Reproduce “LUVE : Latent-Cascaded Ultra-High-Resolution Video Generation with Dual Frequency Experts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chen Zhao, Jiawei Chen, Hongyu Li, Zhuoliang Kang, Shilin Lu, Xiaoming Wei, Kai Zhang, Jian Yang and 1 more
Researchers at Nanjing University, Meituan, and Nanyang Technological University developed LUVE, a latent-cascaded framework featuring a Video Latent Upsampler and dual frequency experts for ultra-high-resolution video generation. This approach achieves superior photorealism, temporal consistency, and content fidelity, outperforming existing methods and demonstrating improved computational efficiency.
Key insights
- Motion dynamics are more effectively learned and established at lower resolutions, providing robust priors that significantly mitigate temporal dynamics degradation in subsequent UHR generation.
- Upsampling video latents directly within the latent space, using a specialized upsampler (VLUer), offers computational efficiency and reduces artifacts compared to conventional RGB or simple latent interpolation methods.
- Video diffusion models exhibit frequency-progressive denoising, where global semantic structures are primarily reconstructed in high-noise stages and fine-grained details in low-noise stages, enabling targeted refinement with frequency-specific experts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.