Flash-VAED: Plug-and-Play VAE Decoders for Efficient Video Generation

Lunjie Zhu; Yushi Huang; Xingtong Ge; Yufei Xue; Zhening Liu; Yumeng Zhang; Zehong Lin; Jun Zhang

cs.CVFeb 22, 2026

Flash-VAED: Plug-and-Play VAE Decoders for Efficient Video Generation

Authors: Lunjie Zhu, Yushi Huang, Xingtong Ge, Yufei Xue, Zhening Liu, Yumeng Zhang, Zehong Lin, Jun Zhang

Organizations: iComAI Lab, Hong Kong University of Science and Technology, Hong Kong SAR, China · School of Data Science, Lingnan University, Hong Kong SAR, China

Abstract

Latent diffusion models have enabled high-quality video synthesis, yet their inference remains costly and time-consuming. As diffusion transformers become increasingly efficient, the latency bottleneck inevitably shifts to VAE decoders. To reduce their latency while maintaining quality, we propose a universal acceleration framework for VAE decoders that preserves full alignment with the original latent distribution. Specifically, we propose (1) an independence-aware channel pruning method to effectively mitigate severe channel redundancy, and (2) a stage-wise dominant operator optimization strategy to address the high inference cost of the widely used causal 3D convolutions in VAE decoders. Based on these innovations, we construct a Flash-VAED family. Moreover, we design a three-phase dynamic distillation framework that efficiently transfers the capabilities of the original VAE decoder to Flash-VAED. Extensive experiments on Wan and LTX-Video VAE decoders demonstrate that our method outperforms baselines in both quality and speed, achieving approximately a 6 $\times$ speedup while maintaining the reconstruction performance up to 96.9%. Notably, Flash-VAED accelerates the end-to-end generation pipeline by up to 36% with negligible quality drops on VBench-2.0. Our code is available at https://github.com/Aoko955/Flash-VAED.

Figures & tables

Figure 2: Overview of the Flash-VAED architecture 2 2 2 We take Flash-VAED-Wan 2.1 as an example. . The proposed stage-wise dominant operator optimization substitutes CausalConv3D with stage-specific efficient operators (left), tailored to each decoding stage. Moreover, the independence-aware channel pruning method (right) reduces the channel count to 12.5% $-$ 25% of the original with minimal quality loss, leveraging channel independence.

Figure 3: Channel-wise similarity analysis. We visualize the top-8 channels most similar to Channel 0. Although the feature maps exhibit visual similarity, the quantitative similarity scores are not high enough to support pruning.

Figure 4: SVD analysis on channel features. The curve for cumulative explained variance ratio reveals the intrinsic low-rank nature of the feature maps. Notably, only 22 components (22.9% of the total) are required to explain 99% of the feature variance, providing strong empirical support for our pruning method.

Model	$(\mathrm{d_{T}},\mathrm{d_{H}},\mathrm{d_{W}})$	FPS $\uparrow$		Reconstruction Quality
Model	$(\mathrm{d_{T}},\mathrm{d_{H}},\mathrm{d_{W}})$	RTX 5090D	Jetson Orin	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Wan 2.1 ( Wan et al., 2025 )	$(4,8,8)$	19.27	0.65	40.40	0.9733	0.0190
LightVAE-Wan 2.1 ( Contributors, 2025 )	$(4,8,8)$	118.60	3.70	32.61	0.9416	0.0892
Flash-VAED-Wan 2.1 (Ours)	$(4,8,8)$	118.77	3.70	37.61	0.9614	0.0285
LTX-Video ( HaCohen et al., 2024 )	$(8,32,32)$	204.55	4.75	33.28	0.9253	0.0497
Turbo-VAED-LTX ( Zou et al., 2026 )	$(8,32,32)$	623.08	23.24	31.52	0.9275	0.0555

Table 1: Quantitative comparison of video reconstruction results with SOTA methods. We compare Flash-VAED against the original VAE decoders and competitive baselines on both a consumer-grade GPU (RTX 5090D) and an edge device (Jetson Orin), where

\mathrm{d_{T}},\mathrm{d_{H}}

, and

\mathrm{d_{W}}

denote the compression ratios for time, height, and width, respectively,

\uparrow

indicates higher is better, while

\downarrow

indicates lower is better. The best results among the methods are highlighted in bold . Flash-VAED outperforms all the baselines in both speed and quality.

Model	$E_{\mathrm{warp}}$ $\downarrow$
Wan 2.1 ( Wan et al., 2025 )	0.017783
LightVAE-Wan 2.1 ( Contributors, 2025 )	0.026640
Flash-VAED-Wan 2.1	0.018662
LTX-Video ( HaCohen et al., 2024 )	0.022071
Turbo-VAED-LTX ( Zou et al., 2026 )	0.026261
Flash-VAED-LTX	0.024267

Table 2: Quantitative comparison of temporal consistency with SOTA methods. We report warping error to measure temporal coherence, where

\downarrow

indicates lower is better. The best results among the methods are highlighted in bold.

Figure 8: Visual comparison of video generation results. Our Flash-VAED (bottom) matches closely with the original Wan 2.1 VAE (top) in fidelity, details and textures, whereas the LightVAE (middle) exhibits severe artifacts that results in invalid video content. The full prompts and additional details are provided in Appendix A.4 .

Panel of Figure 9 — Figure 9: Generation performance of Flash-VAED embedded on the VBench-2.0. We employ Flash-VAED to replace the VAE decoders of: (a) the Self Forcing-Wan 1.3B, and (b) the FastVideo-Wan 1.3B, whose overall generation quality is evaluated on VBench-2.0.

Figure 10: Visualization of samples from the latent distribution.

Figure 11: Ablation results of using different pruning methods.

Prune Ratio	FPS/5090D $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
1	204.55	33.2883	0.9253	0.0497
1/2	625.00	32.5421	0.9307	0.0521
1/4	1167.99	32.2400	0.9293	0.0551
1/8	1907.23	31.0786	0.9162	0.0804

Table 3: Ablation results of using different pruning ratios.

Project Layer	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Random Initialization	30.3061	0.9013	0.0978
Ours	32.2400	0.9293	0.0551

Table 4: Ablation results of different projected layer initialization.

Training Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Without Distillation	30.7957	0.9125	0.0831
Deep-only Distillation	30.8065	0.9126	0.0830
Shallow-only Distillation	31.2386	0.9152	0.0722
Ours	32.3768	0.9317	0.0566

Table 5: Ablation results of using different distillation methods.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 12: Effectiveness of shortcut injection.

Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Wan 2.1 ( Wan et al., 2025 )	34.13	0.9079	0.0487
LightVAE-Wan 2.1 ( Contributors, 2025 )	27.92	0.8397	0.2069
Flash-VAED-Wan 2.1	32.03	0.8825	0.0769
LTX-Video ( HaCohen et al., 2024 )	29.36	0.8134	0.1555
Turbo-VAED-LTX ( Zou et al., 2026 )	26.82	0.7805	0.1783
Flash-VAED-LTX	28.26	0.8031	0.1214

Appendix

Table 6: Reconstruction quality comparison on the DAVIS 2019 Challenge test set at 720p resolution, where

\uparrow

indicates higher is better and

\downarrow

indicates lower is better. The best results are highlighted in bold.

Panel of Figure 13 — Figure 13: Qualitative comparison of video reconstruction results on the DAVIS 2019 Challenge test set for Wan 2.1-based VAE decoders. For each sample, the three rows from top to bottom correspond to the original Wan 2.1 decoder, LightVAE-Wan 2.1, and Flash-VAED-Wan 2.1, respectively. In high-resolution videos with fine-grained details and highly dynamic scenes, Flash-VAED preserves visual details and temporal structures better than LightVAE.

Panel of Figure 14 — Figure 14: Qualitative comparison of video reconstruction results on the DAVIS 2019 Challenge test set for LTX-Video-based VAE decoders. For each sample, the three rows from top to bottom correspond to the original LTX-Video decoder, Turbo-VAED-LTX, and Flash-VAED-LTX, respectively. In high-resolution videos with fine-grained details and highly dynamic scenes, Flash-VAED achieves better reconstruction quality than Turbo-VAED.

Model	Number of Parameters $\downarrow$	Peak VRAM $\downarrow$
Wan 2.1 ( Wan et al., 2025 )	73.30 M	7.11 GB
Wan 2.1 + Channel Pruning	65.20 M	2.31 GB
Wan 2.1 + Operator Optimization	5.63 M	3.51 GB
LightVAE-Wan 2.1 ( Contributors, 2025 )	4.62 M	1.17 GB
Flash-VAED-Wan 2.1	5.86 M	2.00 GB

Appendix

Table 7: Additional efficiency analysis on the Wan 2.1 VAE decoder. We report the parameter count and peak VRAM usage of different lightweight variants and methods.

Model	DSR $\uparrow$	HC $\uparrow$	HId $\uparrow$	HIn $\uparrow$	IR $\uparrow$	MOU $\uparrow$	MVC $\uparrow$
Wan 2.1 ( Wan et al., 2025 )	0.217	0.986	0.785	0.600	0.924	0.131	0.424
LightVAE-Wan 2.1 ( Contributors, 2025 )	0.198	0.953	0.527	0.570	0.819	0.121	0.072
Flash-VAED-Wan 2.1	0.237	0.986	0.741	0.590	0.918	0.141	0.318

Appendix

Table 8: Generation performance of Flash-VAED-integrated Causal Forcing on VBench 2.0. We replace the original Wan 2.1 VAE decoder with different accelerated decoders and report selected VBench 2.0 dimensions, including Dynamic Spatial Relationship (DSR), Human Clothes (HC), Human Identity (HId), Human Interaction (HIn), Instance Preservation (IP), Motion Order Understanding (MOU), and Multi-View Consistency (MVC).

Figure 15: Ablation results of using different resolutions and numbers of frames.

Model	Text Encoding	Denoising	Video Decoding	Decoding Percentage
Wan 1.3B ( Wan et al., 2025 )	0.619	178.066	4.212	2.3%
Self Forcing-Wan 1.3B ( Huang et al., 2026 )	0.598	6.254	3.172	31.6%
FastVideo-Wan 1.3B ( Zhang et al., 2025 )	0.539	5.136	3.925	40.9%

Appendix

Table 9: Latency breakdown of generation pipelines

Model	$(\mathrm{d_{T}},\mathrm{d_{H}},\mathrm{d_{W}})$	FPS $\uparrow$	Reconstruction Quality
Model	$(\mathrm{d_{T}},\mathrm{d_{H}},\mathrm{d_{W}})$	RTX 5090D	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Wan 2.1 ( Wan et al., 2025 )	$(4,8,8)$	19.27	40.40	0.9733	0.0190
HunyuanVideo ( Kong et al., 2024 )	$(4,8,8)$	6.77	33.05	0.9392	0.0272
CogVideoX ( Yang et al., 2025 )	$(4,8,8)$	4.99	32.81	0.9351	0.0302
LTX-Video ( HaCohen et al., 2024 )	$(8,32,32)$	204.55	33.28	0.9253	0.0497

Appendix

Table 10: To validate our choices of Wan ( Wan et al., 2025 ) and LTX-Video ( HaCohen et al., 2024 ) as original VAE decoders, we compare them against current SOTA video VAE decoders. Wan establishes a clear lead in both quality and speed among models with mainstream compression ratios, while LTX-Video dominates in terms of inference speed.

Channel Pruning

See all 10 papers

FlashDecoder: Real-Time Latent-to-Pixel Streaming Decoder with Transformers
Jul 16, 2026Minguk Kang, Suha KwakEfficient Transformer Inference Transformer
Pika Labs · POSTECH
LightVAE: Towards Compact and Efficient Video Autoencoders
Oct 10, 2026Shiqiao Gu, Yang Yong, Fengjun Zhong +5Variational Autoencoders Video Compression
SenseTime · Beihang University · Nanyang Technological University
UltraFlash: Accelerating Megapixel Visual Synthesis
Oct 24, 2026Phuc Lai, Anh Nguyen, Phong Nguyen +1Diffusion Model Inference Acceleration Diffusion-Based Image Generation
Qualcomm AI Research†

Jul 16, 2026J/K move · Enter open · S save

FlashDecoder: Real-Time Latent-to-Pixel Streaming Decoder with Transformers

Minguk Kang, Suha Kwak

Pika Labs · POSTECH

Efficient Transformer Inference Transformer Latent Diffusion Models Streaming Video Generation Streaming Inference

Open paper AskPDF arXiv

Real-time video generation demands fast decoding as much as fast denoising, yet current latent video diffusion models rely on 3D convolutional decoders that are slow and memory-intensive at high resolutions or for long video. We introduce FlashDecoder, a fast, memory-efficient pure-Transformer video decoder that decodes latents to pixels frame by frame. At each step, the current frame attends only to a fixed-size window of past frames through a rolling KV cache. The fixed temporal window keeps decoding fast and memory bounded regardless of video length, enabling constant-latency streaming. Because frames are processed sequentially, temporal causality is enforced without explicit attention masks, enabling training at resolutions up to 1080p and matching the reconstruction quality of convolutional decoders. On the Wan2.1 and Wan2.2 latent spaces, FlashDecoder matches each convolutional decoder in reconstruction quality (e.g., 41.55dB vs. 41.49dB PSNR at 1080p) while decoding 3.6x-4.7x faster with up to 11x less memory on a single H100 GPU. With architecture-aware inference optimizations, the speedup widens to 12x.

Abstract

Channel Pruning

Explore similar work

FlashDecoder: Real-Time Latent-to-Pixel Streaming Decoder with Transformers

LightVAE: Towards Compact and Efficient Video Autoencoders

UltraFlash: Accelerating Megapixel Visual Synthesis