acceptodds
Under review as a conference paper at ICLR 2027

Generative Fractional Interpolants

Abstract

Mandelbrot–Van Ness fractional Brownian motion (Type I fBM) with Hurst index interpolates between Brownian motion, which drives continuous-time diffusion models, and random straight-line paths used in flow matching as . The Hurst index controls the correlation of increments and thereby the temporal memory of the noise process. Away from , this memory makes fBM non-Markovian, motivating Markovian approximations for tractable computation. However, the Markovian approximation of fBM (MA-fBM) used in prior work exhibits increasing error as and fails to preserve the random straight-line limit. We therefore improve MA-fBM by augmenting the approximation process with the terminal value of fBM and derive the optimal approximation coefficients in closed form. The resulting -MA-fBM preserves this limiting behavior while reducing the approximation error. Building on this approximation, we introduce *Generative Fractional Interpolants (GFI)*, a generative model interpolating subdiffusive dynamics (), standard diffusion (), superdiffusive dynamics (), and random straight line paths () used in flow matching. To train our generative model, we derive a KL-divergence of the reverse process and its parameterization and train over the full range of Hurst indicies, allowing the best operating regime to be selected at sampling time. Notably, unlike flow matching, GFI can achieve straight generative trajectories from noise to data, at the cost of trajectory crossings. Empirically, this design choice yields in our experiments lower FD-DINOv2 scores than diffusion models and flow matching on CIFAR-10, the imbalanced CIFAR-10-LT dataset, and CelebA-HQ-256 in latent space.

Primary area:
generative models
Keywords:
diffusion models, fractional brownian motion, fractional noise, generative modeling

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_71k5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper tries to connect diffusion and flow matching using fractional Brownian motion, whose Hurst index controls noise smoothness and temporal correlation. Its key trick is adding the noise process’s terminal value to a finite-dimensional approximation, helping preserve the straight-line limit as the Hurst index approaches one. The main contributions are this improved approximation, closed-form approximation coefficients, and a Hurst-conditioned generative model that lets one choose the noise regime at sampling time without retraining. Verifications include mathematical analysis, approximation-error plots, toy trajectories, and image-generation experiments on MNIST, CIFAR-10 and latent-space CelebA-HQ. Some configurations slightly improve FD-DINOv2 or FID

Critical strengths

The terminal-value augmentation addresses a concrete weakness in earlier noise approximations: losing Type I fractional Brownian motion’s straight-line behavior as the Hurst index approaches one. The idea has a clear mathematical motivation, and the approximation-error plots support it. This gives the paper an interpretable contribution that could be useful beyond its experiments.

Critical weaknesses
  1. The overall evaluation is weak. The main results benchmarking on CIFAR-10 and CelebA-HQ are far from convincing because the improvements are mostly minor. The recalls of many models on CelebA-HQ are too low (around 0.3), not sure if there is an evaluation issue.

  2. Proposition 3 claims convergence to straight-line paths over the full interval [0,1], but Appendix Proposition 12 rules out that convergence for the canonical construction. The main text also uses a regularized drift that differs from the proof’s setup. Since this limit is central to the paper’s diffusion-to-flow story, the authors need a corrected theorem that actually matches their model.

  3. Missing comparisons of training and evaluation efficiency with the baselines.

Questions for authors

none

3 comments
AuthorsOpenReview ↗

Rebuttal (i)

We thank the reviewer for their feedback and address the three concerns raised by the reviewer in detail below.

Proposition 3 claims convergence to straight-line paths over the full interval [0,1]

We thank the reviewer for noticing this mistake. We have updated our claims and proofs for arbitrary T>0T>0 and now explicitly restrict convergence in Proposition 3 to [0,S][0,S] with S<TS<T. The revised results are summarized below, with LpL^p convergence holding for every finite p≥1p\geq1 under the common-terminal coupling (Assumption 1):

  • Proposition 1: BH⟹tεB^H\Longrightarrow t\varepsilon in C([0,T])C([0,T]) for any fixed T>0T>0.
  • Proposition 9: BH→tεB^H\to t\varepsilon in Lp(C([0,T]))L^p(C([0,T])) for any fixed T>0T>0.
  • Proposition 2: XH⟹X⋆X_H\Longrightarrow X^\star in C([0,T])C([0,T]) and XH→X⋆X_H\to X^\star in Lp(C([0,T]))L^p(C([0,T])), for any fixed T>0T>0.
  • Proposition 3: XH⟹X⋆X_H\Longrightarrow X^\star in C([0,S])C([0,S]) and XH→X⋆X_H\to X^\star in Lp(C([0,S]))L^p(C([0,S])), for any fixed T>0T>0 and 0<S<T0<S<T and the generalized linear data-to-noise interpolation Xt=T−tTX(0)+tε,t∈[0,T]X_t = \frac{T-t}{T}X(0)+t\varepsilon, t\in[0,T]

We further note that the extension to arbitrary time horizons was already discussed in our original submission (lines 953–957), where we outlined the proof strategy using the self-similarity of fBM to rescale the time horizon. However, this was not correctly reflected in Proposition 3. The discrepancy happened because we initially formulated the construction with T=1T=1 and simulated up to T−δT-\delta. To evaluate all models on the same interval [0,1][0,1], we subsequently changed the implementation to T=1+δT=1+\delta for GFI(FM), without explictly updating the theoretical statement accordingly. The revised manuscript now explicitly uses S=T=1S=T=1 for GFI(VP) and S=1S=1, T=1+10−2T=1+10^{-2} for GFI(FM) (Secs. 4, 6, and E.2), with the driving noise defined on [0,T][0,T].

We apologize for the incorrect statement and thank the reviewer again for pointing it out. We hope that our revisions resolve the reviewer's concerns regarding the soundness of our theoretical contribution.

Missing comparisons of training and evaluation efficiency with the baselines.

We thank the reviewer for raising this point. The additional computations of our method contribute only a small fraction of the overall computational cost compared to neural network evaluations for models of size DiT-B/2 or larger. We have now included a runtime comparison with the baselines in Sec. K.1 and provide the results below, reporting computational cost (KFLOPs) and wall-clock time (ms) for both training and sampling. All measurements were performed on an NVIDIA RTX 5090 with 32 GB VRAM and we report the median of 30 timings of either a full forward and backward pass or a sampling step. We observe in the table below that GFI incurs only a small additional computational cost during training (+1.1% ms) and sampling (+3.43 ms), when using DiT-B/2, that amortizes over the computations of the score network.

Training stepSBDM [1] (KFLOPs)GFI(K=10) (KFLOPs)Δ\DeltaSBDM [1] (ms)GFI(K=10) (ms)Δ\Delta
DiT-S/48,633,0958,661,508+28,413 (+0.33%)18.4622.21+3.75 (+20%)
DiT-S/237,042,59437,071,004+28,410 (+0.08%)85.3889.13+3.75 (+4.4%)
DiT-B/2139,447,534139,479,007+31,474 (+0.02%)266.35269.22+2.87 (+1.1%)
Sampling stepSBDM [1] (KFLOPs)GFI(K=10) (KFLOPs)Δ\DeltaSBDM (ms)GFI(K=10) (ms)Δ\Delta
DiT-S/42,881,3422,883,757+2,415 (+0.08%)5.109.55+4.45 (+87%)
DiT-S/212,367,35612,369,770+2,415 (+0.02%)23.0227.60+4.58 (+20%)
DiT-B/246,522,05346,524,469+2,417 (+0.005%)80.1383.57+3.43 (+4.3%)

[1] Song et al. Score-Based Generative Modeling through Stochastic Differential Equations. ICLR 2021.

AuthorsOpenReview ↗

Rebuttal (ii)

The overall evaluation is weak. The main results benchmarking on CIFAR-10 and CelebA-HQ are far from convincing because the improvements are mostly minor.

We respectfully disagree with the reviewer that the reported improvements are mostly minor. Comparing GFI against the strongest diffusion and flow-matching baselines in our experiments, we observe a 17.76% relative improvement in FID on CIFAR-10 (5.80 -> 4.77, Table 1) and an 8.96% improvement on CelebA-HQ (7.92 -> 7.21, Table 2). To put these improvements into perspective, influential conceptual contributions to generative modeling have reported relative FID improvements of comparable magnitude, such as 13.3% (DDPM++ -> DDPM++cont.) and 10.2% (NCSN++ -> NCSN++cont.) for score-based generative models [1, Table 3] and 15.1% for flow matching [2, Table 1]. We believe these comparisons demonstrate that the improvements achieved by GFI are substantial, particularly when considered alongside our theoretical and conceptual contributions. Furthermore, we highlight again the low additional computational cost of GFI and the broad evaluation of baseline configurations in the appendix K.

The recalls of many models on CelebA-HQ are too low (around 0.3), not sure if there is an evaluation issue.

In the reported results, precision and recall were evaluated using the same feature extractor and checkpoint selected for FD-DINOv2, following [4]. To address the reviewer's concern and allow a better comparison with related work, we additionally evaluate precision and recall using Inception-V3 features at the checkpoints selected by FID. As shown in the table below, recall improves across all models, reaching 0.37–0.40 compared with the previously reported 0.31–0.34. Compared with LDM [3, Table 1], which reports a recall of 0.49, our models still achieve lower recall, but higher precision. Together with the precision–recall trade-off, we believe these are sufficient explanation for the reported recall values. Together with the precision–recall trade-off, we believe that this is a a reasonable explanation for the reported recall values.

MethodRecall (inceptionv3)Precision (inceptionv3)Recall (dino, Table 2)Precision (dino, Table 2)
LDM [3, Table]0.49*0.72--
Diffusion FM0.380.760.330.78
VP0.370.770.320.80
FM0.390.730.340.75
GFI$_{FM(K=7)}$0.390.760.340.80
GFI$_{FM(K=10)}$0.400.760.310.80

*calculated between the 28k training images and 50k generated images, while we use only 28k generated images.

[1] Song et al. Score-Based Generative Modeling through Stochastic Differential Equations. ICLR 2021.

[2] Lipman et al. Flow Matching for Generative Modeling. ICLR 2023.

[3] Rombach et al. High-Resolution Image Synthesis with Latent Diffusion Models. CVPR 2022.

[4] Stein et al. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. NeurIPS 2023.

[5] Gao et al. Diffusion Meets Flow Matching: Two Sides of the Same Coin, ICLR Blogpost Track 2025.

[6] Nobis et al. Generative Fractional Diffusion Models. NeurIPS 2024.

Presentation: 1: Difficult to follow despite careful reading.

We have revised our manuscript, particularly Section 4 and the appendix, to improve the clarity and readability. We hope these revisions address the reviewer's concerns regarding the presentation of our work.

Lastly, we would like to point the reviewers attention to the qualitative results included now in Appendix B.4, the new abaltion regarding few step samplings from GFI and additional baselines in our rebuttal to Reviewer Lz1L.

AuthorsOpenReview ↗

Rebuttal (iii)

We thank the reviewer again for identifying a mistake in our theoretical contribution. We hope that our revisions have addressed their concerns and would appreciate a reconsideration of their assessment. We would be happy to clarify any remaining questions.

Reviewer_GfxmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Generative Fractional Interpolants, using Type I fractional Brownian noise to connect rough and smooth noising regimes. Its epsilon-MA-fBM approximation adds a terminal-noise component to an OU-based construction. An augmented Gaussian formulation enables score-based training, with one network conditioned on the Hurst index and the index selected at sampling time. Experiments cover MNIST, CIFAR-10, CIFAR-10-LT, and latent-space CelebA-HQ.

Critical strengths
  1. Adding the terminal-noise component is a useful construction for retaining the smooth-driver limit of Type I fBM. The closed-form projection coefficients and data-dimensional learned score make the augmented framework tractable. Training one H-conditioned network also provides a practical way to explore multiple noise regimes without separate training for every H.

  2. The empirical study explores many baseline and driver configurations within a common DiT framework. Tables 7–8 show improvements over Type II across multiple configurations. On CelebA-HQ, Table 13 reports that GFI-FM with K=7 improves FD-DINOv2 from 192.23 to 176.63 and same-checkpoint FID from 9.84 to 8.92 relative to Diffusion FM. These are useful positive results for the complete recipe.

Critical weaknesses
  1. The central straight-path guarantee is internally inconsistent. Proposition 3 / Eq. (15) claims convergence in C([0,1]), whereas Theorem 2 in Appendix C.6 establishes convergence only on fixed intervals [0,T] with T<1. Proposition 12 explicitly shows that the canonical process has no continuous extension to t=1. Furthermore, the main text fixes a positive delta in the drift, while the appendix uses delta=0. With the main-text drift, the initial-signal coefficient is (1+delta-t)/(1+delta), rather than 1-t. The appendix therefore does not justify the main statement as written. Since generation starts at the noise end of the interval, the actual cutoff, regularization and joint prior are consequential. The acknowledged gap between the exact and approximate forward processes further leaves unclear which straight-path guarantee applies to the implemented sampler.

  2. The practical benefit of the proposed path geometry is not sufficiently isolated. The driver experiments compare Type I with terminal augmentation against Type II, and changing H also changes the residual noise levels and training weights. The existing comparisons do not separate these effects from the sampling geometry, and no step-count/quality study establishes a practical advantage of straighter paths. The performance trade-offs also need clearer interpretation: the best FD-DINOv2 configuration on CIFAR-10 has worse same-checkpoint FID and recall than the best FD-DINOv2 baseline (Table 10). On CIFAR-10-LT, the appendix reports FD-DINOv2 140.69 for GFDM, versus 156.52 for the highlighted GFI configuration without a memorization penalty (Tables 11 and 15). The resulting advance over the closest fractional baseline remains insufficiently explained.

Questions for authors
  1. How should Proposition 3 be amended to agree with Theorem 2 and Proposition 12? Please specify the forward coefficients, terminal regularization/cutoff and initialization used in the image experiments, and state the path guarantee that applies to that process.

  2. Under Eq. (24) and Appendix F, with consistent covariance coefficients, it appears that the residual U=X−α⊤YU=X-\alpha^\top Y satisfies, in reverse time ss with t=T−st=T-s,

    dUˉs=[−μ(t)Uˉs+λ(t)2sθ(Uˉs,t,H)] ds+λ(t) dWˉs.d\bar U_s=[-\mu(t)\bar U_s+\lambda(t)^2s_\theta(\bar U_s,t,H)]\,ds+\lambda(t)\,d\bar W_s.

    Do the auxiliary drift terms cancel this way for the full reverse SDE used in your implementation? If so, how should its advantage be understood relative to this effective-noise Brownian representation? This question concerns the full-score SDE, not an automatic extension to the augmented probability-flow ODE.

  3. For one representative existing checkpoint, could you compare quality versus NFE using the actual sampler and, if applicable, the residual SDE above? Please specify the terminal residual SNR, covariance stabilization, time bounds and final output mapping, so that initialization and discretization effects can be distinguished.

8 comments
AuthorsOpenReview ↗

Rebuttal (i)

We thank the reviewer for the detailed read of our work and their feedback. We adress each point raised by the reviewer below.

  1. How should Proposition 3 be amended to agree with Theorem 2 and Proposition 12?

We thank the reviewer for noticing this mistake. We have updated our claims and proofs for arbitrary T>0T>0 and now explicitly restrict convergence in Proposition 3 to [0,S][0,S] with S<TS<T. The revised results are summarized below, with LpL^p convergence holding for every finite p≥1p\geq1 under the common-terminal coupling (Assumption 1):

  • Proposition 1: BH⟹tεB^H\Longrightarrow t\varepsilon in C([0,T])C([0,T]) for any fixed T>0T>0.
  • Proposition 9: BH→tεB^H\to t\varepsilon in Lp(C([0,T]))L^p(C([0,T])) for any fixed T>0T>0.
  • Proposition 2: XH⟹X⋆X_H\Longrightarrow X^\star in C([0,T])C([0,T]) and XH→X⋆X_H\to X^\star in Lp(C([0,T]))L^p(C([0,T])), for any fixed T>0T>0.
  • Proposition 3: XH⟹X⋆X_H\Longrightarrow X^\star in C([0,S])C([0,S]) and XH→X⋆X_H\to X^\star in Lp(C([0,S]))L^p(C([0,S])), for any fixed T>0T>0 and 0<S<T0<S<T and the generalized linear data-to-noise interpolation Xt=T−tTX(0)+tε,t∈[0,T]X_t = \frac{T-t}{T}X(0)+t\varepsilon, t\in[0,T]

We further note that the extension to arbitrary time horizons was already discussed in our original submission (lines 953–957), where we outlined the proof strategy using the self-similarity of fBM to rescale the time horizon. However, this was not correctly reflected in Proposition 3. The discrepancy happened because we initially formulated the construction with T=1T=1 and simulated up to T−δT-\delta. To evaluate all models on the same interval [0,1][0,1], we subsequently changed the implementation to T=1+δT=1+\delta for GFI$_{\mathrm{FM}}$, without explictly updating the theoretical statement accordingly.

Please specify the forward coefficients, terminal regularization/cutoff and initialization used in the image experiments, and state the path guarantee that applies to that process.

The revised manuscript now explicitly uses $ S=T=1 $ for GFI(VP) and S=1S=1, T=1+10−2T=1+10^{-2} for GFI(FM) (Secs. 4, 6, and E.2), with the driving noise defined on [0,T][0,T]. Morever, for GFI(FM) driven by fBM we use μ:[0,S]×[1/2,1),(t,H)↦−(T−t)−1\mu:[0,S]\times [1/2,1), (t,H) \mapsto-(T-t)^{-1} and g:[0,S]×[1/2,1)→R, (t,H)↦cHTHt1−H(T−t)−Hg:[0,S]\times[1/2,1)\to\mathbb{R},\,(t,H)\mapsto c_HT^{H}t^{1-H}(T-t)^{-H}, where cHc_H is given in eq. (63). We restrict to H≥1/2H\geq 1/2 for simplicity as we onlye use the explicit gg in our convergence proof for H↑1H\uparrow 1. In Sec. 5, we use g^=g\hat{g}=g for the approximate forward process with GFI(VP)$ dynamics. For GFI(FM)$ we construct g^\hat{g} by numerically solving an ODE to preserve the one-time marginals of the linear data-to-noise interpolation, as described in Sec. E.2. With these choices of SS and TT, the convergence towards a random ODE now follows for the exact forward process of GFI(VP) from Propositon 2 and the convergence of the exact forward process of GFI(FM) towards liner data-to-noise interpolation paths follows from Proposition 3. As notet in our limitation section, we do not proof convergence of the approximate processes towards the solution of the fBM-driven SDE. We initialze the backward process for GFI(VP) at 1=S=T1=S=T and for GFI(FM) with 1=S<T=1+δ=1+1e−21=S<T=1+\delta=1+1e-2 and simulate in all reported experiments from our intial submission over N=1000N=1000 steps over the interval [1e-3,S].

We apologize for the incorrect statement and thank the reviewer again for pointing it out. We hope that our revisions resolve the reviewer's concerns regarding our theoretical contribution.

On CIFAR-10-LT, the appendix reports FD-DINOv2 140.69 for GFDM, versus 156.52 for the highlighted GFI c

We thank the reviewer for identifying this inconsistency and apologize for the oversight. Initially, Table 2 presented a subset of our best-performing configurations. When we subsequently introduced the memorization penalty, the best eligible GFI configuration no longer outperformed GFDM on CIFAR-10-LT, but we inadvertently overlooked this change. We have corrected the corresponding claim of performance gain in the abstract of our revised manuscript.

AuthorsOpenReview ↗

Rebuttal (ii)

Under Eq. (24) and Appendix F, with consistent covariance coefficients, it appears that the residual U=X−α⊤YU=X-\alpha^\top Y satisfies, in reverse time s with t=(T-s) dUs=−μ(s)Usdt+λ2(t)sθ(Us,t,H)ds+λ(t)dWsdU_s = -\mu(s)U_s dt + \lambda^2(t)s_\theta(U_s,t,H)ds + \lambda(t)dW_s Do the auxiliary drift terms cancel this way for the full reverse SDE used in your implementation? If so, how should its advantage be understood relative to this effective-noise Brownian representation? This question concerns the full-score SDE, not an automatic extension to the augmented probability-flow ODE.

We thank the reviewer for their careful consideration of our work and for raising this interesting question.

All reported results in our initial submission used Euler-Maruyama to simulate the augmented process Z=(X,Y)Z=(X,Y) with d(K+1)d(K+1) dimensions (augmented dynamics) backward in time, not the dynamics that the reviewer proposes. We call the proposed dynamics in the following "residual dynamics", taking values in Rd\mathbb{R}^d, where dd is the dimensionality of the data distribution p0p_0. The stated residual dynamics are correct, and we have included the derivations in Appendix G.2 (former Appendix F). Most importantly, it is possible to sample from p0p_0 using the proposed residual dynamics, since U0=X0U_0=X_0 by independence and the preservation of the one-time marginals. We thank the reviewer for suggesting this interesting idea for an alternative sampler for GFI. Before discussing the results further below, we would like to make three remarks.

(1) Appendix G.1 is formulated at the level of one-time marginals, rather than path distributions. Since one can find a Brownian-driven SDE that matches the one-time marginals of a process of the form X(t)=α(t)X(0)+σ(t)εX(t)=\alpha(t)X(0)+\sigma(t)\varepsilon, where ε∼N(0,1)\varepsilon\sim\mathcal{N}(0,1) [5], this is also possible for the process UU. Similarly, for classical continuous-time diffusion models, sampling from the PF-ODE or the reverse-time SDE yields the same one-time marginals but different path distributions, which lead to different behavior when discretized. In Figure 12 & 13 (Figure 8 & Figure 9 in our initial submission) we already included the path of U (orange) to showcase that the path distribution of U is very different from the path distribution of XX (black), however, we did not consider to sample from this path distribution.

(2) The one-time marginals of UU are, to the best of our authors knowledge, not available in closed form. Instead, we first solve an ODE for the covariance matrix of ZZ and subsequently use it to compute the variance of X∣YX|Y. This variance is implicitly determined by the stochastic dynamics of GFI and, to the best of our knowledge, can not be realized with a classical diffusion model.

(3) Hence, from (1)&(2), we believe that the observed perforamance gains (in our initial submission and in the new results below) can be attributed to the following properties of GFIs stochastic dynamics: (i) favourable one-time marginals of X∣YX|Y over the one-time marginals used in classical diffusion models (ii) a favourable path distribution of X (for augmented dynamics) or U (for residual dynamics) of GFI over realizable path distributions for classical diffusion models (iii) a mixture of favourable path distributions and favourable one-time marginals. Since we empirically observe an advantage of the regime H∈(1/2,1)H\in(1/2,1) in our experiments, in line with the finding of [6], we would be surprised if (1) is the advantage of GFI. We discuss this now in our conclusion and pose this as the central open question for future reseach.

[5] Gao et al. Diffusion Meets Flow Matching: Two Sides of the Same Coin, ICLR Blogpost Track 2025.

[6] Nobis et al. Generative Fractional Diffusion Models. NeurIPS 2024.

AuthorsOpenReview ↗

Rebuttal (iii)

For one representative existing checkpoint, could you compare quality versus NFE using the actual sampler and, if applicable, the residual SDE above?

CIFAR-10

We report in Table 3 (first table below) a comparison on CIFAR-10 of augmented and residual sampler for the SDE and the corresponding PF ODE over NFE = 10,50,100,200,300,500,1000. We observe that the overall best performance is achieved by the proposed residual schedule with K=4 and H=0.73, where differences between both schedules are diminishing for larger N and GFI shows the overall best performance from NFE=100 steps onward. In Table 4 (second table below), we pair the PF ODE of both dynamics with the adaptive step-size sampler Tsit5 and observe that on CIFAR-10, the augmented dynamics perform overall best, reaching a FD-DINOv2 of 334.7 with only 171 steps.

ScheduleDynamicsSamplerHurst10501002003005001000
Diffusion FM–EM–1556.3537.5475.6440.8428.1416.3406.5
GFI VP (K=4)augmentedEM0.733865.7568.3451.3405.7387.7363.6355.8
GFI VP (K=4)residualEM0.731461.8493.9424.0382.7375.2359.4354.3
ScheduleDynamicsHurst0.10.030.010.0030.0010.0001
Diffusion FM––1133.5 (26)586.2 (38)471.4 (47)458.5 (64)447.7 (81)445.6 (108)
GFI VP (K=4)augmented0.733415.8 (81)2796.3 (92)644.3 (142)440.9 (180)334.7 (171)344.4 (282)
GFI VP (K=4)residual0.731033.5 (26)723.6 (26)487.4 (38)495.4 (52)468.4 (61)443.8 (100)

Celeba

On celeba in latent space we observe in Table 5 (first table below) a very similar behaviour of the augmented and the residual GFI sampler when samepled with the SDE, while for the PF-ODE paired with the adaptive step-size sampler Tsit5 in Table 6 (second table below) we observe again that the overall best performance is achieved with the augmented sampler.

ScheduleDynamicsSamplerHurst10501002003005001000
Diffusion FM–EM–2646.3326.2228.8206.9203.5200.7202.2
GFI FM (K=7)augmentedEM–3670.22498.71220.2295.1218.3193.7183.5
GFI FM (K=7)residualEM–3691.42498.21228.6295.3218.7193.9183.1
ScheduleDynamicsHurst0.10.030.010.0030.0010.0001
FM––2564.9 (26)487.7 (42)242.1 (49)218.0 (58)219.1 (70)222.6 (101)
GFI FM (K=7)augmented–2877.7 (85)1789.5 (111)457.3 (135)184.1 (132)182.7 (155)186.6 (224)
GFI FM (K=7)residual–1690.2 (38)2036.7 (38)915.6 (44)232.1 (50)220.1 (63)224.1 (95)
Reviewer_Lz1LOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Generative Fractional Interpolants (GFI), a generative model driven by Type I fractional Brownian motion whose Hurst index HH continuously interpolates between score-based diffusion (H=1/2H=1/2) and the random straight-line paths of flow matching (H→1H\rightarrow 1), turning the diffusion and flow choice into a continuous modeling knob. Because fBM is non-Markovian and admits no score-based time reversal, the authors introduce ε\varepsilon-MA-fBM, a Markovian approximation augmented with the terminal value of fBM that preserves the straight-line limit lost by the MA-fBM of prior work, derive a KL bound for the augmented reverse process, and train a single HH-conditioned model over the full Hurst range so that HH is selected at sampling time. Experiments on CIFAR-10, CIFAR-10-LT and latent CelebA-HQ-256 with DiT backbones report lower FD-DINOv2 than diffusion and flow-matching baselines, at the cost of crossing generative trajectories and no optimal-transport coupling.

Critical strengths

The paper offers an elegant and, to my knowledge, novel unification of diffusion and flow matching: rather than viewing them as two realizations of the same marginals, it connects them at the level of path distributions through a single, physically interpretable knob, the Hurst index of Type 1 fBM. Flow matching emerges as the fully persistent-noise limit (H→1H\rightarrow 1) of a diffusion model, and the framework additionally exposes a sub-diffusive regime (H<1/2H<1/2) that neither paradigm covers.

Also, ε\varepsilon-MA-fBM is a small but substantive methodological contribution. It diagnoses a real deficiency of the existing Markov approximation (its error grows as H→1H\rightarrow 1, Fig. 2), fixes it with a minimal modification — adding the straight-line term to the basis — and provides the optimal coefficients in closed form. Since setting (ω1,ω2)=0(\omega_1, \omega_2)=0 recovers MA-fBM exactly, it is a strict generalization, and it is reusable in other settings that rely on fractional SDEs.

Critical weaknesses

W1. (presentation) The manuscript would benefit from a careful editing pass; several presentation issues make an already dense paper harder to follow.

  • Inconsistent time domain in Definition 1. The process is indexed as BH(I)=(BH(I)(t))t∈[0,1]B_H^{(I)} = (B_H^{(I)}(t))_{t\in [0,1]}, but the covariance in Eq. (6) is stated for t,s∈[0,T]t, s \in [0, T] with TT undefined at that point. Similar confusion between the two time domains appears elsewhere in the paper, and I would ask the authors to carefully check this throughout.
  • Broken cross-references. Unresolved “??” references appear at line 333 (Prop. 5, “given in ??”) and line 1999 (“as defined in ??”), and a missing citation key “(?, Theorem 2.4)” appears around line 1872.
  • Notation table. Given the large number of symbols introduced across the main text and appendix, a notation summary in the appendix would substantially improve readability.
  • Figure 1 lacks orientation cues. The caption states that noise is on the right and data on the left, but the figure itself has no time axis or labels; please annotate the noise and data ends (and the direction of time) directly in the figure.

W2.(Limited experimental evaluation) The experimental evaluation is narrower than the paper’s claims warrant.

  • Quantitative metrics only, single sampler. Results on CIFAR-10, CIFAR-10-LT and CelebA-HQ are reported exclusively through quantitative metrics; no generated samples are shown for any dataset, so sample quality cannot be assessed qualitatively. Moreover, all results use a single stochastic sampler (Euler–Maruyama, 1000 steps). Diffusion models are routinely evaluated with both stochastic and deterministic (PF-ODE) samplers and flow matching with ODE solvers, yet no deterministic-sampler results are provided — which is precisely the setting where the claimed straightness of GFI trajectories should matter.
  • A central selling point of the paper (abstract, Fig. 1, conclusion) is that GFI yields straight generative trajectories even though they cross. However, this is only illustrated on closed-form toy examples (Figs. 1, 5, 6, 8, 9); straightness is never measured on the trained image models. I would ask the authors to quantify trajectory curvature/straightness, e.g., following [1] or the straightness measure of Rectified Flow, and to compare against the diffusion and flow-matching baselines.
  • If GFI trajectories are indeed straighter, the expected practical payoff, as the authors themselves motivate via Karras et al. (2022) is better sample quality at a small number of function evaluations. Yet every experiment uses 1000 sampling steps, which hides exactly this effect. A quality-vs-NFE sweep (e.g., 10–100 steps) against the baselines under comparable solvers would be needed to substantiate the claimed benefit.

W3. (Missing baselines for the central claims)

  • No comparison with trajectory-straightening methods. Straight generative trajectories are the paper’s central selling point, yet none of the methods designed for this purpose, Rectified Flow [2], minibatch-OT flow matching [3], or curvature minimization [1] is compared against, although all three are cited in the related work. Comparing only to vanilla flow matching does not establish that “straight trajectories with crossings” is a better trade-off than existing straightening approaches.
  • No comparison with Lévy-driven models on imbalanced data. The CIFAR-10-LT experiment is motivated by the claim of [4] and [5] that Brownian-driven diffusion fails on heavy-tailed data, but those methods themselves are not included as baselines. Outperforming Brownian baselines alone does not address this motivation.

[1] Lee, Sangyun, Beomsu Kim, and Jong Chul Ye. "Minimizing trajectory curvature of ode-based generative models." International Conference on Machine Learning. PMLR, 2023.

[2] Liu, Xingchao, Chengyue Gong, and Qiang Liu. "Flow straight and fast: Learning to generate and transfer data with rectified flow." arXiv preprint arXiv:2209.03003 (2022).

[3] Tong, Alexander, et al. "Improving and generalizing flow-based generative models with minibatch optimal transport." arXiv preprint arXiv:2302.00482 (2023).

[4] Yoon, Eun Bi, et al. "Score-based generative models with Lévy processes." Advances in neural information processing systems 36 (2023): 40694-40707.

[5] Shariatian, Dario, Umut Simsekli, and Alain Oliviero Durmus. "Denoising levy probabilistic models." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors

Q1 (W2). Could the authors provide qualitative samples for CIFAR-10, CIFAR-10-LT and CelebA-HQ, and report results with a deterministic (PF-ODE) sampler alongside the Euler–Maruyama results?

Q2 (W2). Could the authors quantify trajectory straightness/curvature of the trained image models (e.g., the curvature measure of [1] or the straightness of [2]) and compare it against the diffusion and flow-matching baselines across HH?

Q3 (W2). Could the authors report a quality-vs-NFE sweep (e.g., 10, 20, 50, 100 steps) for GFI and the baselines under comparable solvers to show whether the claimed straightness translates into few-step gains?

Q4 (W3). Could the authors add Rectified Flow [2], minibatch-OT flow matching [3] and curvature minimization [1] as baselines on CIFAR-10 under the same DiT-B/2 setup?

Q5 (W3). Could the authors include the Lévy-driven models [4] and [5] as baselines on CIFAR-10-LT, since this experiment is motivated by their claims?

8 comments
AuthorsOpenReview ↗

Rebuttal (i)

We thank the reviewer for their detailed and constructive feedback. We particularly appreciate their recognition of the novelty and elegance of our unification of diffusion and flow matching, as well as the methodological contribution of our improved approximation of fBM.

W1. (presentation) The manuscript would benefit from a careful editing pass [...]

We have carefully revised the manuscript to improve the clarity and readability of our presentation, particularly in Section 4 and the appendix.

Inconsistent time domain in Definition 1. [...]

We thank the reviewer for identifying these issues. We have corrected the inconsistent time domains and broken references in the revised manuscript.

Notation table [...]

We agree with the reviewer and have added a notation table at the beginning of the appendix. We have also corrected the typos and inconsistencies identified by the reviewer and thank them for their careful reading and attention to detail.

Figure 1 lacks orientation cues.

We agree and updated Figure 1 accordingly.

Q1 (W2). Could the authors provide qualitative samples for CIFAR-10, CIFAR-10-LT and CelebA-HQ, and report results with a deterministic (PF-ODE) sampler alongside the Euler–Maruyama results? Q3 (W2). Could the authors report a quality-vs-NFE sweep (e.g., 10, 20, 50, 100 steps) for GFI and the baselines under comparable solvers to show whether the claimed straightness translates into few-step gains?

We added section B.4 with samples for CIFAR10, CIFAR10-LT and CelebaHQ-256. For each dataset we show 10 samples for 10 different Hurst exponents and samples for two of the best performing baselines for qualitative visual inspection. Moreover, we have added an evaluation of the PF-ODE on CIFAR-10 and CelebA using the adaptive Tsit5 solver. We present selected results from the corresponding tables below, including results obtained with a residual sampler for GFI, as suggested by Reviewer Gfxm and described in detail in Appendix G.2.

GFI achieves the best overall performance across all experiments reported in Table 5-8, although performance at low numbers of function evaluations (NFEs) depends on the dataset and sampling method. On CIFAR-10, GFI achieves the best FD-DINOv2 score of 334.7 with only 171 NFEs (Tables 5–6). On CelebA, GFI achieves the best FD-DINOv2 score of 182.7 with only 155 NFEs (Tables 7–8).

SDE on CIFAR-10

ScheduleDynamicsSamplerHurst10501002003005001000
Diffusion FM–EM–1556.3537.5475.6440.8428.1416.3406.5
GFI VP (K=4)augmentedEM0.733865.7568.3451.3405.7387.7363.6355.8
GFI VP (K=4)residualEM0.731461.8493.9424.0382.7375.2359.4354.3

When comparing GFI with residual SDE sampling against the best baselines for different NFE, we observe that GFI performs better than the best baseline on every tested NFE.

SDE on Celeba

ScheduleDynamicsSamplerHurst10501002003005001000
Diffusion FM–EM–2646.3326.2228.8206.9203.5200.7202.2
GFI FM (K=7)augmentedEM–3670.22498.71220.2295.1218.3193.7183.5
GFI FM (K=7)residualEM–3691.42498.21228.6295.3218.7193.9183.1

When comparing GFI with residual SDE sampling against the best baselines in SDE on Celeba) for different NFE, we observe that GFI performs better for NFE=500,1000, but worse for NFE=10,50,100,200.

PF-ODE and Tsit5 on CIFAR-10

ScheduleDynamicsHurst0.10.030.010.0030.0010.0001
Diffusion FM––1133.5 (26)586.2 (38)471.4 (47)458.5 (64)447.7 (81)445.6 (108)
GFI VP (K=4)augmented0.733415.8 (81)2796.3 (92)644.3 (142)440.9 (180)334.7 (171)344.4 (282)
GFI VP (K=4)residual0.731033.5 (26)723.6 (26)487.4 (38)495.4 (52)468.4 (61)443.8 (100)

GFI reaches a FD-DINOv2 of 334.7 with only 171 NFEs.

ScheduleDynamicsHurst0.10.030.010.0030.0010.0001
FM––2564.9 (26)487.7 (42)242.1 (49)218.0 (58)219.1 (70)222.6 (101)
GFI FM (K=7)augmented–2877.7 (85)1789.5 (111)457.3 (135)184.1 (132)182.7 (155)186.6 (224)
GFI FM (K=7)residual–1690.2 (38)2036.7 (38)915.6 (44)232.1 (50)220.1 (63)224.1 (95)

GFI reaches a FD-DINOv2 of 182.7 with only 171 NFEs.

AuthorsOpenReview ↗

Rebuttal (ii)

Straight generative trajectories are the paper’s central selling point

We would like to clarify that straight trajectories are not the central contribution of our work, but rather one special case. The main practical contribution of GFI is a generative model that supports a broad range of non-Markovian stochastic trajectories with different levels of regularity and roughness, all accessible at sampling time through a single training run (Figures 12–13, formerly Figures 8–9). To emphasize this point, we have also revised Figure 1 to illustrate the two extremes of path regularity and roughness.

Importantly, both GFI plots in Figure 1 depict trajectories generated by an SDE, not an ODE. The nearly straight trajectories at H=0.999H=0.999 illustrate the behavior established theoretically in Proposition 3: as H↑1H\uparrow1, the exact fBM-driven GFI$_{\mathrm{FM}}$ process converges to linear data-to-noise interpolation paths on every [0,S][0,S] with S<TS<T.

Q2 (W2). Could the authors quantify trajectory straightness/curvature of the trained image models (e.g., the curvature measure of [1] or the straightness of [2]) and compare it against the diffusion and flow-matching baselines across

We thank the reviewer for this suggestion. The proposed curvature metrics are primarily designed for differentiable ODE trajectories, whereas our central contribution concerns the regularity of stochastic SDE trajectories, which are almost surely non-differentiable for H<1H<1. Consequently, derivative-based curvature measures are not directly applicable to these trajectories.

For the deterministic PF-ODE, we observe that GFI performs considerably worse than flow matching with fixed-step solvers (Tables 7–8), but achieves better results than the baselines with the adaptive Tsit5 solver. We observed similar sensitivity to discretization when evaluating the proposed curvature metrics. Within the limited discussion period, we were unable to establish a reliable metric that adequately captures the regularity of GFI's stochastic trajectories. Nevertheless, we believe that the strongest practical evidence for the advantages of GFI is its sampling performance at low numbers of function evaluations (NFEs), as reported in our response to the previous point.

Could the authors add Rectified Flow [2], minibatch-OT flow matching [3] and curvature minimization [1] as baselines on CIFAR-10 under the same DiT-B/2 setup?

Due to the limited discussion period, we were able to evaluate two of the three requested baselines: minibatch-OT flow matching (OT-FM) [8] and curvature minimization (CM-FM) [7]. We have added the results to Table 16 and summarize them below. Both OT-FM and CM-FM achieve slightly better FID scores than flow matching with uniform loss weighting, but perform worse than our strongest flow-matching baseline with variance loss weighting. Our best GFI configuration achieves a substantially lower FID than all these baselines. All models were trained and evaluated under the same experimental setup, including a fixed random seed.

MethodHFID* ↓FD-DINOv2 ↓Precision ↑Recall ↑Vendi ↑
Diffusion FM–5.80365.160.670.5517.42
Diffusion VP–6.15394.780.660.5517.60
Flow matching (variance)–6.87399.170.650.5517.11
Flow matching (uniform)–8.63513.870.640.5416.92
OT-FM [8]–8.60514.380.640.5616.90
FM [7]–8.69519.760.640.5616.78
GFI VP (K=4)0.735.30316.650.690.5117.44
GFI FM (K=8)0.934.77380.810.650.5917.19

[7] Lee, Sangyun, Beomsu Kim, and Jong Chul Ye. "Minimizing trajectory curvature of ode-based generative models." International Conference on Machine Learning. PMLR, 2023.

[8] Liu, Xingchao, Chengyue Gong, and Qiang Liu. "Flow straight and fast: Learning to generate and transfer data with rectified flow." arXiv preprint arXiv:2209.03003 (2022).

AuthorsOpenReview ↗

Rebuttal (iii)

Could the authors include the Lévy-driven models [4] and [5] as baselines on CIFAR-10-LT, since this experiment is motivated by their claims?

We thank the reviewer for this suggestion. Due to the limited discussion period, we were able to train and evaluate only one of the proposed baselines and chose [4], as it is also formulated in continuous time. We have added LIM$_{\alpha=1.8}$ as a baseline in Table 4 and report the comparison below. GFI outperforms LIM in both FID and FD-DINOv2, while both models have a memorization penalty of 1.00.

MethodHKFID* ↓FD-DINOv2 ↓Mem. penalty ↓Precision ↑Recall ↑Vendi ↑
LIM (α=1.8)––14.32188.441.000.920.3419.39
GFI-FM (K=3)0.32311.51156.521.000.920.3819.51

[9] Yoon et al. Score-based Generative Models with Lévy Processes. NeurIPS 2023.

Lastly, we would like to point the reviewers attention to the low computational cost of GFI, that we reported in our rebuttal to Reviewer 71k5 and included now in K.1. We report the numbers on DiT-B/2 below.

Training stepSBDM [1] (KFLOPs)GFI(K=10) (KFLOPs)Δ\DeltaSBDM [1] (ms)GFI(K=10) (ms)Δ\Delta
DiT-B/2139,447,534139,479,007+31,474 (+0.02%)266.35269.22+2.87 (+1.1%)
Sampling stepSBDM [1] (KFLOPs)GFI(K=10) (KFLOPs)Δ\DeltaSBDM (ms)GFI(K=10) (ms)Δ\Delta
DiT-B/246,522,05346,524,469+2,417 (+0.005%)80.1383.57+3.43 (+4.3%)

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal summary

We sincerly thank the reviewers for the detailed read of our work and appreciate their constructive feedback and insightful questions. We uploaded a revised manuscript and summarize the main points of our rebuttal here:

  • We added a quantative comparison of GFI against several baselines under different NFEs for both SDE and PF-ODE sampling with multiple solvers. When using adaptive solvers, GFI achieves the overall best results on CIFAR-10 with only 171 NFEs and on Celeba with only 155 NFEs
  • We revised our proofs to formulate all convergence statements for a general time horizon T>0T>0 and corrected Proposition 3 to hold for 0<S<T0<S<T.
  • We added a computational efficiency comparison, demonstrating that GFI incurs only a small additional runtime overhead for DiT-B/2 (1.1% during training and 4.3% during sampling).
  • We included a qualitative comparison of generated images for different values of the hurst index in Sec B.3.
  • We revised in particular Sec. 3 and Sec. 4 to improve clarity of presentation and revised our introduction to better summarize our contribution.

In good spirits, The authors

AuthorsOpenReview ↗

Author follow-up

We thank the reviewer for their engagement and for taking the time to consider our rebuttal. Since the reviewer's follow-up focuses exclusively on the empirical evaluation, we assume that our revisions have addressed the other concerns raised in the initial review. We would therefore appreciate it if the reviewer reconsidered their presentation and soundness ratings in light of the corrections and improvements made during the discussion period. We address the one remaining concern regarding empirical evaluation below.

improving CIFAR-10 from 5.80 to 4.77 and improving CelebA-HQ from 7.92 to 7.21 are not significant in this large-FID area

GFI reduces FID by 17.8% on CIFAR-10 and by 9.0% on CelebA-HQ. We consider this significant relative improvements, particularly given the modest additional computational overhead of 1.1% during training and 4.3% during sampling for DiT-B/2 (Tables 9 and 10). In our rebuttal, we provided examples of relative improvements reported in influential methodological works [1, 2] that are comparable in magnitude. The reviewer now additionally questions the absolute FID values, a concern not raised in the initial review. We believe it is important to distinguish between absolute performance and relative improvements. Achieving state-of-the-art FID scores requires substantial computational resources, even on CIFAR-10, which are not equally accessible to all researchers. We therefore believe that the scientific merit of a novel generative modeling framework should be assessed through comparisons under the same computational budgets, rather than through its absolute FID scores and we provide an extensive set of such baseline experiments in Appendix L. During the discussion period, we further extended these comparisons to include minibatch-OT flow matching [4], curvature minimization [5], and Lévy noise [6] as well as showing GFIs advantage when paired with an adaptive step-size sampler. (See our rebuttal to Reviewer Lz1L and Table 4 and 6)

DDPM++ and NCSN++ are also not the strongest diffusion and flow-matching baselines in 2026.

We agree that DDPM++ and NCSN++ are not the strongest architectures for generative models available today and we do not use these architectures in our experiments, but instead employ Diffusion Transformers (DiTs), as introduced in [3].

Lastly, CIFAR-10 and CelebA-HQ are relatively simple benchmarks compared with ImageNet. The latter is more commonly used in the research community.

We agree that evaluating GFI on ImageNet would provide valuable evidence of scalability. However, this concern was not explicitly raised in the initial review, and conducting such computationally demanding experiments within the remaining discussion period is not feasible. We therefore kindly ask the reviewer to consider that we have focused our rebuttal and additional experiments on addressing the weaknesses identified in their and the other reviewers initial review.

We kindly ask the reviewer to reconsider their assessment of work, taking into account all of our contributions

Our theoretical contribution is a novel unification of diffusion and flow models at the level of path distributions through the design of a generative model driven by Type I fBM. The Hurst index provides a continuous interpolation between Brownian motion and random straight-line paths, as shown in Propositions 1–3.

Our conceptual contribution is the development of an improved approximation of fBM, which may also be of interest beyond the machine learning community, and the design of a generative model with properties distinct from conventional diffusion and flow matching, as illustrated in Figure 1.

Our practical contribution is that GFI achieves lower FID and FD-DINOv2 scores than the strongest baselines in our experiments on Cifar and Celeba, including the best observed FD-DINOv2 performance with adaptive sampling, using only 171 mean NFEs on Cifar and 155 mean NFEs on Celeba.

We would be happy to clarify any remaining questions.

[1] Song et al. Score-Based Generative Modeling through Stochastic Differential Equations. ICLR 2021.

[2] Lipman et al. Flow Matching for Generative Modeling. ICLR 2023.

[3] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers, ICCV 2023.

[4] Tong, Alexander, et al. "Improving and generalizing flow-based generative models with minibatch optimal transport." arXiv preprint arXiv:2302.00482 (2023).

[5] Lee, Sangyun, Beomsu Kim, and Jong Chul Ye. "Minimizing trajectory curvature of ode-based generative models." International Conference on Machine Learning. PMLR, 2023.

[6] Yoon, Eun Bi, et al. "Score-based generative models with Lévy processes." Advances in neural information processing systems 36 (2023): 40694-40707.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.