Wan: Open and Advanced Large-Scale Video Generative Models - Wan-Video/Wan2.2

Hacker News is a community-driven platform for sharing and discussing technology news, startups, and programming-related topics. Through user submissions and comments, Hacker News offers insights into emerging technology trends, industry developments, and entrepreneurial ventures. Readers can participate in discussions, share their insights, and stay informed about the latest advancements in technology and innovation.

Hacker News

Wan2.2 is an open-source large-scale video generation model that introduces Mixture-of-Experts (MoE) architecture to video diffusion models. The release includes three model variants: T2V-A14B and I2V-A14B (27B parameters with 14B active) for text-to-video and image-to-video generation at 480P/720P, and TI2V-5B (5B parameters) for efficient 720P@24fps generation on consumer GPUs. Key improvements include specialized expert models for different denoising stages, enhanced aesthetic control through curated training data, and a high-compression VAE achieving 64x compression ratio. The models support multi-GPU inference with FSDP and DeepSpeed Ulysses, integrate with ComfyUI and Diffusers, and demonstrate superior performance compared to leading commercial models on Wan-Bench 2.0 evaluation.

Wan-Video/Wan2.2: Wan: Open and Advanced Large-Scale Video Generative Models

Computational Efficiency on Different GPUs