Efficient High-Resolution Image-to-Video Generation
via Conditional Segment-wise Generation

YaoYang Liu¹ · Yuechen Zhang² · Wenbo Li³ · Yufei Zhao⁴ · Rui Liu⁵ · Long Chen^1,*

¹HKUST  ·  ²CUHK  ·  ³Joy Future Academy  ·  ⁴HKU  ·  ⁵HUAWEI Research
_{* Corresponding author}

2K (2560×1408) image-to-video generation — 81 frames in ~111s on a single H800, and runnable on a single RTX 4090 (24 GB).

• [2026-05] 📄 arXiv preprint released: arXiv:2605.06356
• [2026-05] 🌐 Project page online: hkust-longgroup.github.io/SwiftI2V
• Code and checkpoints will be released in this repository — stay tuned!

• 🚀 202× less GPU-time than end-to-end 2K I2V baselines
• 🖼️ Native 2K (2560×1408) generation at 81 frames
• 💻 Single consumer RTX 4090 (24 GB) is enough — no data-center GPU required
• 🧩 Conditional Segment-wise Generation (CSG) with bidirectional contextual interaction — memory-bounded regardless of video length
• 🎯 Stage-transition training closes the train–test gap between Stage I and Stage II

Image-to-video (I2V) generation has made rapid progress, yet scaling to high resolution (e.g., 2K) is bottlenecked by the efficiency–fidelity dilemma: end-to-end high-resolution generators deliver strong quality but require tens of thousands of GPU-seconds per clip, while low-resolution generation followed by video super-resolution (VSR) loses the input-image condition and hallucinates details inconsistent with the reference.

SwiftI2V is an efficient two-stage framework that resolves this dilemma:

• Stage I produces a low-resolution motion reference with a large backbone under few-step sampling.
• Stage II refines it to 2K, conditioned on both the input image and the Stage I output.
• Conditional Segment-wise Generation (CSG) divides the temporal axis into bounded segments augmented by neighboring contexts, keeping peak memory roughly constant regardless of total length.
• Stage-transition training simulates Stage I-style artifacts during Stage II training to close the train–test gap.

On VBench-I2V at 2K, SwiftI2V matches strong end-to-end baselines on key I2V metrics while reducing total GPU-time by 202×.

Overall two-stage pipeline of SwiftI2V.

Conditional Segment-wise Generation (CSG) with bidirectional contextual interaction.

For the full gallery, qualitative comparisons, ablations, and RTX 4090 results, please visit our Project Page.

• ✅️ Release project page
• ✅️ Release arXiv paper
• Release inference code
• Release Stage I / Stage II checkpoints

If you find SwiftI2V useful in your research, please consider citing:

@misc{liu2026swifti2vefficienthighresolutionimagetovideo,
      title={SwiftI2V: Efficient High-Resolution Image-to-Video Generation via Conditional Segment-wise Generation},
      author={YaoYang Liu and Yuechen Zhang and Wenbo Li and Yufei Zhao and Rui Liu and Long Chen},
      year={2026},
      eprint={2605.06356},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.06356}
}

We gratefully acknowledge the following open-source projects that made this work possible:

• Wan — the video generation foundation model we build upon.
• DiffSynth-Studio — for its flexible diffusion training and inference framework.
• LightX2V — for its efficient inference utilities.

The project page design is adapted from the Nerfies academic template (CC BY-SA 4.0).