Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch

We are grateful to have received positive feedback. Some of our responses overlap with those addressed to other reviewers, and we provide cross-references for clarity and conciseness. Please feel free to raise your score if you find that your concerns have been addressed or if you see new strengths, interesting insights, or future potential in our work :) .

Q1.Compare to original short-cut model
A1. We attempted to adapt the original shortcut approach to our distillation setting, but it yielded poor convergence and low-quality outputs. For details, we kindly refer the reviewer to our response to Question 1 from Reviewer qe9Q for a detailed explanation.

Q2. Lower steps / ADD
A2. Please see our response to Reviewer VfPo, Question 2, where we address the reasons in detail.

Q3. Other baseline models/ distillation methods
A3. A detailed discussion of this can be found in our response to Question 3 from Reviewer VfPo.

Q4. Code release
A4. We commit to releasing the code following the final decision on the paper. We would also appreciate some time to further develop and refine it to ensure it is released in its best possible form. Please also see our response below for more information.

Q5. Experiment with other modalities
A5. Related to the code releasing question, we are planning to release a general-purpose training library based on our algorithm. This library is designed to support a wide range of customized modalities and models, rather than implementing each modality separately, which we find inefficient. To our knowledge, this would be the first off-the-shelf codebase that supports such flexibility. For example, in the case of video, our method requires only a few self-generated clips. It also supports distillation for modality-conditioned variants, such as the recently released FLUX-Kontext.

We thank the reviewer for the positive feedback and these constructive questions, which are valuable for clarifying key aspects of our work. We hope that our responses help clarify your questions and showcase the potential of our work. If so, we kindly invite you to consider raising your score :).

Q1. $V_\theta(x_t, t)$ can be ill-defined
A.1 To clarify, training a single velocity model $V_\theta(x, t)$ across multiple step sizes (i.e., without explicit $d$-conditioning) is well-defined within the framework of rectified flow training. Specifically, the 1-rectified flow model is derived from a family of $k>1$-rectified flows, each inducing different (and possibly curved) velocity fields. While each $k \ge1$ version has its own teacher and potentially distinct $V_\theta$, in practice, one may recover the 1-rectified version either by (1) gradually updating the model as $k$ decreases, or (2) training a unified model across all $k \ge 1$ trajectories (as we do). These two approaches, when fully optimized, should theoretically yield similar outcomes, and justify each other.

Given this view, we believe it is more precise to reframe the concern: rather than “assuming we can straighten curved velocity,” our method encourages the model to learn to straighten trajectories from curvature. In this sense, straightness is not assumed: it emerges as a learned property from training using our proposed training objective. This is achieved by rectifying curved trajectories as much as possible. Indeed, in worst-case scenarios, e.g. regions with extreme curvature, the training loss can become large and may require special treatment. In these cases, as the reviewer pointed out, the model may struggle to align all the $V_\theta(x, t)$ consistently. A more comprehensive justification of such behavior likely requires further theoretical investigation in future work.

Q2. Why not lower steps
A2. Based on our understanding on existing low-step generation methods, achieving high-quality images with extremely low sampling steps requires two key components: core algorithm and image-level feature learning. Current one/two-step generation methods rely heavily on optimizing feature learning through additional modules such Adversarial Distillation (ADD), which injects richer discriminative visual features into the model. This reflects the fact that reducing sampling steps is not solely a matter of core algorithm design (which is our focus), but also requires enhanced perceptual understanding.

For this reason, all competing baselines focus on ADD to overcome lower step challenges , and may sometimes step in their core algorithm contribution (something we aim to avoid in our approach). Nevertheless, we hypothesize that there may exist more elegant alternatives to replace ADD by focusing on more straightforward image feature learning. We are actively exploring this direction and hope to report promising results in the near future.

Additionally, we envision SCFM as a lightweight yet vital post-training refinement procedure (analogous in spirit to RLHF), that can significantly reduce sampling steps while preserving or even improving generative quality.

Q3. Not comprehensively evaluated; Other baseline models/ distillation methods
A3. Thank you for the suggestion. There are several reasons why we did not include additional experiments with methods such as ReFlow, Progressive Distillation (PD), and Consistency Distillation (CD):

• These distillation techniques have so far only been implemented on EDM-based models, while our method is specifically designed for flow-matching approaches. As a result, a direct and fair comparison is not feasible. The same reasoning applies to SD1.5, which is not pretrained using flow matching and therefore not directly comparable to our setting.
• We believe the most meaningful comparison would require a benchmark consisting of toy-level pretrained flow-matching models that could then be distilled using PD, CD, and other approaches. Creating such a benchmark would likely need broader community support to establish a shared set of FM-pretrained models.
• Our work focuses on large-scale pretrained flow-matching models. Implementing and evaluating all baselines at this scale would require computational resources that we currently cannot afford.
• We note that these distillation techniques have already been shown to perform worse than the baselines listed in Table 1 in their original papers. It is therefore reasonable to expect that our method would also outperform them under similar conditions.

Q4. Some results are not the best
A4. There may be some misunderstanding in our presentation, but as shown in Table 1, our method consistently outperforms all baselines across the reported metrics. The doubled latency in the SD3.5 experiments, as explained in lines 222–226, results from skipping the CFG distillation (unlike SD3.5L-Turbo which includes it), and consequently performing inference with double batch sizes. We will revise the phrasing to clarify this distinction and avoid any further confusion.

Q5. Model weights initialization; training/testing prompt difference
A5. The model is trained using LoRA, with the $B$ matrix initialized to zero. Therefore, in all our experiments, the model weights are indeed initialized from the pretrained backbone.

Regarding training data, the training and testing prompts are different, in other words, the model is evaluated in a fully blind setting. The few-shot learning experiments are intended to show that a general distillation capability can be obtained using a small, randomly selected dataset, or even one generated by the model itself. The benefits are obvious: (1) unlike all other methods, we do not require big training sets especially for uncommon modalities; (2) less computational resources; (3) appealing flexibility for model expansion (see our response to Reviewer SAAF for more details); (4) customizable to any self-maintained/private data requirements, and etc.

We are very grateful to the reviewer for the feedback, and very pleased to have received positive reviews.

Q1. A direct shortcut implementation
A1. We have attempted to implement the original shortcut method for comparison. However, we observed that convergence was significantly slower and the resulting performance was not competitive with any other baselines (i.e., it failed to produce reasonable outputs at low sampling steps, such as 8-12). Henthforce, applying the original shortcut formulation to large models such as FLUX or SD3.5 appears to require substantial computational resources and/or non-trivial pretraining techniques, and we are concerned that our current implementation is not sufficient to support a fair comparison.

Q2. Notation consistency/ambiguous
A2. Thank you for the suggestion! We are now aware that some notational inconsistencies and ambiguities are indeed present, and we will do a careful revision in the next version of the paper to ensure clarity.