Teaser: FreeSwim generates ultra-high-resolution videos with rich details and coherent global structure. ✨

FreeSwim introduces a novel training-free approach for generating ultra-high-resolution videos, tackling the high computational cost of training large video models. Our method leverages pre-trained Diffusion Transformers (DiTs) at lower resolutions and synthesizes high-resolution videos efficiently without additional training. At the heart of FreeSwim is a unique sliding-window attention mechanism that preserves fine-grained visual details and high fidelity, addressing the inherent issues of repetitive patterns and lack of global coherence in video generation. This paper presents an innovative dual-path pipeline with a cross-attention override strategy, ensuring both local details and semantic consistency are maintained in high-resolution video generation.

🌟 Key Highlights:

• Training-Free: Generate high-quality, high-resolution videos without additional model training.
• Efficient: Achieves 2x speedup with cross-attention caching.
• High-Resolution: Synthesize 4K resolution videos with fine-grained details.

• Training-Free Ultra-High-Resolution Video Generation: 🎥 We introduce FreeSwim, a method that utilizes pre-trained DiTs to generate videos at high resolutions without the need for further training or model adaptation.
• Inward Sliding-Window Attention: 🔍 A critical component that maintains a local receptive field size during inference, improving detail preservation and reducing artifacts.
• Dual-Path Architecture: 🔄 Combines local window attention with a full-attention branch to ensure semantic accuracy and avoid content repetition.
• Efficiency Improvements: ⚡ Implements a cross-attention caching strategy, significantly speeding up inference without sacrificing video quality, achieving over 2x speedup.

• Coarse-to-Fine Generation: 🖥️ The initial video is generated at the model's native resolution and refined through a high-resolution upsampling process.
• Inward Sliding-Window Attention: 🔄 Ensures that the receptive field during inference matches the training resolution, maintaining local coherence.
• Cross-Attention Override: 🌐 A dual-path mechanism where the full-attention branch provides global semantic guidance to the window attention branch, ensuring high-quality output without repetitive patterns.
• Feature Reuse Strategy: 🔁 Reduces computational costs by reusing cross-attention features from the full-attention branch during inference, enabling high-resolution generation with minimal performance loss.

Figure: Overview of the FreeSwim framework with inward sliding-window attention, dual-path design, and cross-attention cache & reuse. 🧩

FreeSwim outperforms previous state-of-the-art methods in both video quality and efficiency on VBench, with significant improvements in aesthetic appeal, imaging quality, and overall consistency. It provides a training-free solution for generating 4K resolution videos, achieving fine-grained details and semantic consistency. 🌈

We recommend the following core library versions (FlexAttention behavior is sensitive to the PyTorch version):

pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu118

pip install diffusers==0.35.2

To run inference with FreeSwim, you need to specify the following command-line arguments:

python inference.py --mode nocache --target_height 1080 --target_width 1920

• A single 24 GB GPU (e.g., RTX 3090 / RTX 4090) is sufficient to generate 1080P, 81-frame videos with the default FreeSwim configuration.
• For the cache-based variant, if GPU memory becomes a bottleneck, the cached cross-attention features can be offloaded to the CPU to further reduce VRAM usage, at the cost of a slight slowdown in inference.

If you use FreeSwim in your research, please cite our paper:

@misc{wu2025freeswimrevisitingslidingwindowattention,
      title={FreeSwim: Revisiting Sliding-Window Attention Mechanisms for Training-Free Ultra-High-Resolution Video Generation}, 
      author={Yunfeng Wu and Jiayi Song and Zhenxiong Tan and Zihao He and Songhua Liu},
      year={2025},
      eprint={2511.14712},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.14712}, 
}