What is davinci-magihuman?

daVinci-MagiHuman is a research project that generates synchronized video and audio from text or a reference image using a single unified transformer, rather than separate models stitched together. The paper behind it is titled Speed by Simplicity, and the core idea is a fifteen billion parameter, forty layer transformer that processes text, video, and audio tokens together through self attention alone, with no separate cross attention step or multiple parallel model streams. The architecture uses what the authors call a sandwich design, where the first and last four layers have modality specific processing while the middle thirty two layers share parameters across text, video, and audio. It skips explicit timestep embeddings during the denoising process, instead inferring the current state directly from the input, and uses learned per head gating for training stability. The model supports six languages, including Mandarin, Cantonese, English, Japanese, Korean, German, and French, and reports fast inference: a five second video at 256p resolution in about two seconds, and a five second video at 1080p in under forty seconds, both on a single H100 GPU. In human evaluation across two thousand paired comparisons, the authors report an eighty percent win rate against one existing avatar model and a sixty one percent win rate against another. Setup is available through a prebuilt Docker image that supports the full pipeline including high resolution output, or a manual Conda environment covering PyTorch, an attention library appropriate to the GPU, and a companion compiler project for speed. Running the model requires downloading several external components separately, including a text encoder, an audio model, and a video encoder, all hosted on Hugging Face, in addition to the project's own released weights. The project supports both text to video and image plus text to video generation modes. It is released under the Apache 2.0 license, and the full model stack, including base, distilled, and super resolution components, along with inference code, is openly available.