Search Results for "mixture"

Wan2.2 is a major upgrade to the Wan series of open and advanced large-scale video generative models, incorporating cutting-edge innovations to boost video generation quality and efficiency. It introduces a Mixture-of-Experts (MoE) architecture that splits the denoising process across specialized expert models, increasing total model capacity without raising computational costs. Wan2.2 integrates meticulously curated cinematic aesthetic data, enabling precise control over lighting, composition, color tone, and more, for high-quality, customizable video styles. ...

...The model supports text-to-video and image-to-video generation tasks with flexible resolution options suitable for various GPU hardware configurations. Wan2.1’s architecture balances generation quality and inference cost, paving the way for later improvements seen in Wan2.2 such as Mixture-of-Experts and enhanced aesthetics. It was trained on large-scale video and image datasets, providing generalization across diverse scenes and motion patterns.