Magnitude-Direction Decoupling for Fast Video Generation with Flow Matching Models

Haonan Xu; Feiyang Chen; Songkui Chen; Hongpeng Pan; Zhefeng Wang; Xinyu Duan; Baoxing Huai; Yang Yang

cs.CVAug 18, 2026

Magnitude-Direction Decoupling for Fast Video Generation with Flow Matching Models

Authors: Haonan Xu, Feiyang Chen, Songkui Chen, Hongpeng Pan, Zhefeng Wang, Xinyu Duan, Baoxing Huai, Yang Yang

Organizations: Nanjing University of Science and Technology, Nanjing, China · Huawei, Shanghai & Hangzhou, China

Abstract

Flow matching models for video generation achieve impressive performance but suffer from high computational overhead due to iterative denoising. In fact, the original model is not necessary for all denoising steps, allowing some steps to use lightweight alternatives for faster sampling. However, directly using caching or lightweight models can deviate from the original denoising trajectory, resulting in suboptimal performance. Through empirical analysis, we find that lightweight models can robustly capture the magnitude components of the original model's output, while caching provides reliable directional guidance. Building on this insight, we propose the Magnitude-Direction Decoupling (MDD) method, which adaptively employs a direction-calibrated lightweight model as a substitute for the original model to accelerate inference and effectively correct deviations in the denoising trajectory. Moreover, MDD further reduces inference costs by reusing magnitude information under classifier-free guidance (CFG). As a result, MDD offers a more reliable and lightweight solution to accelerate sampling. Experiments show that MDD outperforms existing acceleration methods, delivering promising speedups (e.g., up to 2.95x on Wan2.1) while preserving high visual fidelity and content richness.

Figures & tables

Panel of Figure 2 — Figure 2: Comparison of visual fidelity and sampling speed with competing methods. Latency is measured on a single A100 GPU. LPIPS denotes learned perceptual image patch similarity. Video synthesis configuration: 81 frames, 5s, 480P on Wan2.1-14B. Our method consistently generates outputs that better preserve the visual fidelity and richness of the original video content while achieving enhanced acceleration performance. The limitations of competing methods are highlighted in red boxes and are best examined under zoomed-in view.

Panel of Figure 3 — Figure 3: Visualization of decoupled model output relationships. (a,d) directional similarities of $v_{\theta}(x_{t},t)$ with $v_{\varphi}(x_{t},t)$ and $\hat{v}_{\theta}(x_{t},t)$ ; (b,e) magnitude ratios of $v_{\theta}(x_{t},t)$ to $v_{\varphi}(x_{t},t)$ and to $\hat{v}_{\theta}(x_{t},t)$ ; (c,f) magnitudes of $v_{\varphi}(x_{t},t\mid c)$ (Cond) and $v_{\varphi}(x_{t},t\mid c=\emptyset)$ (Uncond). Directional similarities are evaluated using cosine similarity, and magnitudes are measured using $\ell_{2}$ norms.

Figure 4: Comparison of our method with the adaptive caching method, TeaCache [ 22 ] , and the large-small model collaborative inference strategy, SRDiffusion [ 6 ] . Our method improves the approximation of the original large output by combining direction estimation from residual reuse with magnitude estimation from the small model, offering a more reliable lightweight alternative for accelerating inference.

Wan2.1 (832 $\times$ 480, 81 frames, $T=50$ )
Method	Efficiency		Visual Quality
Method	Speedup $\uparrow$	Latency $\downarrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	VBench $\uparrow$
Wan2.1-14B	1 $\times$	948 s	-	-	-	83.09%
Wan2.1-1.3B	-	192 s	0.597	11.81	0.336	81.00%
TeaCache [ 22 ]	2.62 $\times$	362 s	0.208	20.95	0.680	82.41%
SRDiffusion [ 6 ]	2.75 $\times$	344 s	0.240	19.10	0.636	83.01%

Table 1: Quantitative evaluation of inference efficiency and visual quality in video generation models. Visual retention metrics, including LPIPS, SSIM, and PSNR, are calculated against the original large models (Wan2.1-14B and EasyAnimateV5.1-12B).

Threshold $\tau$	Efficiency		Visual Quality
Threshold $\tau$	Speedup $\uparrow$	Latency $\downarrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	VBench $\uparrow$
Wan2.1-14B	1 $\times$	948 s	-	-	-	83.09%
$\tau=0.001$	2.21 $\times$	428 s	0.131	23.76	0.786	82.74%
$\tau=0.002$	2.76 $\times$	344 s	0.148	23.33	0.771	82.85%
$\tau=0.005$	2.95 $\times$	321 s	0.178	22.72	0.748	82.62%
$\tau=0.010$	3.12 $\times$	303 s	0.207	22.21	0.728	82.20%

Table 2: Ablation study of threshold

\tau

in Eq. ( 6 ) on Wan2.1-14B. A larger threshold enables a greater proportion of steps to be approximated by the lightweight small-model, yielding higher acceleration at the expense of visual quality.

Methods	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	VBench $\uparrow$
MDD w/o rescaling	0.197	21.40	0.689	81.16%
MDD	0.178	22.72	0.748	82.62%

Table 3: Ablation study of rescaling with the lightweight model on Wan2.1-14B.

CFG Reuse	Efficiency		Visual Quality
CFG Reuse	Speedup $\uparrow$	Latency $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	VBench $\uparrow$
$\times$	-	385 s	0.177	22.73	0.750	82.64%
$\checkmark$	1.20 $\times$	321 s	0.178	22.72	0.748	82.62%

Table 4: Ablation study of CFG Reuse on Wan2.1. Visual retention metrics (LPIPS, SSIM, and PSNR) are computed relative to the Wan2.1-14B model, with

\tau

set to 0.005.

Interval	Efficiency		Visual Quality
Interval	Speedup $\uparrow$	Latency $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
Wan2.1-14B	1 $\times$	476 s	-	-	-
TeaCache [ 22 ]	2.37 $\times$	201 s	0.367	17.54	0.607
SRDiffusion [ 6 ]	2.47 $\times$	193 s	0.381	16.75	0.585
Ours ( $\tau$ = 0.002)	2.50 $\times$	190 s	0.276	18.93	0.661
Ours ( $\tau$ = 0.005)	2.77 $\times$	172 s	0.292	18.48	0.655

Table 5: Results for Wan2.1-14B 720p, with latency measured under an 8

\times

NVIDIA A100 GPU setup.

Methods	ODE solver	Efficiency		Visual Quality
Methods	ODE solver	Speedup $\uparrow$	Latency $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	VBench $\uparrow$
Wan2.1-14B (T=50)	UniPC	1 $\times$	948 s	-	-	-	83.09%
Ours (T=50)	UniPC	2.95 $\times$	321 s	0.178	22.72	0.748	82.62%
Wan2.1-14B (T=40)	DPM++	1 $\times$	766 s	-	-	-	82.86%
Ours (T=40)	DPM++	2.68 $\times$	286 s	0.169	23.00	0.765	82.45%

Table 6: Empirical analysis of lightweight substitution intervals.

Interval	Efficiency		Visual Quality
Interval	Speedup $\uparrow$	Latency $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	VBench $\uparrow$
0%–95%	3.85 $\times$	246 s	0.567	13.43	0.431	79.02%
20%–95% (Ours)	2.95 $\times$	321 s	0.178	22.72	0.748	82.62%

Table 7: Compatibility analysis results between MDD and efficient ODE solvers.

Video Generation

Video generation studies models that create and edit moving images from text, images, audio, and structured controls.

See all 364 papers

VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation
May 22, 2026Junwen Tan, Jinglin Liang, Hongyuan Chen +1Rectified Flow Image Generation
South China University of Technology
Parallel Decoding Distillation for Fast Image and Video Generation
Jul 28, 2026Neta Shaul, Chao Liu, Arash Vahdat +1Video Diffusion Models Diffusion Model Distillation
2Weizmann Institute of Science · 1NVIDIA
RhymeFlow: Training-Free Acceleration for Video Generation with Asynchronous Denoising Flow Scheduling
Jun 4, 2026Chensheng Dai, Shengjun Zhang, Yifan Li +3Video Diffusion Models Diffusion Transformer
1Tsinghua University · 2GigaAI

May 22, 2026J/K move · Enter open · S save

VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation

Junwen Tan, Jinglin Liang, Hongyuan Chen, Shuangping Huang

South China University of Technology

Rectified Flow Image Generation Video Generation Diffusion Model Inference Acceleration

Open paper AskPDF arXiv

Though rectified flow models have achieved remarkable performance in image, video, and 3D generation, their practical deployments are challenged by slow inference speeds. Prior acceleration methods reuse cached features from previous steps, which neglects the growing mismatch between static caches and the evolving input, leading to reduced output fidelity. This work proposes Velocity Decomposition and Estimation (VDE), a training-free acceleration method that shifts the paradigm from caching-and-reusing to decomposing-and-estimating. Specifically, VDE decomposes the model's velocity into components parallel and orthogonal to the input, exploiting their temporal predictability and directional stability for precise, input-adaptive estimation. To prevent error accumulation, it periodically anchors the model's state via full forward passes. Extensive experiments on image and video generation tasks demonstrate that VDE achieves substantial acceleration with minimal loss in visual quality. Notably, VDE accelerates Flux by 3.22 times and achieves an LPIPS of 0.069 on Qwen-Image, outperforming the best baseline with a 52.2% reduction.

Abstract

Video Generation

Explore similar work

VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation

Parallel Decoding Distillation for Fast Image and Video Generation

RhymeFlow: Training-Free Acceleration for Video Generation with Asynchronous Denoising Flow Scheduling