acceptodds
Under review as a conference paper at ICLR 2027

Flow-Coupled Drifting for One-Step Generation

58 views

Abstract

One-step generation produces samples with a single network evaluation, but training high-quality one-step generators remains challenging. Drifting Models train such generators with attraction and repulsion fields, yet estimating these fields requires many real and generated samples per condition and therefore large batches. With only one real sample, the positive, and one other generated sample, the negative, the normalized field reduces to the positive–negative difference, which ignores where the current output lies relative to the positive and the negative. We propose Flow-Coupled Drifting (FCD) to make such single-pair training effective. First, we scale the attraction and repulsion displacements separately according to their distances from the current output. The resulting force depends on the output, and its expectation is proportional to the formal 2-Wasserstein gradient-flow velocity of a smoothed energy distance in pretrained feature spaces. Second, to pair each generator input with its own positive, we form two inputs by interpolating the same positive with independent Gaussian noises at a shared time, and the generator maps them to the output and the negative. We also train at the pure-noise endpoint used for one-step inference. On class-conditional ImageNet , FCD adapts pretrained SiT models into one-step generators, achieves an FID of 1.38, and its performance improves consistently with model size. Ablations show that flow-coupled training accelerates convergence and remains effective with a batch size of 16. The code will be made publicly available.

Primary area:
generative models
Keywords:
few-step generation, diffusion model

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_RUsbOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this work, the authors study an extension of Drifting Models, a new paradigm for one-step generation that has received much recent attention. The authors modify the usual drifting procedure in two ways:

  1. They allow it to work even with single samples from the positive and negative distributions in each batch. In particular, in Eq (8) they modify the normalizing factors in the usual prescription (Eq (4)) in a manner that permits a non-trivial velocity field even when only a single positive and negative sample exist. Interestingly, this is then shown to be a gradient on a particular energy distance between probability distributions, hinting at good convergence properties.
  2. The ability to use single samples allows for a parallel to be made to flow matching: effectively there is now an option to regress a given input towards "it's own" positive, rather than the collective halo of positives as is done for the regular drifting mechanism.

This leads to immediate empirical benefits which are well-documented in the paper. Overall I think this is a strong paper and a valuable contribution to the literature.

Critical strengths

The methodology in the paper is very clearly explained and the connection to flow-matching seems natural. The theoretical description and justification of the results in terms of global scalar energies (e.g. in Eq (12)) is well done and important for drifting models, which do not always have such guarantees.

Operationally the method seems to work very well empirically, e.g. I was quite impressed by the rapid convergence displayed in Figure 3. The strong result on ImageNet in Table 5 is also quite impressive.

Critical weaknesses

Though the method clearly works, ultimately I ended up being a bit confused by the motivation behind it: though I understand the analogy to flow matching, over there one integrates from t = 0 to t = 1 at inference time. Whereas if I have understood the proposed method of this work correctly, in this approach at inference one simply evaluates the noise -> data map at t = 1. Why then does training a t-conditioned generator help? I feel that this is important to understand despite the empirical success of the method.

Questions for authors
  1. Clarification request: it seems to me that the true motivation of lifting the single-pair degeneracy in Section 3.2 and 3.3 is to make it possible to implement the flow-matched training algorithm in Section 3.4. Is this true? If so, it would be good to emphasize this conceptual point. It is loosely suggested in the abstract that conventional drifting "requires large batches" and the new method allows for more flexibility -- should this be viewed as a significant benefit of the new method? If so, I think that Table 4 would then benefit from an ablation showing how the FID for vanilla drifting changes as one alters the batch size.

  2. Clarification request: as mentioned above, could the authors provide a bit more explanation of why the flow-coupled training method explained in Algorithm 1 works better? If I understand correctly, we do not integrate across t at inference time, so why does the having a t-conditioned extension of the generator help? A pessimistic person might have thought that all the training of the generator f(z,t) at points at t < 1 simply is never really used at inference time and is somewhat superfluous. I think this is clearly not the case, but it would be good to understand clearly why. (Perhaps an extra picture in an appendix showing how the generator on some easy-to-visualize synthetic data evolves?) Overall I think this is a very strong paper and I would just like to request some further explanation of the motivation and conceptual background of the method.

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

This paper addresses the problem that, in Drifting Models, when each generated sample is paired with only one positive and one negative, the normalized drifting field degenerates and the update direction no longer depends on the current generated sample. To address this, the paper proposes Flow-Coupled Drifting (FCD). It studies one-step image generation with a single positive and negative sample per update. It replaces the self-normalized single-pair Drifting field with separately distance-scaled attraction and repulsion, relates the expected force to a smoothed energy-distance objective, and adds a flow-coupled training branch using two noisy views of the same positive example. Starting from a pretrained SiT, the XL/2model achieves FID 1.38 on class-conditional ImageNet 256. Overall, the manuscript clearly articulates the problem, provides a relatively complete theoretical explanation, and presents competitive results, while some details still need clarification and supplementation.

Critical strengths
  • The theoretical analysis is sound. The analysis of the energy functional and the gradient relationship is mathematically careful, with explicit caveats about normalization and feature-space scope.
  • The analysis of the degeneration of the single-positive/single-negative Drifting field to the positive-negative difference is clear, and the problem definition is accurate.
  • The main experiments and ablations are sufficient, and sufficient supplementary material is provided.
  • The main empirical result is strong, and the method is tested at several model sizes and batch sizes.
Critical weaknesses
  • The energy-distance interpretation is established prior theory and needs clearer positioning. At the pure-noise endpoint, the attraction–repulsion objective is closely related to energy distance and its unbiased stochastic-gradient estimators, as in the Cramér-distance and proper-scoring-rule literature. The single-pair degeneracy diagnosed here is specific to the normalized Drifting estimator; it does not imply that distribution matching intrinsically requires many samples per condition. The force construction may be useful in this setting, but the paper should distinguish the established energy-gradient identity from what is genuinely new: the particular flow coupling, optimization recipe, and adaptation of a pretrained SiT.
  • The theory establishes an identity, but does not explain the benefit of the coupling branch. The paper proves a semi-gradient relationship and an endpoint distribution-matching statement in the non parametric setting. Yet the endpoint and earlier-time objectives act on different slices of a freely measurable time-conditioned map. Improvements at inference time must therefore be explained through shared parameters, finite-capacity effects, and training dynamics. These are precisely the mechanisms for which a theoretical or empirical explanation would be most useful. I would also welcome a characterization of what distribution the earlier-time conditional energy encourages the model to fit.
  • The main empirical comparison leaves two important confounds. FCD begins from a pretrained SiT, whereas the cited original Drifting training uses a different initialization and training history. A Drifting baseline adapted from the same SiT checkpoint would be particularly informative. Likewise, replacing the energy term with ordinary flow-matching or sample-prediction regression, while retaining the end point objective, would test whether the gain comes from the proposed coupling rather than an auxiliary denoising task.
  • Cost and evaluation need a more balanced presentation. Training involves multiple generator evaluations and large frozen feature encoders, with gradients propagated through their feature computations. The small-batch result is interesting, but batch size alone does not measure total training efficiency. GPU-hours, optimizer steps, and the cost of the SiT initialization should be reported.
  • Some ablation differences are too small to interpret from single runs. Some Results call for variability across seeds before drawing strong conclusions. The method would also benefit from a clearer description of the SiT output reparameterization and the precise meaning of one network-function evaluation at test time.
  • The rationale for the default p_{flow}=0.5 is insufficient.
  • Significance analysis and algorithm overhead are not reported.
Questions for authors
  • How do τ and γ affect stability and FID across different feature encoders, different feature scales, and large distances in early training?
  • With multiple encoders, do the forces in different feature spaces require additional normalization?
  • How does original Drifting perform when adapted from the same SiT checkpoint with a comparable compute budget?
  • What happens when the t < 1 branch is replaced by a conventional flow-matching or sample-prediction loss?
  • Can the authors characterize the population target of the conditional objective for t < 1?
  • What are the full GPU costs and encoder-independent generation metrics?
Reviewer_evLPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper builds on Drifting Models and tries to make them work with only one positive and one negative sample per generated sample. With a single pair, the normalized drifting field reduces to y+−y−y^+-y^-, which does not depend on the generated sample. The authors instead scale the attraction and repulsion terms separately by their distances, and show that in expectation this gives the gradient flow of (half) a smoothed energy distance. They also propose flow coupling: part of the time, the generator gets two noisy versions of the same real image at the same tt, and the two outputs are used as the sample and the negative. Starting from pretrained SiT, FCD gets one-step FID 2.45 / 1.72 / 1.38 for B/L/XL on ImageNet 256. In the 20-epoch XL/2 ablation, batch size 16 gives 2.41 vs. 2.10 at batch size 128.

Critical strengths
  1. I don't think the loss by itself is fundemetally novel, but the whole system is interesting and usefull. The way positives and negatives are built (two noisy copies of the same real image, with the second output used as the negative) seems to be what makes single-pair, small-batch training work. It turns a pretrained multi-step model into a good one-step generator without a teacher, a discriminator, or JVPs, and the small-batch result is useful in practice.

  2. The ablations are good. Flow coupling helps a lot (2.89 to 2.10, and much faster convergence in Fig. 3), and the ablations on gamma, the repulsion term, the encoders, and batch size support the design choices.

Critical weaknesses
  1. There is no matched comparison with multi-sample Drifting. The ablations isolate the force and flow coupling given SiT initialization, and the paper does say that FCD starts from SiT while Drifting is trained from scratch. But this means Table 5 mixes the effect of the method with the effect of the initialization. At about the same size, FCD has worse FID than Drifting (L/2: 1.72 vs. 1.53), although IS and FDr⁶ are better. A Drifting baseline initialized from SiT, with the same encoders and training budget, would make it much clearer how much of the gain comes from the method.

  2. The connection to energy scores and the related generator literature is not discussed. The paper cites energy distance and MMD gradient flows, but the objective is an expected energy score (up to a factor of 2 and a constant), and generators have been trained with energy scores before. Please cite and discuss at least [1], where the energy score was introduced, and [2], which trains generators with an energy score on diffusion-noised inputs. [2] builds its samples differently (several samples from one noisy input), so it does not cover flow coupling, but it is close. I would also suggest looking for other related work on energy-score and MMD-based generator training.

[1] T. Gneiting and A. E. Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477):359–378, 2007.

[2] V. De Bortoli, A. Galashov, J. S. Guntupalli, G. Zhou, K. Murphy, A. Gretton, and A. Doucet. Distributional Diffusion Models with Scoring Rules. ICML 2025

Questions for authors
  1. Could you add Drifting initialized from SiT with the same encoders and a similar training budget, for example at B/2? A from-scratch FCD run would also be nice to see.
  2. FDr⁶ includes MAE and SigLIP, which are also used for training. Could you report FDr per encoder, and an average without the training encoders?
  3. The appendix says the theory holds before the per-encoder normalization in Eq. 17. How much does this normalization change the results?
  4. Some details are unclear: is any guidance used during training? How do you get the sample output from SiT's velocity output? Which layers, tokens/pooling, and preprocessing are used for each encoder?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.