Predict to Skip: Linear Multistep Feature Forecasting for Efficient Diffusion Transformers

Hanshuai Cui; Zhiqing Tang; Qianli Ma; Zhi Yao; Weijia Jia; Wei Zhao

cs.CVFeb 20, 2026

Predict to Skip: Linear Multistep Feature Forecasting for Efficient Diffusion Transformers

Authors: Hanshuai Cui, Zhiqing Tang, Qianli Ma, Zhi Yao, Weijia Jia, Wei Zhao

Organizations: School of Artificial Intelligence, Beijing Normal University, Beijing, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai, China · Shenzhen University of Advanced Technology, Shenzhen, China

Abstract

Diffusion Transformers (DiT) have emerged as a widely adopted backbone for high-fidelity image and video generation, yet their iterative denoising process incurs high computational costs. Existing training-free acceleration methods rely on feature caching and reuse under the assumption of temporal stability. However, reusing features for multiple steps may lead to latent drift and visual degradation. We observe that model outputs evolve smoothly along much of the diffusion trajectory, enabling principled predictions rather than naive reuse. Based on this insight, we propose PrediT, a training-free acceleration framework that formulates feature prediction as a linear multistep problem. We employ classical linear multistep methods to forecast future model outputs from historical information, combined with a corrector that activates in high-dynamics regions to prevent error accumulation. A dynamic step modulation mechanism adaptively adjusts the prediction horizon by monitoring the feature change rate. Together, these components enable substantial acceleration while preserving generation fidelity. Extensive experiments validate that our method achieves up to $5.54\times$ latency reduction across various DiT-based image and video generation models, while incurring negligible quality degradation.

Figures & tables

Figure 1. Accelerating Diffusion Transformer inference across multiple models. PrediT achieves significant speedup on various DiT-based architectures for both image and video generation while preserving visual quality. A teaser montage of image and video generation examples from several Diffusion Transformer models, illustrating that PrediT reduces inference cost while preserving visual quality.

Panel of Figure 2 — Figure 2. Analysis of direct feature reuse limitations. (a) Feature trajectory visualization via PCA shows local smoothness along the diffusion process. (b) Comparison of per-step approximation error and accumulated drift: naive reuse incurs significant error accumulation, while higher-order prediction methods reduce drift. A two-panel analysis figure. The left panel shows a PCA visualization of feature trajectories along the diffusion process. The right panel plots approximation error and accumulated drift, showing that higher-order prediction reduces drift compared with naive reuse.

Figure 3. Overview of the PrediT framework. PrediT formulates feature prediction as a linear multistep problem, using historical model outputs to extrapolate future features and reduce redundant model calls. The threshold $\tau$ and correction ratio $r$ control the selection between ABM for high-dynamics and AB for low-dynamics regions. A pipeline diagram of the PrediT framework showing historical model outputs feeding an Adams-Bashforth predictor, an Adams-Moulton corrector for high-dynamics regions, and dynamic step modulation to control skip intervals.

Method	Acceleration				Visual Quality
Method	Latency(s) $\downarrow$	Speedup $\uparrow$	FLOP(T) $\downarrow$	Speedup $\uparrow$	Img. $\uparrow$	CLIP $\uparrow$	Aes. $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
Original: 50 Steps	23.71	1.00 $\times$	3719.50	1.00 $\times$	0.9767	31.94	6.7448	0.0000	1.000	$\infty$
$\Delta$ -DiT( $\mathcal{N}=2$ )	15.86	1.50 $\times$	1859.75	2.00 $\times$	0.9357	32.02	6.5344	0.3074	0.6783	18.63
TeaCache( $\tau=0.4$ )	12.31	1.92 $\times$	1438.92	2.58 $\times$	0.8710	31.79	6.4400	0.6960	0.4175	11.54
ABM-Solver( $\tau=0.1$ )	13.05	1.82 $\times$	952.22	3.90 $\times$	0.9378	32.08	6.7260	0.4762	0.6014	17.72
Original: 25 Steps	11.58	2.05 $\times$	1859.75	2.00 $\times$	0.9467	31.64	6.7307	0.2970	0.6818	18.85

Table 1. Comparison of visual quality and efficiency in text-to-image generation on FLUX.

Figure 4. Visual quality comparison of different acceleration methods on FLUX. PrediT achieves higher acceleration while maintaining comparable visual quality. Zoom in for more details. A qualitative comparison grid of FLUX text-to-image results across several prompts, contrasting the original model, baseline acceleration methods, and PrediT.

Method	Acceleration		Image	CLIP
FLUX.1	Latency(s) $\downarrow$	Speedup $\uparrow$	Reward $\uparrow$	Score $\uparrow$
Original: 50 Steps	23.71	1.00 $\times$	0.9767	31.9448
Only AB w/o DSM	6.49	3.65 $\times$	0.9308	31.8125
Only ABM w/o DSM	11.89	1.99 $\times$	0.9318	31.8634
Only AB + DSM	3.65	6.49 $\times$	0.9374	31.8534
Only ABM + DSM	6.15	3.86 $\times$	0.9671	32.1778

Table 2. Ablation study of predictor-corrector components and DSM, with comparison to static skipping baselines on FLUX.1.

Settings	Method	Acceleration		Visual Quality
Settings	Method	Latency(s) $\downarrow$	Speedup $\uparrow$	VBench $\uparrow$	Quality $\uparrow$	Semantic $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
544p $\times$ 960p, 17 frames	Original: 50 Steps	76.71	1.00 $\times$	82.08	82.86	78.98	0.0000	1.000	$\infty$
	PAB( $\mathcal{N}=8$ )	66.21	1.16 $\times$	80.81	81.36	78.60	0.1433	0.8370	27.41
	TeaCache( $\tau=0.14$ )	57.71	1.33 $\times$	82.18	82.36	81.44	0.1466	0.8536	25.64
	Original: 25 Steps	40.29	1.90 $\times$	81.51	82.10	78.68	0.1973	0.8196	23.17
	ProfilingDiT( $\mathcal{N}=6$ )	44.60	1.72 $\times$	80.33	80.02	81.58	0.2610	0.6662	18.77

Table 3. Comparison of visual quality and efficiency in text-to-video generation on HunyuanVideo on a single A800 GPU.

Panel of Figure 5 — Figure 5. Latency comparison between the original method and PrediT on HunyuanVideo at 544P and 720P resolutions across different frame counts. Two line plots comparing latency versus frame count for the original HunyuanVideo pipeline and PrediT at 544P and 720P resolutions.

Figure 6. Visual quality comparison of different acceleration methods on HunyuanVideo. Other acceleration methods suffer from blurriness or loss of details, while PrediT maintains comparable visual quality with higher acceleration. Zoom in for more details. A qualitative comparison grid of video frames from HunyuanVideo, showing the original method, baseline accelerators, and PrediT across several prompts.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Order	$\beta_{0}$	$\beta_{1}$	$\beta_{2}$	$\beta_{3}$	LTE
AB1	$1$	–	–	–	$\frac{1}{2}h^{2}f^{\prime\prime}$
AB2	$\frac{3}{2}$	$-\frac{1}{2}$	–	–	$\frac{5}{12}h^{3}f^{\prime\prime\prime}$
AB3	$\frac{23}{12}$	$-\frac{16}{12}$	$\frac{5}{12}$	–	$\frac{9}{24}h^{4}f^{(4)}$
AB4	$\frac{55}{24}$	$-\frac{59}{24}$	$\frac{37}{24}$	$-\frac{9}{24}$	$\frac{251}{720}h^{5}f^{(5)}$

Appendix

Table 4. Adams-Bashforth coefficients and local truncation errors.

Order	$\gamma_{-1}$	$\gamma_{0}$	$\gamma_{1}$	$\gamma_{2}$	$\gamma_{3}$	LTE
AM1	$1$	–	–	–	–	$-\frac{1}{2}h^{2}f^{\prime\prime}$
AM2	$\frac{5}{12}$	$\frac{8}{12}$	$-\frac{1}{12}$	–	–	$-\frac{1}{24}h^{3}f^{\prime\prime\prime}$
AM3	$\frac{9}{24}$	$\frac{19}{24}$	$-\frac{5}{24}$	$\frac{1}{24}$	–	$-\frac{19}{720}h^{4}f^{(4)}$
AM4	$\frac{251}{720}$	$\frac{646}{720}$	$-\frac{264}{720}$	$\frac{106}{720}$	$-\frac{19}{720}$	$-\frac{3}{160}h^{5}f^{(5)}$

Appendix

Table 5. Adams-Moulton coefficients and local truncation errors.

Method	480p $\times$ 640p $\times$ 45f	480p $\times$ 640p $\times$ 65f	544p $\times$ 960p $\times$ 17f	544p $\times$ 960p $\times$ 65f	720p $\times$ 1280p $\times$ 129f
Original	44295	45949	43131	49477	74465
PrediT	44819	46313	44043	49715	74525
PAB	79831	OOM	55045	OOM	OOM
TeaCache	48619	51499	48227	53319	80083
ProfilingDiT	OOM	OOM	57237	OOM	OOM
TaylorSeer	OOM	OOM	61557	OOM	OOM

Appendix

Table 6. Memory Usage (MB) on Different Resolutions and Frames using A800 80GB.

Settings	Latency(s) $\downarrow$	Speedup $\uparrow$	FID $\downarrow$	sFID $\downarrow$	IS $\uparrow$
DDIM-50	0.569	1.00 $\times$	2.28	4.25	242.3
( $\mathcal{O}=2$ , $\tau=1.25$ , $r=0.3$ )	0.219	2.60 $\times$	2.97	5.33	222.6
( $\mathcal{O}=2$ , $\tau=1.25$ , $r=0.1$ )	0.250	2.28 $\times$	3.01	5.24	222.5
( $\mathcal{O}=2$ , $\tau=1.00$ , $r=0.3$ )	0.251	2.27 $\times$	2.71	5.31	228.4
( $\mathcal{O}=2$ , $\tau=1.00$ , $r=0.2$ )	0.297	1.92 $\times$	2.72	5.28	228.3
( $\mathcal{O}=2$ , $\tau=1.00$ , $r=0.1$ )	0.293	1.94 $\times$	2.78	5.30	227.7

Appendix

Table 7. Ablation Study with Different Configurations on ImageNet with DiT-XL/2.

Figure 7. More Visual Comparisons on FLUX.1. Zoom in for best viewing. A supplementary comparison grid of FLUX text-to-image results across additional prompts, showing the original model, baseline accelerators, and PrediT.

Figure 8. More Visual Comparisons on HunyuanVideo. Zoom in for best viewing. A supplementary comparison grid of HunyuanVideo frames across additional prompts, contrasting baseline acceleration methods with PrediT.

Explore similar work

LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration
Feb 24, 2026Peiliang Cai, Jiacheng Liu, Haowen Xu +3Diffusion Model Inference Acceleration Diffusion Models
Shanghai Jiao Tong University
Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models
Apr 29, 2026Zhirong Shen, Rui Huang, Jiacheng Liu +6Diffusion Transformer Diffusion Model Caching
Shanghai Jiao Tong University · University of Electronic Science and Technology of China · Shandong University +2
Memory-Efficient Training-Free Acceleration of Diffusion Transformers with BaryCache
Aug 24, 2026Chengjie Lu, Tianchi Deng, Zhengqi He +3Diffusion Model Inference Acceleration Diffusion-Based Image Generation
College of Electronics and Information Engineering, Shenzhen University · School of Artificial Intelligence, Shenzhen University · College of Computer Science and Software Engineering, Shenzhen University

Feb 24, 2026J/K move · Enter open · S save

LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration

Peiliang Cai, Jiacheng Liu, Haowen Xu, Xinyu Wang+2

Shanghai Jiao Tong University

Diffusion Model Inference Acceleration Diffusion Models Diffusion Model Caching T2I Diffusion Models Diffusion Transformer Video Diffusion Models

Open paper AskPDF arXiv

Diffusion models have achieved remarkable success in image and video generation tasks. However, the high computational demands of Diffusion Transformers (DiTs) pose a significant challenge to their practical deployment. While feature caching is a promising acceleration strategy, existing methods based on simple reusing or training-free forecasting struggle to adapt to the complex, stage-dependent dynamics of the diffusion process, often resulting in quality degradation and failing to maintain consistency with the standard denoising process. To address this, we propose a LEarnable Stage-Aware (LESA) predictor framework based on two-stage training. Our approach leverages a Kolmogorov-Arnold Network (KAN) to accurately learn temporal feature mappings from data. We further introduce a multi-stage, multi-expert architecture that assigns specialized predictors to different noise-level stages, enabling more precise and robust feature forecasting. Extensive experiments show our method achieves significant acceleration while maintaining high-fidelity generation. Experiments demonstrate 5.00x acceleration on FLUX.1-dev with minimal quality degradation (1.0% drop), 6.25x speedup on Qwen-Image with a 20.2% quality improvement over the previous SOTA (TaylorSeer), and 5.00x acceleration on HunyuanVideo with a 24.7% PSNR improvement over TaylorSeer. State-of-the-art performance on both text-to-image and text-to-video synthesis validates the effectiveness and generalization capability of our training-based framework across different models. Our code is available at https://github.com/caipeiliang2004/LESA.

Abstract

Explore similar work

LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration

Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models

Memory-Efficient Training-Free Acceleration of Diffusion Transformers with BaryCache