FastFlow: Accelerating The Generative Flow Matching Models with Bandit Inference
Divya Jyoti Bajpai, Dhruv Bhardwaj, Soumya Roy, Tejas Duseja, Harsh Agarwal, Aashay Sandansing, Manjesh Kumar Hanawal
TL;DR
FastFlow is an adaptive inference framework that accelerates flow matching models by intelligently skipping denoising steps using bandit algorithms, achieving over 2.6x speedup without sacrificing quality.
Contribution
It introduces a plug-and-play method that approximates minor denoising steps and uses bandit learning to determine optimal skipping, generalizing across tasks.
Findings
Achieves over 2.6x speedup in generation tasks.
Maintains high-quality outputs comparable to full models.
Seamlessly integrates with existing flow matching pipelines.
Abstract
Flow-matching models deliver state-of-the-art fidelity in image and video generation, but the inherent sequential denoising process renders them slower. Existing acceleration methods like distillation, trajectory truncation, and consistency approaches are static, require retraining, and often fail to generalize across tasks. We propose FastFlow, a plug-and-play adaptive inference framework that accelerates generation in flow matching models. FastFlow identifies denoising steps that produce only minor adjustments to the denoising path and approximates them without using the full neural network models used for velocity predictions. The approximation utilizes finite-difference velocity estimates from prior predictions to efficiently extrapolate future states, enabling faster advancements along the denoising path at zero compute cost. This enables skipping computation at intermediary steps.…
Peer Reviews
Decision·ICLR 2026 Poster
I think the most notable point is that the method is training-free and easy to integrate into existing flow matching pipelines. The speedup figures are plausible given the cost model of flow samplers where every velocity evaluation dominates wall time. I also like recasting the step-size selection as a bandit objective, whichh directly encodes the speed-accuracy tradeoff
- I think the novelty is thinner than the paper suggests. The velocity extrapolator collapses to a two step Adams Bashforth style predictor in uniform time-step. The work should at least acknowledge this equivalence and position itself relative to, for example, PNDM [1], and other linear multistep sampling strategies already common in diffusion code bases. Empirically, a direct comparison to a simple two step predictor that still evaluates the model at checkpoints would be informative. Moreover
- the motivation to accelerate flow-matching models is reasonable. - the proposed method is plug-and-play and introduce negalectable extra costs
- missing comparisons on ImageNet 256
Unlike static acceleration schemes, FastFlow adapts to each sample’s complexity. The multi-armed bandit dynamically decides per timestep how many steps to skip, meaning simpler cases automatically run faster while complex cases get more compute. This adaptive inference is novel and ensures no one-size-fits-all schedule, leading to greater robustness across diverse inputs. The proposed framework is model-agnostic and plug-and-play, so it can be applied to existing pretrained flow-matching models
It requires some exploration to learn the optimal skipping policy. You acknowledge that the speedup may not fully materialize in the very first steps or first few samples due to this exploration phase. In practice, you mitigate this by seeding the bandit with one full generation, but if a user only generates a handful of samples, the adaptive policy might not have time to reach peak efficiency. In scenarios with very few inference runs, the benefit of FastFlow could be less pronounced. The effe
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsGenerative Adversarial Networks and Image Synthesis · Human Pose and Action Recognition · Image Enhancement Techniques