daVinci-MagiHuman Released

daVinci-MagiHuman is a 15B parameter single-stream transformer released by SII-GAIR and Sand.ai that jointly processes text, video, and audio in one model, improving lip-sync and facial expression alignment in human videos.

• Supported languages: English, Mandarin/Cantonese, Japanese, Korean, German, French
• Release configuration: base / distilled / super resolution model stack
• License: Apache 2.0

The architecture is a sandwich design where the front and back 4 layers each handle modality-specific processing while the middle 32 layers are shared. With DMD-2 distillation, it uses only 8 denoising steps, and the super resolution stage upscales 256p output to 540p or 1080p in latent space, avoiding additional VAE round-trips.

In official evaluation, across 2000 pairwise comparisons it outperformed Ovi 1.1 at 80% and LTX 2.3 at 60.9%, with metrics of visual quality 4.80, text alignment 4.18, physical consistency 4.52, and voice WER 14.60%. On a single H100 GPU, generation of 5 seconds/256p takes 2 seconds, and 1080p takes 38 seconds, as presented.

There are also constraints for practical use. An H100 GPU is essentially required, 3 external models must be downloaded separately before running, and the first generation may be slower than the official speed due to compile warmup. An online demo is available on HuggingFace Spaces, and a getting-started guide is organized on the official GitHub.