Wan-Weaver: Interleaved Multi-modal Generation (T2I & I2I)

Wan-Weaver is a new AI model that can generate text and images interactively, enabling applications like illustrated stories, fashion lookbooks, and children's books.

Details

Wan-Weaver is a novel AI model developed by Tongyi Lab at Tsinghua University, designed specifically for interleaved text and image generation. Unlike traditional text-to-image or image-to-image models, Wan-Weaver can generate text and images in an interactive, back-and-forth manner, similar to how humans create illustrated stories or social media posts. The key innovation is its Planner + Visualizer architecture, which decouples the text and image generation processes during training, allowing the model to learn the interplay between the two modalities without requiring real interleaved data. Instead, the researchers synthesized 'textual proxy' data to train the model. Wan-Weaver demonstrates strong long-range consistency, ensuring the text and images match across multiple steps. In benchmarks, it outperforms most open-source models and even rivals Google's commercial Nano Banana model in some metrics. This capability enables new applications like illustrated stories, fashion lookbooks, and children's books, where the text and visuals are tightly integrated.