🚀 We open-sourced LongLive — interactive, real-time long-video generation. 👥Generates video in real time as users enter text prompts. ⚡️20.7 FPS on a single H100,⏱️up to 240s per clip. 🎬Fine-tunes SOTA short-video models (e.g., Wan) into long-video generators. 🌍One step closer to World Models. All code for training &... inference, model weights, demo page, and videos released! Paper: Code: Model: Demo Page: Introduction Video:show more

Here is the technical introduction video.

All code for training & inference, model weight, demo page, and videos released! Paper: Code: Model: Page: Video: #NVIDIA #MIT #AI

Cool work! qq: How is the VAE used to encode/decode? Is the output latent frame chunks decoded on the fly (which is troublesome since Wan2.1's VAE takes the 1st latent frame as a single RGB frame and the other latent frames to be 4 RGB frames), or are the latent frames buffered to be decoded, e.g., encode/decode 21 (or more) latent frames each time?

Cool work! Did you analyze what the bottleneck is if you want to higher than 27 fps? Surely that's not the limit

Thanks for your interest in our work. Because we already used short window attention, thus the bottle neck is not in the attention, but in other FFN layers. We can achieve around 25 FPS, if we use FP8 quantization and torch.compile on H100 GPUs.