Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow Matching

Divya Jyoti Bajpai, Shubham Agarwal, Apoorv Saxena, Kuldeep Kulkarni, Subrata Mitra, Manjesh Kumar Hanawal

TL;DR

FlowCast tackles the slow inference bottleneck of Flow Matching by introducing a training-free speculative decoding framework that leverages FM's constant-velocity tendency. It reuses the model’s own velocity predictions as zero-cost drafts to extrapolate future states and validates them via a mean-squared error threshold, enabling adaptive skipping of redundant steps. The authors provide a formal error bound showing how speculative updates affect trajectory fidelity and demonstrate over 2.5× speedups across image generation, editing, and video tasks without measurable quality loss. The approach is plug-and-play for existing FM models and comes with theoretical guarantees, broad applicability, and compatibility with other acceleration methods. Practically, FlowCast enables real-time or interactive use of Flow Matching in diverse modalities while preserving high-fidelity trajectories.

Abstract

Flow Matching (FM) has recently emerged as a powerful approach for high-quality visual generation. However, their prohibitively slow inference due to a large number of denoising steps limits their potential use in real-time or interactive applications. Existing acceleration methods, like distillation, truncation, or consistency training, either degrade quality, incur costly retraining, or lack generalization. We propose FlowCast, a training-free speculative generation framework that accelerates inference by exploiting the fact that FM models are trained to preserve constant velocity. FlowCast speculates future velocity by extrapolating current velocity without incurring additional time cost, and accepts it if it is within a mean-squared error threshold. This constant-velocity forecasting allows redundant steps in stable regions to be aggressively skipped while retaining precision in complex ones. FlowCast is a plug-and-play framework that integrates seamlessly with any FM model and requires no auxiliary networks. We also present a theoretical analysis and bound the worst-case deviation between speculative and full FM trajectories. Empirical evaluations demonstrate that FlowCast achieves $>2.5\times$ speedup in image generation, video generation, and editing tasks, outperforming existing baselines with no quality loss as compared to standard full generation.

FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow Matching

TL;DR

FlowCast tackles the slow inference bottleneck of Flow Matching by introducing a training-free speculative decoding framework that leverages FM's constant-velocity tendency. It reuses the model’s own velocity predictions as zero-cost drafts to extrapolate future states and validates them via a mean-squared error threshold, enabling adaptive skipping of redundant steps. The authors provide a formal error bound showing how speculative updates affect trajectory fidelity and demonstrate over 2.5× speedups across image generation, editing, and video tasks without measurable quality loss. The approach is plug-and-play for existing FM models and comes with theoretical guarantees, broad applicability, and compatibility with other acceleration methods. Practically, FlowCast enables real-time or interactive use of Flow Matching in diverse modalities while preserving high-fidelity trajectories.

Abstract

Flow Matching (FM) has recently emerged as a powerful approach for high-quality visual generation. However, their prohibitively slow inference due to a large number of denoising steps limits their potential use in real-time or interactive applications. Existing acceleration methods, like distillation, truncation, or consistency training, either degrade quality, incur costly retraining, or lack generalization. We propose FlowCast, a training-free speculative generation framework that accelerates inference by exploiting the fact that FM models are trained to preserve constant velocity. FlowCast speculates future velocity by extrapolating current velocity without incurring additional time cost, and accepts it if it is within a mean-squared error threshold. This constant-velocity forecasting allows redundant steps in stable regions to be aggressively skipped while retaining precision in complex ones. FlowCast is a plug-and-play framework that integrates seamlessly with any FM model and requires no auxiliary networks. We also present a theoretical analysis and bound the worst-case deviation between speculative and full FM trajectories. Empirical evaluations demonstrate that FlowCast achieves speedup in image generation, video generation, and editing tasks, outperforming existing baselines with no quality loss as compared to standard full generation.
Paper Structure (19 sections, 2 theorems, 25 equations, 13 figures, 7 tables, 1 algorithm)

This paper contains 19 sections, 2 theorems, 25 equations, 13 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related works
  3. Setup
  4. Speculative Decoding Framework
  5. Flow Matching Overview
  6. Speculative Decoding for Flow models
  7. Analysis
  8. Experiments
  9. Results
  10. Conclusion
  11. Ethics and Reproducibility statement
  12. Appendix
  13. Proof of Lemma \ref{['thm:lemma']}
  14. Proof of Theorem \ref{['thm:theorem']}
  15. Multi-iteration Editing
  16. ...and 4 more sections

Key Result

Lemma 4.1

Let $x(t)$ be the solution of $\frac{dx}{dt} = v(x,t), x(0)=x_0,$ where $v$ is Lipschitz in $x$ with constant $M$. Assume: 1) $\|x"(t)\|\leq N$ for all $t\in[0,1]$, 2) $\|\tfrac{\partial v}{\partial x}(x,t)\|\leq M$ for all $(x,t)$, 3) a forward Euler scheme with step size $h$ is run for $k$ steps,

Figures (13)

  • Figure 1: Comparison of normal vs. speculative image generation: The left side is the normal iterative process of generation in FM model. The right side is our approach which introduces intermediate speculative drafts that are rapidly proposed and verified in parallel by the backbone, enabling faster sampling while preserving quality.
  • Figure 2: GEdit Scores across three models: BAGEL, FLUX, and Step-Edit. Each subfigure reports semantic consistency (G_SC), perceptual quality (G_PQ), and overall score (G_O) versus speedup.
  • Figure 3: Results on Hunyuan Video model where we report the Vbench score and specifically for quality we use the BRISQUE metrics that assess the quality of individual frames.
  • Figure 4: Speculative step reduction maintains image quality across models, whereas direct step reduction leads to noticeable degradation.
  • Figure 5: Speculative step reduction maintains edit fidelity and consistency better than direct step reduction.
  • ...and 8 more figures

Theorems & Definitions (3)

  • Lemma 4.1
  • Theorem 4.2
  • proof