TimeRipple: Accelerating vDiTs by Understanding the Spatio-Temporal Correlations in Latent Space

CVPR 2026

Wenxuan Miao¹, Yulin Sun¹, Aiyue Chen³, Jing Lin³, Yiwu Yao³, Yiming Gan⁴, Jieru Zhao¹, Jingwen Leng^1,2, Minyi Guo^1,2, Yu Feng^1,2,†

¹ Shanghai Jiao Tong University ² Shanghai Qi Zhi Institute ³ Huawei Technologies Co., Ltd. ⁴ Institute of Computing Technology, Chinese Academy of Sciences

^† Corresponding author

Paper

Demo Video

Comparisons

The comparisons show generation quality and acceleration performance on HunyuanVideo. PSNR is measured against the original dense-attention output.

“A 3D model of a 1800s Victorian house.”

TimeRipple_85%

31.28 dB PSNR · 2.66× Speedup

Original

PSNR Reference · 1.00× Speedup

Δ-DiT

26.09 dB PSNR · 1.22× Speedup

PAB_5,9

26.07 dB PSNR · 1.23× Speedup

MInference

25.00 dB PSNR · 1.44× Speedup

SVG_70%

25.78 dB PSNR · 1.69× Speedup

“A truck stuck in traffic during rush hour.”

TimeRipple_85%

31.28 dB PSNR · 2.66× Speedup

Original

PSNR Reference · 1.00× Speedup

Δ-DiT

26.09 dB PSNR · 1.22× Speedup

PAB_5,9

26.07 dB PSNR · 1.23× Speedup

MInference

25.00 dB PSNR · 1.44× Speedup

SVG_70%

25.78 dB PSNR · 1.69× Speedup

Sensitivity Study

Sensitivity of TurboDiffusion generation quality to the Q/K reuse detection window size. Results use Wan2.1-14B at 832×480 with 81 frames; each group shares the same prompt and random seed.

(1) “A hummingbird flaps its wings and hovers in front of a flower.”

2×2

4×4

8×8

16×16

Enlarged detail: flower center

2×2

16×16

(2) “An Arctic fox runs lightly across a vast expanse of snow.”

2×2

4×4

8×8

16×16

Enlarged detail: background sand texture

2×2

16×16

(3) “Fiery red maple trees reflected on an autumn lake, with a small drifting boat.”

2×2

4×4

8×8

16×16

Enlarged detail: boat and lake reflection

2×2

16×16

BibTeX

@misc{miao2025timeripple,
  title={TIMERIPPLE: Accelerating vDiTs by Understanding the Spatio-Temporal Correlations in Latent Space},
  author={Wenxuan Miao and Yulin Sun and Aiyue Chen and Jing Lin and Yiwu Yao and Yiming Gan and Jieru Zhao and Jingwen Leng and Minyi Guo and Yu Feng},
  year={2025},
  eprint={2511.12035},
  archivePrefix={arXiv},
  primaryClass={cs.AR},
  url={https://arxiv.org/abs/2511.12035}
}