Alibaba has introduced Wan2.2-S2V, an open-source model focused on creating digital human videos from speech inputs. This Speech-to-Video tool animates portrait photos into avatars that can speak, sing, and perform actions.

As part of the Wan2.2 video generation series, the model produces animated videos using a single image and an audio clip. It supports various framing options, such as portrait, bust, and full-body views. The system generates character actions and environmental elements based on prompt instructions, allowing creators to align visuals with specific narrative needs.

The model uses audio-driven animation to create performances, including dialogue and music, and can manage multiple characters in a scene. It accommodates different avatar types, from cartoons and animals to stylized figures.

Output resolutions include 480P and 720P, suitable for social media and professional uses.

Example Prompt and Application

For a prompt like: "In the video, a man is walking alongside the railway tracks, singing to express his emotions as he goes. A train slowly passes by him," the model can generate corresponding animated content.

Technical Features

Wan2.2-S2V combines text-guided global motion control with audio-driven local movements for character performances in various scenarios. It compresses historical frames into a compact latent representation to reduce computational needs and support stable long-video creation.

The model was trained on a large-scale audio-visual dataset tailored to film and television contexts, using a multi-resolution approach for flexible formats, including vertical short videos and horizontal productions.

Availability and Impact

The Wan2.2-S2V model is available for download on Hugging Face, GitHub, and Alibaba Cloud’s ModelScope community. Alibaba previously open-sourced Wan2.1 models in February 2025 and Wan2.2 models in July 2025. The Wan series has accumulated over 6.9 million downloads on Hugging Face and ModelScope.