SandAI-org/MAGI-1

Implements a Transformer-based VAE architecture with causal temporal modeling to predict video in autoregressive chunks, enabling streaming generation and long-horizon synthesis. Supports diverse conditioning modalities including text prompts, images, and chunk-wise instructions for fine-grained control over scene transitions and video attributes. Available as distilled 4.5B variants and integrates with ComfyUI workflows plus Hugging Face ecosystem for inference and deployment.

3,663 stars. No commits in the last 6 months.