acceptodds
Under review as a conference paper at ICLR 2027

LightVAE: Towards Compact and Efficient Video Autoencoders

12 views

Abstract

Latent diffusion models enable high-quality video generation, but video VAE encoding and decoding remain computationally expensive. Existing methods use stage-specific lightweight operators, yet either constrain spatial and temporal computation jointly or remove temporal computation altogether, without distinguishing their compression sensitivities in pretrained Conv3D. To address these limitations, we present LightVAE, a compression framework that exploits redundancy in temporal modeling and network depth. LightVAE exactly reparameterizes eligible causal Conv3D layers into an aggregate spatial response and a difference-driven temporal correction. Based on their distinct compression sensitivities, temporal low-rank compression (TLR) keeps the aggregate spatial kernel dense and applies layer-specific low-rank approximation only to the temporal correction. An algebraically equivalent folded implementation converts the arithmetic savings into practical acceleration. LightVAE further uses width-preserving block pruning to reduce depth and probe-guided feature alignment, guided by feedback from a frozen teacher suffix, to recover quality after compression. The framework also extends to video encoders. Across three video VAE backbones and two GPU platforms, LightVAE consistently accelerates decoding. For the Wan VAE family, it achieves up to and the decoding throughput of state-of-the-art efficient VAE decoders on NVIDIA H100 and RTX 5090, respectively, with higher DAVIS PSNR and lower LPIPS. On MiniMax-H3, it achieves up to the original decoding throughput while retaining 99.51% of the original model's overall VBench-2.0 score.

Primary area:
generative models
Keywords:
Video Variational Autoencoders, Efficient Video Generation, Model Compression

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KYmLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LightVAE, an approach to making existing video VAEs more computationally efficient without training a new architecture from scratch. The main idea is to reduce unnecessary computation in pretrained models while limiting the degradation in video reconstruction quality.

The proposed method consists of four key components: 1) Temporal Low-Rank Compression (TLR) preserves spatial processing while reducing temporal computation through low-rank approximation. 2) Weight-Folded Execution reuses cached features and combines operations to accelerate inference. 3) Width-Preserving Block Pruning removes redundant blocks without reducing channel dimensions. 4) Probe-Guided Feature Alignment uses teacher feedback to improve student reconstruction quality.

Critical strengths
  • The proposed framework can directly compress pretrained video VAEs while preserving their latent interfaces, making it practical to integrate into existing video generation pipelines without retraining the entire model.
  • The paper provides a solid analysis of the different compression sensitivities of spatial and temporal components, leading to a principled compression strategy. Furthermore, the authors connect the mathematical formulation with practical engineering optimizations, such as weight folding and caching, to achieve actual inference speedups rather than merely reducing theoretical computational costs.
Critical weaknesses
  • Since temporal reconstruction quality is also important for video VAEs, additional evaluation of temporal consistency would strengthen the experimental validation. While the paper reports frame-level reconstruction metrics and VBench-2.0 scores, including complementary video-level metrics, such as FVD, could help assess whether temporal information is preserved after compression.
  • The paper motivates TLR by temporal continuity and shows that temporal-only compression yields lower approximation errors than spatial or joint compression. However, it remains unclear whether this observation consistently holds across individual videos with different levels of temporal variation. In particular, it would be valuable to examine whether the proposed compression strategy remains effective under large inter-frame changes, and whether the stage-wise temporal sensitivity observed in Figure 5 still holds in such cases.
Questions for authors
  • The paper presents an interesting motivation and overall approach. However, I would like to see additional evaluation of temporal consistency to better assess video reconstruction quality. I am also curious whether the underlying assumption of temporal redundancy holds consistently for both highly similar consecutive frames and frames with substantial temporal changes.
5 comments
AuthorsOpenReview ↗

Response to Reviewer KYmL

Response to Weaknesses 1–2 and Question 1

Thank you for suggesting additional video-level evaluation and analysis of inter-frame variation. We followed your suggestions by adding reconstruction FVD evaluations on UCF101 and DAVIS and per-video comparisons of compression strategies under low and high inter-frame variation. The experiments show that LightVAE-Pro achieves the lowest reconstruction FVD among the evaluated compressed decoders within each backbone, while temporal-only compression retains its error advantage on every tested video, including those with large inter-frame changes.

1. Video-level reconstruction quality

We added reconstruction FVD (rFVD) evaluations on UCF101 and DAVIS, using features from an I3D model pretrained on Kinetics-400 [1] to compare the distributions of source and reconstructed videos.

BackboneDecoderUCF101 rFVD ↓DAVIS rFVD ↓
Wan2.1Original VAE2.3614.29
Wan2.1Flash-VAED8.9927.04
Wan2.1LightVAE-Pro2.8615.63
Wan2.1LightVAE-Lite8.2023.44
Wan2.2Original VAE2.6612.08
Wan2.2Turbo-VAED6.1337.55
Wan2.2LightVAE-Pro4.4918.06
Wan2.2LightVAE-Lite9.3225.96

Combined with Table 1, Wan2.2 Pro and Wan2.1 Lite offer higher H100/5090 throughput and lower rFVD on both datasets than Turbo-VAED and Flash-VAED, respectively. These results complement the frame-level reconstruction metrics with a video-level distributional assessment.

2. Robustness to large inter-frame changes

We quantify inter-frame variation as the mean absolute RGB difference between consecutive frames, with pixel values normalized to [0, 1]. Higher scores indicate larger changes in appearance. Before examining compression errors, we selected 128 low-variation and 128 high-variation videos from the same Wan2.1-generated corpus using this measure. Their median scores were 0.0004 and 0.1153, respectively. The table reports mean relative L₂ error against the original decoder’s output for the same latent input.

Across both groups, temporal-only compression produces lower error than spatial-only and joint compression for every tested video at each of the five decoder stages. Here, “up” is short for “upsampling” and labels stages along the decoder’s upsampling path. The stage-wise trend in Figure 5 also holds: in both groups, mean temporal-only error decreases from up1 to up3, toward later stages at higher spatial resolutions.

Inter-frame variationCompressed branchmidup0up1up2up3
LowSpatial only0.01220.02200.02950.01950.0280
LowTemporal only0.00180.00350.00960.00180.0018
LowBoth0.01200.02220.03000.01940.0278
HighSpatial only0.04740.07440.10790.06410.0460
HighTemporal only0.02070.02330.06100.02480.0116
HighBoth0.04660.07340.11570.06580.0468

We have included these results, evaluation settings, and per-video comparisons in the revised manuscript.

References

[1] João Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6299–6308, 2017.

Thank you for the additional video-level results. I found that LightVAE-Lite has a higher rFVD than Turbo-VAED on UCF101 (9.32 vs. 6.13), but a lower rFVD on DAVIS (25.96 vs. 37.55) using Wan2.2. Could the authors provide some explanation for this difference?

AuthorsOpenReview ↗

Response to Reviewer KYmL

Thank you for pointing out this difference. This is because LightVAE-Lite adopts a more aggressive compression strategy to target real-time decoding. It reduces the number of residual blocks in the Wan2.2 decoder from 14 to 5 and achieves 2.24× the decoding throughput of Turbo-VAED (Table 1). The trade-off is that LightVAE-Lite’s reconstruction performance is less consistent across datasets than that of LightVAE-Pro. Its rFVD is higher than Turbo-VAED’s on UCF101 and lower on DAVIS.

LightVAE-Pro offers a more direct comparison with Turbo-VAED because the two decoders are closer in depth. LightVAE-Pro retains all 14 residual branches of the original Wan2.2 decoder. It achieves higher decoding throughput than Turbo-VAED (38.84 vs. 29.24 FPS, Table 1) and lower rFVD on both UCF101 (4.49 vs. 6.13) and DAVIS (18.06 vs. 37.55).

These results show that LightVAE can accelerate decoding at different parameter budgets, with different trade-offs between throughput and reconstruction quality to suit deployment needs.

Reviewer_eTbrOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LightVAE is a post-training compression pipeline for 3D causal video autoencoders. It decomposes causal Conv3D layers into a dense spatial kernel and a low-rank SVD temporal correction, folded into one convolution at inference. It also combines Gumble-softmax block pruning with teacher-suffix probe feature distillation for quality recovery. Experiments reports up to 4.2x decoding speedup on Wan VAEs and 2.97x on MiniMax-H3.

Critical strengths
  • The paper target a practical deployment deployment bottleneck in latent video generation, where high-resolution caucal Conv3D decoding dominates latency after diffusion step-distll.
  • The Section 4.2 introduces folds current-frame temporal correction into the spatial kernel without extra rank truncation.
  • The paper is well-organized with clear notation. The component ablations in Table3 isolate depth pruning, low-rank conversion, and zero-rank stage replacement.
Critical weaknesses
  • The proposed pipeline mainly adopts existing techniques in model compression and distillation.
    • Low-Rank Filter Decomposition and Folding follow standard SVD factorization [1] and linear re-parameterization [2].
    • Block Pruning and Feature Distillation is commonly adopted in prior works. The depth pruning is similar to TinyFusion while feature recovery combines a CrossKD frozen teacher with gradient-guided weights.
  • The experiment has unfair comparison. Table 1 benchmarks Wan2.2 VAE on unchunked clips, however Wan2.2 VAE is a causal-VAE that support streaming inference.
  • The downstream generation evaluation in Table 2 does not reliably measure decoder reconstruction fidelity. Because the DiT backbone and sampled latents are frozon, a lossy decoder cannot improve high-level video semantics. The table is misleading by connecting the decoder performance with video generation quality.

  1. Denton, E. L., Zaremba, W., Bruna, J., LeCun, Y., & Fergus, R. (2014). Exploiting linear structure within convolutional networks for efficient evaluation. In Advances in Neural Information Processing Systems (NeurIPS) (Vol. 27, pp. 1269–1277).

  2. Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., & Sun, J. (2021). RepVGG: Making VGG-style ConvNets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 13733–13742).

Questions for authors
  • Since causal VAEs (Wan2.1 and Wan2.2) decode chunk-by-chunk with feature caching while Turbo-VAED is non-causal, how do these model compare across both backbones under a same evaluation setting?
  • The DiT latents are frozen, why should VBench-2.0 increases in Table 2 be interpreted as real quality gains rather than metric noise, especially given the PSNR is lower than original decoder?
7 comments
AuthorsOpenReview ↗

Response to Reviewer eTbr (part 1/4)

Response to Weakness 1

Thank you for raising this concern. Below, we explain how our method differs from these prior works.

(1) Low-rank compression: TLR vs. Denton et al. [1]

Both methods use low-rank approximation. The key differences are the architectures they target and how they identify the weights to compress. Directly extending Denton et al.’s low-rank approximation to Conv3D in video VAEs may change the sum of the temporal kernel slices, S=∑jWjS=\sum_j W_j (Eq. 4), and introduce larger output errors. Our ablations show that the spatial branch is more sensitive to compression (Figs. 5 and 6).

Denton et al. target 2D CNNs and construct the weights to compress through channel clustering. Their biclustering scheme groups input and output channels by weight similarity, then applies low-rank approximation to each resulting weight subtensor ([1], Section 3.4).

TLR targets 3D convolutions in video VAEs and identifies the weights to compress through exact spatiotemporal decomposition. We decompose causal Conv3D into a spatial response and an inter-frame correction, where the spatial kernel is S=∑jWjS=\sum_j W_j (Eq. 4). At conversion, we preserve SS exactly and apply low-rank approximation only to the inter-frame correction weights, so the approximation error acts only on inter-frame feature differences (Eqs. 5–6), protecting the spatial branch while exploiting inter-frame redundancy.

(2) Execution: TLR vs. RepVGG [2]

TLR differs from RepVGG in the scope of reparameterization, kernel organization, and intermediate-feature reuse. This design is tailored to causal video VAE decoding: it reduces spatial-convolution calls and reuses cached low-rank projections across frames, avoiding explicit inter-frame differences during inference.

Reparameterization scope. RepVGG merges parallel branches operating on the same input into a single dense convolution. TLR takes current-frame features XtX_t and historical features Xt−1,Xt−2X_{t-1},X_{t-2}. After expanding the inter-frame differences, we merge the weights of all terms acting on XtX_t in the current output into a single spatial kernel, while retaining low-rank computation for the historical terms.

Kernel organization. RepVGG adds the equivalent convolution kernels elementwise, keeping CoutC_{\mathrm{out}} output channels. TLR concatenates the merged spatial kernel and two low-rank projection kernels along the output-channel dimension. One packed Conv2D performs all three computations, producing Cout+2rC_{\mathrm{out}}+2r intermediate channels while keeping the three output groups separate.

Intermediate-feature reuse. RepVGG directly produces the summed output of its branches. TLR produces both results for the current frame and low-rank projections for subsequent frames: the first group is added to a correction computed from cached historical projections through a shared pointwise convolution; the other two groups are cached for reuse in subsequent frames. The final output still has CoutC_{\mathrm{out}} channels (Eqs. 8–11; Fig. 3).

(3) Block pruning and feature recovery: differences from TinyFusion and CrossKD

Our block pruning builds on TinyFusion’s search strategy, which we extend to convolutional residual blocks in video VAEs.

Both CrossKD and LightVAE use a frozen teacher suffix, but use its loss differently. CrossKD-style distillation may not improve student reconstruction due to teacher–student suffix mismatch after pruning. LightVAE instead selects feature-matching directions that locally agree with probe-loss descent, achieving better recovery than ordinary feature MSE and reconstruction-only training (Fig. 8).

CrossKD uses the suffix loss as a direct training objective. It passes student features through frozen downstream teacher modules and optimizes student parameters by matching the resulting predictions to the teacher’s predictions.

LightVAE uses the suffix loss to construct feature-matching weights. We pass student features through the frozen teacher suffix, compute the probe-loss gradient with respect to those features, and combine it with signed student–teacher feature differences to assign a weight to each component. If moving a component toward its teacher counterpart would locally increase the probe loss, we set its matching weight to zero, avoiding the uniform supervision imposed by ordinary feature MSE. The weights are detached during training, and the probe loss itself is not directly included in the training objective.

AuthorsOpenReview ↗

Response to Reviewer eTbr (part 2/4)

Response to Weakness 2

Thank you for raising this concern. The comparison in Table 1 is fair: the Wan2.2 VAE baseline follows its official causal decoding setup, using chunk-by-chunk decoding with feature caching. The 81-frame setting denotes the total output video length, not the decoding chunk size. This decoding path follows the official implementation (wan/modules/vae2_2.py) and is preserved in our submitted anonymous code (third_party/wan22/model.py). We have clarified this setting in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer eTbr (part 3/4)

Response to Question 1

Thank you for your comments. Table 1 fairly compares decoding speed and reconstruction quality under matched conditions.

For speed, decoders within each backbone receive identical complete latent inputs before timing starts, with matched hardware, output length, and resolution, while retaining their native execution paths. Both causal and non-causal decoders can begin decoding immediately, without waiting for additional inputs.

For quality, each decoder uses its native decoding procedure and receives identical latents from the same encoder. Reconstructions are evaluated against the same source videos using the same metrics, allowing a fair comparison of reconstruction quality.

We also examined video length as a potential source of bias. Turbo-VAED’s original paper [1] uses a standard test length of 17 frames, and its official Wan example also defaults to 17 frames, whereas Table 1’s Wan latency benchmarks use 81 frames. We therefore added a matched 17-frame evaluation, retaining each method’s native decoding path and Table 1’s resolution for latency measurements.

In this short-video setting, Pro outperforms Turbo-VAED on all three DAVIS reconstruction metrics and has lower measured decoding latency on both GPUs.

DecoderDAVIS PSNR ↑DAVIS SSIM ↑DAVIS LPIPS ↓H100 latency (ms) ↓RTX 5090 latency (ms) ↓
Original Wan2.2 VAE34.98280.935040.06871820.5721817.668
Turbo-VAED32.83220.919460.10155566.818898.523
LightVAE-Pro33.95290.926410.08433413.080877.719
LightVAE-Lite31.89650.900490.12009247.415507.019

We have included these results and clarified each method’s decoding protocol in the revised manuscript.

References

[1] Ya Zou et al. Turbo-VAED: Fast and Stable Transfer of Video-VAEs to Mobile Devices. Proceedings of the AAAI Conference on Artificial Intelligence, 40(16), 14086–14094, 2026.

Reviewer_EBSPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the high inference cost of video VAEs in video generation systems and proposes LightVAE, a compression framework featuring Temporal Low-Rank Compression (TLR). TLR preserves dense spatial modeling while compressing only the temporal component. In addition, the paper introduces a general structural compression-and-recovery framework applicable to both convolutional and Transformer-based autoencoders.

Critical strengths
  1. The research problem is well-motivated and practically relevant. As diffusion denoising becomes faster, the computational overhead of VAE encoding and decoding becomes increasingly significant. The paper directly targets high-resolution causal Conv3D architectures in pretrained video VAEs while preserving the original latent interface, making its practical use cases clear.
  2. The core claim that spatial and temporal branches exhibit different sensitivities to compression is directly analyzed. Figure 5 reports stage-wise results showing that temporal-only approximation yields lower output errors than both spatial-only and joint approximation.
  3. The experimental validation is reasonably comprehensive. Across multiple tables, LightVAE demonstrates a favorable trade-off between inference speed and output quality.
Critical weaknesses
  1. The evidence for TLR's cross-architecture generalization remains limited. Although the paper covers three backbones, the core operator, TLR, is evaluated only on the closely related Wan2.1/2.2 VAE family, while MiniMax-H3 is used only to validate block pruning adapted from TinyFusion. The paper acknowledges this limitation in the Appendix, but it remains a key factor in assessing the broader impact of the contribution.
  2. If ranks can be globally allocated between the spatial and temporal branches under a shared MAC budget, rather than requiring each branch to account for half of the absolute savings, what is the optimal error of the joint baseline?
  3. Please quantify the actual end-to-end performance gains. For at least one Wan configuration, please report a latency breakdown for full video generation, including DiT denoising, VAE decoding, and other components, along with the end-to-end speedup achieved by replacing the original VAE with Pro/Lite. This would help clarify the practical impact of the Introduction's claim that the VAE is becoming a more significant bottleneck.
Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer EBSP (part 1/3)

Response to Weakness 1: Generalization beyond the Wan family

Thank you for your interest in TLR’s generalization across architectures. TLR compresses pretrained Conv3D operators and can be applied to Conv3D-based video VAEs without relying on Wan’s specific decoder layout. Such VAEs are used by more than ten open-weight video model families, including HunyuanVideo [1], LTX [2], CogVideoX [3], SkyReels [4], and LongCat-Video [5]. Exceptions include MiniMax-H3 [6], which uses a Transformer decoder.

To validate TLR’s generalization across architectures, we added compression experiments on the VAEs of HunyuanVideo-1.5 and LTX-2.5. Compared with the corresponding original decoders, LightVAE-Pro achieves 1.74× and 1.35× H100 decoding speedups, respectively, with PSNR drops of 0.31–0.88 dB (smaller drops are better) across UCF101 and DAVIS. We have included these results and clarified TLR’s scope in the revised manuscript.

BackboneDecoderUCF101 PSNR ↑SSIM ↑LPIPS ↓DAVIS PSNR ↑SSIM ↑LPIPS ↓H100 FPS ↑RTX 5090 FPS ↑
HunyuanVideoOriginal33.4070.93640.087134.0910.91590.131227.0013.71
HunyuanVideoLightVAE-Pro33.0990.93370.088633.5760.91080.130146.9518.78
LTXOriginal31.7570.91930.095830.6490.86650.1701404.60175.45
LTXLightVAE-Pro31.1530.90760.102529.7670.85870.1905546.88266.35

References

[1] Bing Wu et al. HunyuanVideo 1.5 Technical Report. arXiv:2511.18870, 2025.

[2] Yoav HaCohen et al. LTX-2: Efficient Joint Audio-Visual Foundation Model. arXiv:2601.03233, 2026.

[3] Zhuoyi Yang et al. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. arXiv:2408.06072, 2024.

[4] Guibin Chen et al. SkyReels-V2: Infinite-length Film Generative Model. arXiv:2504.13074, 2025.

[5] Meituan LongCat Team et al. LongCat-Video Technical Report. arXiv:2510.22200, 2025.

[6] MiniMax. MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities. Official release article, 2026.

AuthorsOpenReview ↗

Response to Reviewer EBSP (part 2/3)

Response to Weakness 2: Budget allocation between spatial and temporal branches

Thank you for your question about joint budget allocation. Jointly selecting ranks across layers and branches creates a large discrete search space, making exhaustive evaluation costly. We therefore conducted a controlled sweep of fixed MAC-savings splits for joint spatial–temporal compression: the spatial branch’s share of the required savings increases from 10% to 50% in 10-percentage-point increments, with the remainder assigned to the temporal branch. These correspond to the 1:9 through 5:5 splits below, compared with temporal-only compression (TLR) and spatial-only compression at each target MAC budget.

Across all tested MAC budgets, TLR achieves the lowest error among the tested allocations, while error generally increases as the spatial branch takes on a larger share of the savings.

The table reports mean relative L2L_2 error against the original decoder. At 50% of the original MACs, TLR achieves an error of 0.087, while increasing the spatial share from 10% to 50% raises the error from 0.326 to 0.398. This supports concentrating compression on MtempM_{\mathrm{temp}} while preserving dense SS.

Allocation40%50%60%70%80%90%95%
TLR (0:10)0.1590.0870.0660.0520.0420.0300.026
Joint compression, savings split 1:90.3340.3260.3210.1470.1410.0470.038
Joint compression, savings split 2:80.3330.3270.3230.3200.3170.1380.045
Joint compression, savings split 3:70.3460.3350.3260.3220.3180.1410.135
Joint compression, savings split 4:60.3820.3540.3350.3240.3200.3160.138
Joint compression, savings split 5:51.1620.3980.3540.3320.3210.3180.139
Spatial-only (10:0)———1.1590.3550.3250.321

We have included these results in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer EBSP (part 3/3)

Response to Weakness 3: End-to-end performance gains

Thank you for suggesting an evaluation of the actual end-to-end gains. Using LightX2V [3], we tested two Wan2.2-TI2V-5B variants on a single consumer RTX 5090: Wan2.2-TI2V-5B-Turbo [1], which uses bidirectional temporal attention and four denoising steps, and the autoregressive EdenLive [2], which uses four denoising steps per chunk.

In bidirectional generation, the original VAE accounts for 60.0% of total latency. Pro/Lite provide 1.43×/1.75× end-to-end speedups and reduce VAE decoding time below DiT denoising time. In autoregressive generation, the original VAE takes approximately 555 ms per regular 12-frame chunk, exceeding the 500 ms budget for 24 FPS generation. Pro/Lite reduce this to approximately 249/143 ms and raise end-to-end throughput from 15.72 FPS to 26.22/34.16 FPS, enabling real-time generation at 24 FPS on a single GPU in steady state. We have included these results in the revised manuscript.

(1) Four-step bidirectional generation (121 frames, 1280 × 704, W × H)

DecoderDiT denoising (s)VAE decoding (s)Other components (s)End-to-end latency (s)End-to-end speedup
Original Wan2.2 VAE7.937 ± 0.05012.716 ± 0.0310.526 ± 0.01421.179 ± 0.0871.00×
LightVAE-Pro7.895 ± 0.0486.420 ± 0.0240.522 ± 0.00414.837 ± 0.0701.43×
LightVAE-Lite7.886 ± 0.0493.709 ± 0.0110.525 ± 0.00512.120 ± 0.0651.75×

(2) Four-step autoregressive generation: meeting the 24 FPS real-time target (717 frames, 832 × 480, W × H)

DecoderVAE decoding per regular chunk (ms)End-to-end latency (s)End-to-end speedupEnd-to-end FPSSteady-state chunk interval, mean / P95 (ms)
Original Wan2.2 VAE≈55545.611 ± 0.0091.00×15.72762.0 / 791.8
LightVAE-Pro≈24927.346 ± 0.0201.67×26.22456.5 / 486.7
LightVAE-Lite≈14320.988 ± 0.0382.17×34.16350.1 / 380.1

References

[1] Quanhao Li et al. Wan2.2-TI2V-5B-Turbo. Model and code repository, 2025.

[2] Zehuan Huang et al. EdenLive: Real-Time Generative Rendering from Any Visual Reference. Technical report, 2026.

[3] LightX2V Contributors. LightX2V: Light Video Generation Inference Framework. Software, 2025.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.