DaVinci MagiHuman

Give DaVinci MagiHuman one portrait photo and either a script or an audio file, and you get back a talking video with the lips matched to the speech.

DaVinci MagiHuman is an open-source AI avatar video generator, a 15B-parameter model developed by Sand.ai and GAIR Lab at Shanghai Jiao Tong University.

Most lip-sync setups chain a text-to-speech step to a separate video step.

DaVinci MagiHuman takes a different route: a single-stream Transformer denoises video and audio tokens together in one pass.

That design is the main thing that sets it apart from pipeline-style tools.

The license is Apache 2.0, so you can inspect the weights, run inference on your own machine and use the output commercially.

Speed is a selling point too.

Short clips can come out in seconds on professional hardware such as an NVIDIA H100.

That also hints at the catch: this is a model for people comfortable with local inference and serious GPUs, not a point-and-click app for casual use.