DaVinci MagiHuman

DaVinci MagiHuman is an open audio-video generation model for talking avatars. Given one portrait photo as the visual anchor and either a text script or an audio file as the driver, it produces a video of the person speaking with lips aligned to the speech. Unlike pipelines that run a text-to-speech step and a separate lip-sync step, MagiHuman uses a single-stream Transformer that jointly denoises video and audio tokens, aligning speech and frames within one model pass.

The model has 15B parameters and is released by Sand.ai and GAIR Lab (Shanghai Jiao Tong University) under the Apache 2.0 license, so you can inspect the weights, run inference locally or on a server, and use it commercially within the license terms. On a single NVIDIA H100-class GPU, the project reports roughly two seconds of wall time to generate a short 256p clip, and it supports multiple languages for lip sync (coverage depends on the training data and release notes). Published benchmark-style comparisons report strong word-error-rate and human-preference results against public baselines such as Ovi 1.1 and LTX 2.3.

This particular website is an independent, third-party page rather than the official research lab site. It offers an online studio where you can upload a portrait, enter a script or audio, choose a resolution and generate a clip, and it links to the official open-source demos and Hugging Face checkpoints. Because it runs the open model, results and hardware requirements follow the upstream release.