MAGI-1

Why we included this project

MAGI-1 takes a different path from the usual one-shot video generators: it produces clips autoregressively, chunk by chunk, so it can stream frames as they finish and keep temporal consistency over longer sequences. The chunk-wise approach also makes scene transitions easier to control, which matters when you are building an actual pipeline rather than rendering a single demo clip. Pretrained weights ship at 4.5B and 24B parameters, with distilled and FP8-quantized versions, plus ComfyUI nodes so the model drops into existing workflows without much glue code. The repo also documents GPU footprints for each size and includes a T5 text encoder and a transformer-based VAE, the kind of practical detail that makes it runnable, not just readable.