cell RNA sequencing and heritable lineage barcodes (Weinreb et al., 2020). We fit published monocyte/neutrophil day-2-to-day-6 transport, withholding day-4 observations and lineage labels.

Local-mass cost ↓FMY-FlowsReduction
TEDSim43.73 ± 5.6329.18 ± 0.2133.3%
LARRY46.03 ± 3.6332.61 ± 0.5429.1%

The cost is Aϵ=∫01N−1∑i∥x˙i(t)∥[miϵ(t)]−1/2 dtA_\epsilon=\int_0^1 N^{-1}\sum_i\|\dot x_i(t)\|[m_i^\epsilon(t)]^{-1/2}\,dt, with miϵ(t)=N−1∑jexp⁡(−∥xi(t)−xj(t)∥2/(2ϵ2))m_i^\epsilon(t)=N^{-1}\sum_j\exp(-\|x_i(t)-x_j(t)\|^2/(2\epsilon^2)). We use N=256N=256 (including self), fixing ϵ\epsilon before evaluation at 0.2 times median pairwise distance among 256 source/256 observed-target cells.

This supports lower local-mass cost here, not classical BOT optimality or developmental-path recovery.

W5. Independent CFM and the OT-CFM comparison

There are no empirical comparisons with the OT-CFM of Tong et. al. (going beyond the independent couplings). It is not clear if the "CFM" in Table 1-2 refers to OT-CFM or the one with independent coupling (I-CFM)

We compare FM, OT-CFM, and YF-KD on 512D FFHQ female-to-male latents using matched generators and 100,000 training updates. OT-CFM uses linear paths from exact quadratic OT on equal-size uniform minibatches; YF-KD retains action and cohesion (training K=8K=8). These new experiments are separate from Table 4.

At one/two Euler steps, YF-KD lowers mean FD by 78.1%/52.0% versus OT-CFM and 92.3%/88.4% versus FM.

Held-out latent Fréchet distance (FD ↓; mean ± sample SD, three seeds):

Method1 step2 steps4 steps8 steps16 steps32 steps64 steps128 steps
FM121.69 ± 2.3365.38 ± 2.2427.90 ± 2.7312.43 ± 2.528.36 ± 2.467.65 ± 2.387.64 ± 2.357.74 ± 2.35
OT-CFM42.82 ± 0.1915.83 ± 0.196.49 ± 0.154.71 ± 0.454.72 ± 0.604.99 ± 0.705.19 ± 0.755.31 ± 0.78
YF-KD9.38 ± 0.477.60 ± 1.017.39 ± 1.077.36 ± 1.077.36 ± 1.077.36 ± 1.077.37 ± 1.077.37 ± 1.07

FD uses all 5,572 source and 4,351 target held-out latents in raw ALAE coordinates.

At four steps, YF-KD achieves endpoint RMS 0.0461 ± 0.0011, 75.9% below OT-CFM and 88.3% below FM. Evaluation uses 256 shared sources, target-centered RMS normalization, and model-specific RK4 references checked by two successive refinements.

Part 3/3

W6. Velocity notation in the continuity equation

The pushforward density evolving under this dynamics necessarily satisfies continuity equation with the parameterized velocity field

In Eq. (1), vt(x)v_t(x) denotes the learned velocity vθ(x,t)v_\theta(x,t), so the continuity equation is

∂tρt(x)+∇x⋅(ρt(x)vθ(x,t))=0.\partial_t\rho_t(x)+\nabla_x\cdot\bigl(\rho_t(x)v_\theta(x,t)\bigr)=0.

This is the continuity equation for the parameterized dynamics used in training.

Answers for questions

Q1. Compactness, convexity, and Gaussian sources

Why are we restricting to compact and convex domain

This seems too restrictive and does not cover commonly used distributions such as Gaussian

Our training algorithm directly supports Gaussian sources in Rd\mathbb R^d. It evaluates the objective from sampled trajectories without imposing compact support. Compactness and convexity specify the setting of the stated theorem; convexity concerns the ambient domain and permits multimodal distributions.

Extending the separation theorem to Rd\mathbb R^d would require finite moments, sufficient flow regularity and integrability, the stated speed bound, and a uniform weighted Poincaré inequality along the transported distributions. The practical algorithm already accommodates Gaussian sampling.

Q2. Controls without the proposed trajectory bias

This is a basic ablation study that should be included (I could not find these in Appendix I: they are only ablating on the transport exponent and cohesion weight)

Our matched ablation demonstrates cohesion's contribution to accurate few-step transport. On 30D LARRY, adding cohesion at fixed α=0.5\alpha=0.5 reduces four-step numerical endpoint RMS by 94.0% and four-step target W2W_{2} by 64.9%. Both metrics improve in all three seeds:

MethodNumerical endpoint RMS ↓Target W2W_{2} ↓
Full Y-Flows0.0820 ± 0.00510.4888 ± 0.0138
Y-Flows without cohesion1.3555 ± 0.38081.3938 ± 0.3425

Both models use training K=16K=16, identical backbone/batch size, and 2,000 updates (mean ± sample SD, three seeds). Numerical RMS uses model-specific convergence-checked RK4 references; target W2W_{2} compares four-step generated endpoints with fixed observed targets. Both normalize by target-centered RMS. This comparison isolates cohesion's benefit at the same sampling budget.

The FFHQ ablations also show strong few-step performance with linear action and cohesion. Appendix I, Table 8 reports single-seed latent FD for zero-cohesion and linear-action variants:

Configuration1 step2 steps4 steps8 steps
α=0.7, λ=0\alpha=0.7,\ \lambda=01607201.710.314.70
α=0.5, λ=0\alpha=0.5,\ \lambda=01131111.46.195.04
α=1, λ=0.01\alpha=1,\ \lambda=0.017.634.314.314.34
α=1, λ=0.1\alpha=1,\ \lambda=0.14.394.244.234.23

At (α,λ)=(1,0.1)(\alpha,\lambda)=(1,0.1), FD is already 4.39 at one step and 4.24 at two, close to 4.23 at four/eight steps. Thus, our objective retains strong few-step quality even with linear action; sublinear action supplies the additional bias toward temporally concentrated motion.

Reviewer_Tb9cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper work trains a neural ODE with an endpoint distribution loss, a sublinear velocity penalty, and a density-weighted Jacobian penalty. It gives a bound relating terminal group separation to the transport and Jacobian energies. It implements the endpoint constraint with either Sinkhorn method or a learned critic model. Experiments cover cross domains include cell differentiation, FFHQ latents, MNIST and CelebA. The experiments validate the main paper claim that it is improving sample quality at small inference budgets.

Critical strengths

(1) The method proposes a fairly direct way to learn the transport path rather than prescribe it. The derivation of the objective is concrete, and the Jacobian estimator makes it practical.

(2) The paper gives a comprehensive experimental comparison to illustrate the method's perfromance, which goes beyond illustrative two-dimensional examples. The appendix section includes inference-step curves, ablations, and raw-pixel generation. The paper also acknowledges the better converged FM result on FFHQ and the discretization sensitivity of the eight-step CelebA model.

Critical weaknesses

(1) Training is computationally expensive. Each update requires an ODE rollout, backpropagation through the solver, and spatial derivatives for the Jacobian penalty. This is substantially more involved than standard flow matching.

(2) I am not convinced that the theory and experiments establish the claimed Y-shaped mechanism. Theorem 1 limits how quickly trajectories can separate, but it does not show that moving together before branching is better than separating gradually from the start. The more direct argument in Eq. (19) assumes that (K_{\mathrm{split}}) is independent of the branching time, even though changing that time also changes the required velocity. For a fixed branching motion compressed into duration (T), the integrated Jacobian penalty scales as (1/T), rather than (T). This step needs more justification. The experiments leave similar questions. Equation (38) measures how much motion happens in the first step, rather than whether trajectories move together. Also, the linear control (\alpha=1) in Table 8 outperforms the sublinear results in Table 4, so it is unclear how much the concave velocity cost contributes. Some reported numbers also need clarification: Table 9 gives a mean speed of 9.5 under a hard bound of 5, and its bound-20 row implies a maximum of 54.4. Table 2 reports (W_1>W_2), which should not happen for standard Wasserstein distances using the same ground metric. These could be reporting or estimation issues, but they make the results harder to interpret.

Questions for authors

(1) How does the total training time and memory usage compare with standard flow matching under the same hardware and model size? Do the gains from fewer sampling steps offset the additional cost of differentiating through the ODE solver and computing the Jacobian penalty?

(2) Could you clarify why (K_{\mathrm{split}}) in Eq. (19) can be treated as independent of the branching time, and provide a controlled comparison showing that the sublinear cost actually encourages shared motion before branching?

4 comments

Part 1/4

Dear Reviewer, thank you for your review.

Answers for weaknesses

1. Training cost

(1) Training is computationally expensive. Each update requires an ODE rollout, backpropagation through the solver, and spatial derivatives for the Jacobian penalty. This is substantially more involved than standard flow matching.

Y-Flows has a higher cost per update, but both variants achieve lower endpoint W1W_{1} and W2W_{2} than FM at a matched 100-second active training budget on 30D TEDSim. Training K=4K=4 gives 15.8% lower W2W_{2} with four sampling steps.

All methods use identical 222,334-parameter networks, batch size 256, float32, and one RTX A6000. Results are mean ± sample SD over three seeds. KK denotes Euler rollout steps during training; evaluation uses four or sixteen sampling steps. Both Y variants include the velocity action, cohesion/Jacobian penalty, and Sinkhorn endpoint loss.

Endpoint quality at 100 active training seconds.

MethodW1W_{1}, 4 steps ↓W2W_{2}, 4 steps ↓W1W_{1}, 16 steps ↓W2W_{2}, 16 steps ↓
FM3.866 ± 0.0953.948 ± 0.1063.934 ± 0.1524.060 ± 0.181
Y-Flows, training K=4K=43.263 ± 0.0163.326 ± 0.0263.278 ± 0.0163.339 ± 0.025
Y-Flows, training K=16K=163.336 ± 0.0133.403 ± 0.0183.333 ± 0.0123.398 ± 0.018

Peak GPU memory during the same runs.

MethodPeak allocated GPU memory (MiB)Relative to FM
FM23.29 ± 0.001.00×
Y-Flows, training K=4K=455.91 ± 0.002.40×
Y-Flows, training K=16K=16154.84 ± 0.006.65×

Peaks include resident data, model, and training tensors allocated by PyTorch; unused allocator cache and CUDA overhead are excluded. The trade-off is better endpoint quality at matched training time, with 2.4–6.65× higher allocated memory.

FFHQ latent Fréchet distance (FD ↓).

Method1 step2 steps4 steps8 steps
FM121.69 ± 2.3365.38 ± 2.2427.90 ± 2.7312.43 ± 2.52
OT-CFM42.82 ± 0.1915.83 ± 0.196.49 ± 0.154.71 ± 0.45
YF-KD9.38 ± 0.477.60 ± 1.017.39 ± 1.077.36 ± 1.07

Y-Flows improves TEDSim endpoint quality at comparable training time, with higher memory use. On FFHQ, YF-KD achieves the lowest FD at one and two sampling steps.

Part 2/4

2.1. Transport-cost evidence and Theorem 1

(2) I am not convinced that the theory and experiments establish the claimed Y-shaped mechanism. Theorem 1 limits how quickly trajectories can separate, but it does not show that moving together before branching is better than separating gradually from the start.

Figure 4 (50D TEDSim) shows a more pronounced shared segment at α=0.5\alpha=0.5 than at α=1\alpha=1, complementing the new 30D transport-cost measurements below.

Classical BOT charges Eα(Γ)=∑emeαℓeE_\alpha(\Gamma)=\sum_e m_e^\alpha\ell_e, 0<α<10<\alpha<1, for edge masses mem_e and lengths ℓe\ell_e in a mass-conserving network (Xia, 2003). For learned particles of mass 1/N1/N, we evaluate the BOT-inspired local-mass cost

Aϵ=∫011N∑i=1N∥x˙i(t)∥[miϵ(t)]α−1 dt,miϵ(t)=1N∑j=1Nexp⁡ ⁣(−∥xi(t)−xj(t)∥22ϵ2).A_\epsilon=\int_0^1\frac{1}{N}\sum_{i=1}^N\|\dot x_i(t)\|[m_i^\epsilon(t)]^{\alpha-1}\,dt, \qquad m_i^\epsilon(t)=\frac{1}{N}\sum_{j=1}^N\exp\!\left(-\frac{\|x_i(t)-x_j(t)\|^2}{2\epsilon^2}\right).

Here miϵm_i^\epsilon includes self-mass. Y-Flows lowers this cost and improves endpoint W2W_2 on both datasets (Response 2.3):

DatasetFMY-FlowsReduction vs. FM
TEDSim43.73 ± 5.6329.18 ± 0.2133.3%
LARRY46.03 ± 3.6332.61 ± 0.5429.1%

Mean ± sample SD over three seeds at 2,000 updates; Y-Flows uses training K=16K=16. These evaluations are separate from the 100-second study.

For a coherent mass-mm bundle with internal distances ≪ϵ\ll\epsilon and negligible contributions from other bundles, miϵ≈mm_i^\epsilon\approx m, giving cost mαLm^\alpha L over distance LL. For positive masses, L>0L>0, and 0<α<10<\alpha<1,

(m1+m2)αL<(m1α+m2α)L.(m_1+m_2)^\alpha L<(m_1^\alpha+m_2^\alpha)L.

Thus equal-length sharing lowers cost, before accounting for joining or detour costs.

2.2. Branching-time scaling and a revised proof

The more direct argument in Eq. (19) assumes that (K_{\mathrm{split}}) is independent of the branching time, even though changing that time also changes the required velocity. For a fixed branching motion compressed into duration (T), the integrated Jacobian penalty scales as (1/T), rather than (T). This step needs more justification.

Fixed-motion compression gives C∗/TC_*/T, correcting Eq. (19). Theorem 1 remains valid under its stated assumptions; its proof is independent of Eq. (19). We will add the following proof of later population separation for Jλ=Aα+λCJ_\lambda=A_\alpha+\lambda C, 0<α≤10<\alpha\le1.

Let Z∼U[−a,a]Z\sim U[-a,a] and Xt=(Ft(Z),Y+ht)X_t=(F_t(Z),Y+ht), with bounded independent YY and h>0h>0. Consider smooth odd increasing maps satisfying F0(z)=zF_0(z)=z, F1(z)=f(z)F_1(z)=f(z), 0<m≤Ft′≤L0<m\le F_t'\le L, 1≤f′≤L1\le f'\le L, and any imposed speed cap, with strict expansion on a set of positive measure. Define

Vt(z)=(1−t)z+tf(z),Gt(z)=∫0zf′(r)t dr.V_t(z)=(1-t)z+tf(z),\qquad G_t(z)=\int_0^z f'(r)^t\,dr.

With b=log⁡f′b=\log f', the exact transported Jacobian energy satisfies

C(F)−C(G)=E∫01∣∂tlog⁡Ft′−b∣2dt,C(G)=Eb2<C(V)=E(f′−1)2f′.C(F)-C(G)=\mathbb E\int_0^1|\partial_t\log F_t'-b|^2dt,\qquad C(G)=\mathbb E b^2<C(V)=\mathbb E\frac{(f'-1)^2}{f'}.

Expanding the square proves the identity; (r−1)2/r>(log⁡r)2(r-1)^2/r>(\log r)^2 for r≠1r\ne1 gives strictness. Hence Jλ(G)<Jλ(V)J_\lambda(G)<J_\lambda(V) for sufficiently large λ\lambda. Weighted AM–GM gives Gt′≤Vt′G_t'\le V_t' and dG(t)<dV(t)d_G(t)<d_V(t) for 0<t<10<t<1, where dF(t)=E[Ft(Z)∣Z>0]d_F(t)=\mathbb E[F_t(Z)\mid Z>0].

For any FF with Jλ(F)≤Jλ(G)+ηJ_\lambda(F)\le J_\lambda(G)+\eta, the identity, Aα(F)≥hαA_\alpha(F)\ge h^\alpha, and Cauchy–Schwarz give

sup⁡t∣dF(t)−dG(t)∣≤aL23Aα(G)−hα+ηλ.\sup_t|d_F(t)-d_G(t)|\le\frac{aL}{2\sqrt3}\sqrt{\frac{A_\alpha(G)-h^\alpha+\eta}{\lambda}}.

Using Appendix E, Eq. (29), set BF(t)=dF(t)2B_F(t)=d_F(t)^2 and, for fixed q∈(0,1)q\in(0,1), define

τq(F)=inf⁡{t∈[0,1]:BF(t)−BF(0)≥q[BF(1)−BF(0)]}.\tau_q(F)=\inf\{t\in[0,1]:B_F(t)-B_F(0)\ge q[B_F(1)-B_F(0)]\}.

Since dGd_G increases and lies below the threshold at τq(V)\tau_q(V), the uniform bound gives τq(F)>τq(V)\tau_q(F)>\tau_q(V) for fixed η≥0\eta\ge0 and sufficiently large λ\lambda, including all minimizers when they exist.

This replaces the exact-onset claim with population separation in the stated class, also at α=1\alpha=1; Figure 4 addresses the exponent empirically.

Part 3/4

2.3. Measuring shared transport

The experiments leave similar questions. Equation (38) measures how much motion happens in the first step, rather than whether trajectories move together.

Eq. (38) measures temporal concentration. The local-mass cost in Response 2.1 additionally accounts for proximity during movement, complementing Figure 4's trajectory evidence. The lower costs reported there are accompanied by better endpoint fit in the same evaluation:

DatasetFM endpoint W2W_2 ↓Y-Flows endpoint W2W_2 ↓
TEDSim4.443 ± 0.2683.493 ± 0.011
LARRY3.597 ± 0.0172.890 ± 0.055

Mean ± sample SD over three seeds; the same 2,000-update, 30D experiments as Response 2.1, separate from Response 1's timed study. The finite-bandwidth cost supports the full-method comparison; it does not alone establish a shared trunk or isolate the exponent's contribution.

2.4. Linear-exponent control

Also, the linear control (\alpha=1) in Table 8 outperforms the sublinear results in Table 4, so it is unclear how much the concave velocity cost contributes.

The revised result in Response 2.2 also holds at α=1\alpha=1. Sufficient cohesion can favor later population separation in the stated class with either exponent. The exponent's additional geometric influence is assessed empirically in Figure 4, where the sublinear setting shows a more pronounced shared segment in the PCA-projected trajectories.

The image-quality benefit depends on the setting: the single-seed FFHQ linear controls perform well (Table 8), whereas MNIST favors sublinearity (Table 14). These quality rankings do not determine shared-transport geometry; Figure 4 provides the complementary qualitative exponent comparison.

Part 4/4

Answers for questions

Q1. Do inference savings offset training cost?

(1) How does the total training time and memory usage compare with standard flow matching under the same hardware and model size? Do the gains from fewer sampling steps offset the additional cost of differentiating through the ODE solver and computing the Jacobian penalty?

The benefit is already measurable at matched training time. On TEDSim, Y-Flows achieves 15.8% lower four-step W2W_2 at 100 active training seconds, using identical networks and hardware (Response 1). This budget includes rollout and Jacobian computation; allocated memory is 2.40–6.65× FM's across the two Y-Flows variants.

On FFHQ, YF-KD achieves latent FD 7.60±1.017.60\pm1.01 in two sampling steps, comparable to FM's best tested mean, 7.64±2.357.64\pm2.35 at 64 steps (100,000 generator updates per method; mean ± sample SD over three seeds). The critic is unnecessary at inference.

These results demonstrate better quality at matched training time on TEDSim and fewer sampling steps at comparable quality on FFHQ. Total cost savings additionally depend on deployment volume and measured sampling time at matched quality; the numerical break-even point remains to be measured.

Q2. Corrected scaling and controlled shared motion

(2) Could you clarify why (K_{\mathrm{split}}) in Eq. (19) can be treated as independent of the branching time, and provide a controlled comparison showing that the sublinear cost actually encourages shared motion before branching?

Response 2.2 supplies an exact energy comparison that resolves the scaling objection. Fixed-motion compression gives C∗/TC_*/T, so KsplitK_{\mathrm{split}} cannot be duration-independent. The revised proof establishes lower objective value than straight transport and later population separation for sufficiently large cohesion in the stated nonlinear flow class. Its validity at α=1\alpha=1 identifies cohesion's contribution.

Figure 4 separately tests the exponent: on 50D TEDSim, α=0.5\alpha=0.5 shows a more pronounced shared segment than α=1\alpha=1 in projected trajectories. The full method also reduces local-mass transport cost by 33.3% on TEDSim and 29.1% on LARRY, while improving endpoint fit (Responses 2.1 and 2.3). Together, the controlled theoretical comparison, qualitative exponent ablation, and quantitative costs support shared transport and delayed population separation within the stated settings.

Thank you again for your thoughtful review. We hope this addresses your concerns, and we would appreciate your reconsideration of the score.

Reviewer_oT7ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Y-shaped generative flows, encouraging samples to move together before separating toward different targets. It combines a sublinear velocity cost with a cohesion regularizer and an endpoint matching loss. The paper analyzes when separation can occur and reports improved few-step generation across several datasets.

Critical strengths
  1. Adding velocity coherence to continuous normalizing flows leads to an interesting formulation. The combination with a sublinear transport cost gives a different way to control the generation path, and the analysis helps explain how these two terms affect the motion.
Critical weaknesses
  1. I am still not fully convinced why we need Y-shaped flows for distribution modeling. Different paths can reach the same target distribution, so a shared transport path is not necessarily better. The method adds regularization to intermediate trajectories, but the quantitative evaluation mostly focuses on distributional distances. The visualizations show branching, but do not tell us whether these paths are more useful or more accurate. Biology seems like a reasonable application, but I would like to see whether the method actually recovers the underlying developmental paths. For image generation, the motivation is less clear, especially since the converged quality is only marginally better than FM / MF.

  2. The explanation of why Y-shaped flows help few-step generation is also not clear to me. Time compression makes fast movement over a short interval cheaper under the proposed objective, but this does not necessarily mean that fewer integration steps are sufficient. There may be a trade-off: concentrating movement into a short interval could require rapid changes in velocity, making coarse numerical integration more difficult.

  3. Training seems considerably more involved than standard flow matching. It requires ODE rollouts and backpropagation through them, Jacobian estimation, and an endpoint loss, with an additional critic for YF-KD. While the experiments cover several settings, I am still unsure whether the additional training cost would remain manageable for much larger models and higher resolutions.

Questions for authors
  1. How much training time and GPU memory does the method require compared with FM for the same number of updates? Could you also show quality versus elapsed training time, including all the additional training costs?

  2. Could you evaluate a setting where the intermediate path itself matters? For example, Tedsim provides known developmental structure. Comparing the learned paths with intermediate states or lineage information would help show what the Y-shaped regularization offers beyond matching the final distribution.

Please also address the concerns raised in the Weaknesses section. I am open to revising my rating upward if the authors provide convincing explanations and supporting evidence that adequately address these concerns.

4 comments

Part 1/4

Dear Reviewer, thank you for your review.

Answers for weaknesses

1.1. Motivation for shared paths

Different paths can reach the same target distribution, so a shared transport path is not necessarily better. The method adds regularization to intermediate trajectories, but the quantitative evaluation mostly focuses on distributional distances.

Endpoint matching alone does not favor shared paths. Our motivation is that sharing can lower branched optimal transport (BOT) cost through its concave mass penalty, while supporting better sample quality at small integration budgets. The FFHQ latent Fréchet distance (FD, lower is better) results are shown below as mean ± standard deviation over three seeds.

Method1248163264128Best reported
FM35.74±4.1526.33±3.4620.57±1.6621.13±0.5314.90±0.154.56±0.4112.76±0.584.97±0.684.56±0.41
MF35.24±10.923.56±1.5216.98±1.7513.43±2.0411.74±2.1311.05±2.1110.77±2.0810.66±2.0510.66±2.05
YF-S35.23±13.723.45±1.4020.21±1.3920.82±0.7719.58±0.6719.04±0.6418.78±0.6316.64±0.6312.64±0.63
YF-KD38.71±16.46.54±1.195.36±0.285.37±0.425.39±0.495.41±0.525.42±0.545.43±0.555.22±0.22

MF denotes Mean Flows; YF-S and YF-KD use Sinkhorn and Kantorovich-dual endpoint losses. “Best reported” reproduces the manuscript’s separate best scores, not necessarily the minima of the displayed columns.

Local-mass transport-cost evaluation. To quantify the benefit of shared transport, we compared FM and Y-Flows on simulated TEDSim data and the experimental LARRY dataset.

LARRY combines single-cell RNA sequencing with heritable lineage barcodes to track mouse hematopoietic differentiation at days 2, 4, and 6 (Weinreb et al., 2020). We use the published monocyte/neutrophil trajectory subset, learning transport from day 2 to day 6 while withholding day-4 observations and lineage labels from model fitting.

For both datasets, we evaluate a local-mass transport cost that makes motion less expensive when particles travel in spatially close bundles:

Aϵ=∫011N∑i=1N∣x˙i(t)∣,[miϵ(t)]α−1,dt,miϵ(t)=1N∑j=1Nexp⁡!(−∣xi(t)−xj(t)∣22ϵ2).A_\epsilon = \int_0^1 \frac{1}{N}\sum_{i=1}^{N} \\|\dot{x}_i(t)\\|\\,[m_i^\epsilon(t)]^{\alpha-1}\\,dt, \qquad m_i^\epsilon(t) = \frac{1}{N}\sum_{j=1}^{N} \exp\\!\left(-\frac{\\|x_i(t)-x_j(t)\\|^2}{2\epsilon^2}\right).

We use N=256N=256 particles and α=0.5\alpha=0.5, including each particle itself in its local mass. For each dataset, we fix ϵ\epsilon before model evaluation to 0.2 times the median pairwise distance in a balanced pool of 256 source and 256 observed-target cells. Results are mean ± sample SD over three training seeds.

DatasetFMY-FlowsReduction vs. FM
TEDSim43.73 ± 5.6329.18 ± 0.2133.3%
LARRY46.03 ± 3.6332.61 ± 0.5429.1%

Y-Flows achieves lower local-mass transport cost on both datasets, providing quantitative support for our shared-transport motivation in simulated and experimental settings.

Part 2/4

1.2. Path accuracy and usefulness

The visualizations show branching, but do not tell us whether these paths are more useful or more accurate. Biology seems like a reasonable application, but I would like to see whether the method actually recovers the underlying developmental paths.

Dataset and evaluation. We evaluated path accuracy on Dyngen, a stochastic gene-regulatory simulation with 135 genes, a shared trunk, two terminal fates, and 21 recorded states per trajectory. We use 30 principal components fitted to the original training endpoints. We selected α=0.05\alpha=0.05 using intermediate states of the original 256 trajectories, before evaluating on fresh realizations of the same simulated network.

We compare the predicted and ground-truth mean paths for each terminal fate using normalized symmetric curve distance. This metric measures spatial agreement by averaging squared nearest-segment distances in both directions at 256 uniformly spaced arc-length samples, including endpoints. The two fates receive equal weight, and distances are normalized by training-endpoint variance. Results are reported as mean ± sample SD across three training seeds.

Metric ↓FMY-Flows (α=0.05\alpha=0.05)Y-Flows vs. FM
Normalized symmetric curve distance0.217 ± 0.0020.176 ± 0.00719.3% lower

The lower curve distance indicates that Y-Flows more closely match the average spatial geometry of the reference developmental paths. Trajectory figure: https://anonymous.4open.science/r/Anonymous-CD2C/.

1.3. Motivation for image generation

For image generation, the motivation is less clear, especially since the converged quality is only marginally better than FM / MF.

The image experiments test feasibility and quality at small inference budgets in 512D FFHQ latents, 784D MNIST pixels, and 12,288D CelebA-64 pixels. For CelebA-64, we report FID (↓) at different numbers of inference steps below. Here, KK denotes the number of training rollout steps, and parameter counts refer to the generators.

MethodParameters2 steps4 steps8 steps16 steps32 steps
FM45.1M179.786.342.421.212.8
Y-Flows (K=10K=10)21.4M75.237.922.816.112.6
Y-Flows (K=8K=8)21.4M116.039.412.314.812.8

Y-Flows improve quality at small inference budgets, while the scores become comparable at larger budgets. FM also has a lower reported best FFHQ FD. These results support few-step generation as the main motivation for the image experiments.

2.1. Why can Y-shaped flows help few-step generation?

The explanation of why Y-shaped flows help few-step generation is also not clear to me.

The intended mechanism combines time compression, velocity cohesion, and training through finite-step rollouts. For 0<α<10<\alpha<1, the velocity action for moving distance DD at constant speed over duration TT is DαT1−αD^\alpha T^{1-\alpha}, favoring shorter movement intervals. Cohesion penalizes spatial velocity differences; the endpoint loss evaluates the training-budget rollout directly.

Cohesion can tighten an Euler-error bound. For a continuously differentiable velocity field, globally LL-Lipschitz in space and bounded by MM, a source with finite second moment, and

C=∫01E∣Dxv(t,Xt)∣F2,dt,H=∫01E∣∂tv(t,Xt)∣2,dt<∞,C=\int_0^1\mathbb E\\|D_xv(t,X_t)\\|_F^2\\,dt, \qquad H=\int_0^1\mathbb E\\|\partial_t v(t,X_t)\\|^2\\,dt<\infty,

the exact and uniform NN-step Euler endpoint laws satisfy

W2(ρ^1,N,ρ1)≤eL3,N(H+MC).W_2(\widehat\rho_1^{\\,N},\rho_1) \le\frac{e^L}{\sqrt3\\,N}\left(\sqrt H+M\sqrt C\right).

The proof couples trajectories from the same source, bounds acceleration ∂tv+(Dxv)v\partial_t v+(D_xv)v by H+MC\sqrt H+M\sqrt C in integrated L2L^2, and propagates Euler remainders using Lipschitz stability. At fixed L,M,HL,M,H, smaller CC tightens the bound; it does not guarantee smaller actual error. Cohesion alone does not control temporal variation, and compression can increase HH.

Part 3/4

2.2. Does time compression impair coarse integration?

Time compression makes fast movement over a short interval cheaper under the proposed objective, but this does not necessarily mean that fewer integration steps are sufficient. There may be a trade-off: concentrating movement into a short interval could require rapid changes in velocity, making coarse numerical integration more difficult.

Y-Flows has lower mean integration error than FM and OT-CFM at every tested budget from 4–128 steps on TEDSim and LARRY. The upper bound above explains how cohesion can help control integration error.

Entries are mean ± sample SD over three seeds of endpoint RMS error against each model’s convergence-checked RK4 reference (512–1024 steps), normalized by target centered RMS; columns give Euler steps. Y-Flows uses training K=16K=16. These 30D tests use an earlier TEDSim endpoint cohort, separate from TEDSim in Sections 1.1 and 3.1.

DatasetMethod481516173264128
TEDSimFM0.3419 ± 0.04680.2471 ± 0.05130.1947 ± 0.04930.1841 ± 0.03530.1759 ± 0.04660.1197 ± 0.02790.0786 ± 0.01550.0433 ± 0.0078
TEDSimOT-CFM0.1564 ± 0.01220.0969 ± 0.00860.0589 ± 0.00560.0557 ± 0.00530.0529 ± 0.00510.0302 ± 0.00300.0158 ± 0.00160.0081 ± 0.0008
TEDSimY-Flows0.0632 ± 0.00340.0304 ± 0.00170.0159 ± 0.00090.0149 ± 0.00080.0140 ± 0.00080.0074 ± 0.00040.0037 ± 0.00020.0018 ± 0.0001
LARRYFM0.2425 ± 0.01920.1400 ± 0.01370.0818 ± 0.00810.0773 ± 0.00760.0732 ± 0.00720.0411 ± 0.00400.0213 ± 0.00210.0108 ± 0.0011
LARRYOT-CFM0.0926 ± 0.01620.0543 ± 0.01020.0319 ± 0.00610.0301 ± 0.00570.0286 ± 0.00540.0160 ± 0.00300.0083 ± 0.00160.0042 ± 0.0008
LARRYY-Flows0.0820 ± 0.00510.0413 ± 0.00300.0221 ± 0.00180.0207 ± 0.00170.0195 ± 0.00160.0104 ± 0.00090.0052 ± 0.00050.0026 ± 0.0002

Y-Flows achieve lower integration error even below the training step budget, consistent with the role of cohesion in the upper bound above.

3.1. Training cost and scalability

It requires ODE rollouts and backpropagation through them, Jacobian estimation, and an endpoint loss, with an additional critic for YF-KD.

While the experiments cover several settings, I am still unsure whether the additional training cost would remain manageable for much larger models and higher resolutions.

On 30D TEDSim, both Y-Flows variants achieve lower W1W_1 and W2W_2 than FM at a matched 100-second active training budget. Training K=4K=4 gives 15.8% lower W2W_2 with four sampling steps.

All methods use identical 222,334-parameter networks, batch size 256, float32, and one RTX A6000. Results are mean ± sample SD over three seeds with fixed evaluation samples. K denotes training rollout steps; evaluation uses 4 or 16 sampling steps. Active time includes complete synchronized training updates; setup, warm-up, logging, checkpointing, and evaluation are excluded. Zero denotes initialization; later rows use the first completed update at or beyond each budget.

TEDSim: peak GPU memory during the same 100-second training runs.

MethodPeak allocated GPU memory (MiB)Relative to FM
FM23.29 ± 0.001.00×
Y-Flows, training K=4K=455.91 ± 0.002.40×
Y-Flows, training K=16K=16154.84 ± 0.006.65×

Mean ± sample SD over three seeds. Peaks include resident data, model, and training tensors allocated by PyTorch; unused allocator cache and CUDA overhead are excluded.

Part 4/4

3.1. Training cost and scalability (continued)

Values are mean ± sample SD over three seeds; KK denotes training rollout steps.

TEDSim: endpoint W1W_1 ↓ with 4 sampling steps.

Active training budget (s)FMY-Flows, training K=4K=4Y-Flows, training K=16K=16
04.551 ± 0.0074.551 ± 0.0074.551 ± 0.007
24.245 ± 0.0904.234 ± 0.0654.388 ± 0.039
44.179 ± 0.0913.909 ± 0.0154.191 ± 0.077
83.986 ± 0.0593.720 ± 0.0143.864 ± 0.025
163.908 ± 0.1623.479 ± 0.0543.687 ± 0.040
323.886 ± 0.1433.389 ± 0.0053.479 ± 0.029
643.867 ± 0.0393.298 ± 0.0223.388 ± 0.030
1003.866 ± 0.0953.263 ± 0.0163.336 ± 0.013

TEDSim: endpoint W2W_2 ↓ with 4 sampling steps.

Active training budget (s)FMY-Flows, training K=4K=4Y-Flows, training K=16K=16
04.735 ± 0.0104.735 ± 0.0104.735 ± 0.010
24.318 ± 0.1024.376 ± 0.0674.532 ± 0.047
44.266 ± 0.0714.023 ± 0.0164.331 ± 0.087
84.067 ± 0.0723.818 ± 0.0153.972 ± 0.026
163.997 ± 0.1823.549 ± 0.0673.777 ± 0.052
323.981 ± 0.1513.461 ± 0.0093.551 ± 0.037
643.974 ± 0.0533.361 ± 0.0323.461 ± 0.041
1003.948 ± 0.1063.326 ± 0.0263.403 ± 0.018

TEDSim: endpoint W1W_1 ↓ with 16 sampling steps.

Active training budget (s)FMY-Flows, training K=4K=4Y-Flows, training K=16K=16
04.551 ± 0.0074.551 ± 0.0074.551 ± 0.007
24.160 ± 0.1124.238 ± 0.0644.390 ± 0.038
44.196 ± 0.0213.917 ± 0.0164.195 ± 0.077
84.045 ± 0.1653.733 ± 0.0143.872 ± 0.025
163.991 ± 0.2513.494 ± 0.0563.694 ± 0.041
323.899 ± 0.2163.408 ± 0.0053.480 ± 0.034
643.855 ± 0.0413.311 ± 0.0233.386 ± 0.028
1003.934 ± 0.1523.278 ± 0.0163.333 ± 0.012

TEDSim: endpoint W2W_2 ↓ with 16 sampling steps.

Active training budget (s)FMY-Flows, training K=4K=4Y-Flows, training K=16K=16
04.735 ± 0.0104.735 ± 0.0104.735 ± 0.010
24.244 ± 0.1344.380 ± 0.0664.534 ± 0.046
44.320 ± 0.0124.033 ± 0.0164.336 ± 0.086
84.163 ± 0.2073.833 ± 0.0163.982 ± 0.025
164.118 ± 0.2933.563 ± 0.0683.785 ± 0.051
324.022 ± 0.2753.477 ± 0.0073.551 ± 0.040
643.994 ± 0.0613.372 ± 0.0323.455 ± 0.038
1004.060 ± 0.1813.339 ± 0.0253.398 ± 0.018

At the matched 100-second budget, Y-Flows achieves better endpoint quality than FM, with higher peak allocated GPU memory.

Answers for questions

Q1. Training time, memory, and quality versus elapsed time

How much training time and GPU memory does the method require compared with FM for the same number of updates? Could you also show quality versus elapsed training time, including all the additional training costs?

Response 3.1 reports peak allocated GPU memory and endpoint W1W_1 and W2W_2 versus active training time through 100 seconds for FM and both Y-Flows variants.

Q2. Intermediate paths and developmental structure

Could you evaluate a setting where the intermediate path itself matters? For example, Tedsim provides known developmental structure. Comparing the learned paths with intermediate states or lineage information would help show what the Y-shaped regularization offers beyond matching the final distribution.

Yes. Response 1.2 reports Dyngen mean-path geometry evaluation. Y-Flows improves dyngen mean-curve geometry by 19.3%.

Q3. Remaining weaknesses

Please also address the concerns raised in the Weaknesses section. I am open to revising my rating upward if the authors provide convincing explanations and supporting evidence that adequately address these concerns.

Responses 1.1–3.1 address each weakness with supporting evidence.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.