15B Parameter Audio-Video Portrait Model Open-Sourced: Sand.ai Joint Team Builds, 2 Seconds to Generate Video on Single H100

Sand.ai and Shanghai Chuangzhi Academy GAIR Lab jointly open-sourced the audio-video joint generation base model daVinci-MagiHuman, with 15 billion parameters. The model uses a single-stream Transformer architecture, unifying text, video, and audio into the same token sequence, and completes joint modeling via self-attention without cross-attention or modality-specific branches.

Focused on portrait scenes, the model can generate natural facial expressions and voices, achieving precise synchronization of lip movements, expressions, and actions, supporting six languages including Chinese, English, and Japanese. With techniques like distillation and super-resolution during inference, a single H100 GPU can generate a 5-second 256p video in just 2 seconds.

Objective evaluations show the model ranks first in visual quality and text consistency among open-source models, with better speech intelligibility than LTX 2.3 and Ovi 1.1. The complete model stack (base model, distilled model, super-resolution model, and inference code) has been fully open-sourced.