DaVinci MagiHuman

DaVinci MagiHuman: daVinci-MagiHuman is an advanced, open-source 15B-parameter AI model developed by Sand.ai and GAIR Lab at Shanghai Jiao Tong University. It is designed to generate high-quality, lip-synced talking videos from a single portrait image and a script or audio file. Unlike traditional methods that combine separate text-to-speech and video pipelines, daVinci-MagiHuman utilizes a unified single-stream Transformer to jointly denoise video and audio tokens simultaneously. Released under the Apache 2.0 license, it allows users to inspect weights, run inference locally, and use the technology for commercial purposes. It is optimized for speed, capable of generating short clips in just seconds on professional-grade hardware like the NVIDIA H100.