LongLive

Why we included this project

Most video models fall apart when you ask for more than a few seconds of output: memory blows up, inference slows down, and the context window limits how much you can generate. LongLive, an NVIDIA infrastructure project rather than a single model, is built for that problem. It pairs NVFP4 low-precision quantization with parallel paths for training, distillation, and inference, which is how multi-shot generations reach interactive frame rates. The repo also ships FP8 inference support, KV-cache compression through TriAttention, and a retrieval-augmented framework for long-video generation, so you can adopt pieces one at a time instead of reworking your whole pipeline. If you already run diffusion video models like Wan and want minutes-long output, or you need teacher-forcing and DMD distillation workflows, the code is a practical reference for production-grade optimization.