FlashI2V: Fourier-Guided Latent Shifting Prevents Conditional Image Leakage in Image-to-Video Generation

If you like our project, please give us a star ⭐ on GitHub for latest update.

💡 We also have other generation projects that may interest you ✨.

Open-Sora Plan: Open-Source Large Video Generation Model
Bin Lin, Yunyang Ge and Xinhua Cheng etc.

UniWorld-V1: High-Resolution Semantic Encoders for Unified Visual Understanding and Generation
Bin Lin, Zongjian Li, Xinhua Cheng etc.

Identity-Preserving Text-to-Video Generation by Frequency Decomposition
Shenghai Yuan, Jinfa Huang, Xianyi He etc.

📣 News

[2025.09.30] We have uploaded the Ascend version of the training and inference code, along with the model weights. For details, please refer to the NPU branch.

🗓️ TODO

😍 Gallery

Image-to-Video Results of FlashI2V-1.3B

video_000007.mp4

video_000021.mp4

video_000027.mp4

video_000035.mp4

video_000109.mp4

video_000142.mp4

video_000149.mp4

video_000151.mp4

video_000163.mp4

video_000191.mp4

video_000273.mp4

video_000280.mp4

video_000172.mp4

video_000184.mp4

video_000214.mp4

video_000352.mp4

😮 Highlights

Overfitting to In-domain Data Causes Performance Degradation

Existing I2V Methods involve Conditional image leakage. (a) Conditional image leakage causes performance degradation issues, where the videos are sampled from Wan2.1-I2V-14B-480P with Vbench-I2V text-image pairs. (b) In the existing I2V paradigm, we observe that chunk-wise FVD on in-domain data increases over time, while chunk-wise FVD on out-of-domain data remains consistently high, indicating that the law learned on in-domain data by the existing paradigm fails to generalize to out-of-domain data.

Model Overview

We propose FlashI2V to introduce conditions implicitly. We extract features from the conditional image latents using a learnable projection, followed by the latent shifting to obtain a renewed intermediate state that implicitly contains the condition. Simultaneously, the conditional image latents undergo the Fourier Transform to extract high-frequency magnitude features as guidance, which are concatenated with noisy latents and injected into DiT. During inference, we begin with the shifted noise and progressively denoise following the ODE, ultimately decoding the video.

Best Generalization and Performance across Different I2V Paradigms

Comparing the chunk-wise FVD variation patterns of different I2V paradigms on both the training and validation sets, it is observed that only FlashI2V exhibits the same time-increasing FVD variation pattern in both sets. This suggests that only FlashI2V is capable of applying the generation law learned from in-domain data to out-of-domain data. Additionally, FlashI2V has the lowest out-of-domain FVD, demonstrating its performance advantage.

Vbench Results

Model	I2V Paradigm	Subject Consistency↑	Background Consistency↑	Motion Smoothness↑	Dynamic Degree↑	Aesthetic Quality↑	Imaging Quality↑	I2V Subject Consistency↑	I2V Background Consistency↑
SVD-XT-1.0 (1.5B)	Repeating Concat and Adding Noise	95.52	96.61	98.09	52.36	60.15	69.80	97.52	97.63
SVD-XT-1.1 (1.5B)	Repeating Concat and Adding Noise	95.42	96.77	98.12	43.17	60.23	70.23	97.51	97.62
SEINE-512x512 (1.8B)	Inpainting	95.28	97.12	97.12	27.07	64.55	71.39	97.15	96.94
CogVideoX-5B-I2V	Zero-padding Concat and Adding Noise	94.34	96.42	98.40	33.17	61.87	70.01	97.19	96.74
Wan2.1-I2V-14B-720P	Inpainting	94.86	97.07	97.90	51.38	64.75	70.44	96.95	96.44
CogVideoX1.5-5B-I2V†	Zero-padding Concat and Adding Noise	95.04	96.52	98.47	37.48	62.68	70.99	97.78	98.73
Wan2.1-I2V-14B-480P†	Inpainting	95.68	97.44	98.46	45.20	61.44	70.37	97.83	99.08
FlashI2V† (1.3B)	FlashI2V	95.13	96.36	98.35	53.01	62.34	69.41	97.67	98.72

† means testing with recaptioned text-image-pairs in Vbench-I2V.

🔒 License

See LICENSE for details. For Ascend version, you can see LICENSE in NPU branch.

🤝 Contributors

🙏 Acknowledgements

Wan2.1 - https://github.com/Wan-Video/Wan2.1
Open-Sora Plan - https://github.com/PKU-YuanGroup/Open-Sora-Plan
Mindspeed-MM - https://gitee.com/ascend/MindSpeed-MM
Megatron-LM - https://github.com/NVIDIA/Megatron-LM

✏️ Citation

If you want to cite our work, please follow:

@misc{ge2025flashi2v,
      title={FlashI2V: Fourier-Guided Latent Shifting Prevents Conditional Image Leakage in Image-to-Video Generation}, 
      author={Yunyang Ge and Xinhua Cheng and Chengshu Zhao and Xianyi He and Shenghai Yuan and Bin Lin and Bin Zhu and Li Yuan},
      year={2025},
      eprint={2509.25187},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.25187}, 
}

splinter21/FlashI2V