Wan 2.2 Releases With a Plethora Of New Features
Blog post from RunPod
Wan 2.2 marks a significant advancement in video generation technology, building upon its predecessor Wan 2.1 by introducing a Mixture-of-Experts (MoE) architecture and expanding its training dataset with 65.6% more images and 83.2% more videos. This new model employs a dual "high noise" and "low noise" approach to manage early and later stages of video denoising, allowing for enhanced customization and complexity in video generation. The release also includes the Text-Image-to-Video 5B (TI2V-5B) model, which supports 720P resolution video generation using text or image prompts on consumer-grade graphics cards. Despite the increased dataset size, Wan 2.2 maintains similar compute costs and memory usage as its predecessor, while improving performance and versatility. This update, which is compatible with previous tools like LoRAs, offers a range of new configuration options for developers and hobbyists, emphasizing both technical enhancements and practical deployment considerations on platforms like Runpod's GPU cloud.