LongLive is a long video generation and World Action Model infrastructure repository launched by NVIDIA NVlabs, with research directions covering video generation, world models, and world action modeling in embodied scenarios. Each project in the repository is self-contained, each with its own directory, code, documentation, and model weights, and the desired projects can be downloaded on demand via partial clone and sparse checkout, and can run without other projects.
[Main Projects] Long-WAM: a long-context world action model, pretrained on robot videos based on LongLive 2.0, with a long-context robot policy featuring streaming visual memory. Benchmarks include LIBERO, RoboTwin 2.0, Domino, RoboCasa GR1, and RoboCasa365; inference supports streaming IDM/COD, a configurable Agentic API, and image-to-video demos; deployment provides runtime configurations for RTX 5090, DGX Spark, and AGX Thor, as well as YAM, Franka, and Unitree G1 interfaces. LongLive-Plug: a once-for-all distillation method for video generation that separates reusable capabilities from task-specific customization—training functional LoRA on a base model, then mounting it onto compatible downstream models while retaining the downstream models' task-specific weights, supporting Wan2.1-14B, Wan2.2-TI2V-5B, and MiniMax-H3. LongLive 2.0: NVFP4 parallel infrastructure for long video generation. On the training side, it supports balanced sequence parallelism for T2V/I2V autoregressive (AR) training (teacher-forcing), T2V/I2V AR training on multi-shot (or single-shot) videos, and NVFP4 (or BF16) in AR training and few-step distillation. On the inference side, it supports NVFP4 inference (W4A4) and NVFP4 KV Cache, TorchAO FP8 PTQ inference (W8A8) from BF16 checkpoints, multi-shot attention sink, sequence-parallel inference, and asynchronous decoding. LongLive 1.0: real-time interactive long video generation, which can receive continuous user prompts and generate corresponding videos in real time, enabling people to continuously guide the direction of long videos during the video generation process. The core ideaIncluding attention sink, KV-recache, and streaming long-term fine-tuning. LongLive-RAG: A general retrieval-augmented framework for long video generation.
[Version and feature updates (according to official release records)]
- 2026.10.08: The Long-WAM paper goes live on arXiv.
- 2026.10.07: Release of Long-WAM.
- 2026.09.28: Release of LongLive-Plug training and inference code, including four recipes covering CFG and few-step distillation.
- 2026.07.08: LongLive 2.0 supports FP8 inference.
- 2026.06.01: Release of LongLive-RAG.
- 2026.05.30: LongLive 2.0 supports I2V AR teacher-forcing training and I2V DMD distillation for Wan2.2-TI2V-5B.
- 2026.05.25: Optimized the NVFP4 inference path, improving overall throughput by 18.6%.
- 2026.05.13: Release of LongLive 2.0, supporting NVFP4, parallel and multi-shot AR training, DMD distillation and inference (45.7 FPS).
- 2026.04.12: Support for TriAttention-based KV cache compression, reducing KV by 50% with no loss in quality.
- 2026.01.27: LongLive was accepted by ICLR-2026.
- 2026.01.11: Support for adapting the original RoPE to KV-cache relative RoPE, enabling generation of infinitely long videos.
- 2025.11.03: Implemented LongLive on the linear attention model SANA-Video, enabling real-time generation of 60-second interactive videos.
- 2025.09.29: Release of the paper, GitHub repository, model weights LongLive-1.3B, and demo page.
[Usage and ecosystem] The project provides training and inference code, public checkpoints (hosted on Efficient-Large-Model), download and inference instructions, as well as benchmark training/evaluation and real-device deployment demos. FPS statistics are calculated according to the LongLive 1.0 and Self Forcing protocols for the generation frame rate within diffusion time, excluding VAE decoding. The repository has spawned multiple works based on LongLive, such as DreamForge-World 0.1, DreamX-World 1.0, SANA-Video (LongSANA), Daydream Scope, MemFlow, ShotStream, Stream-T1, KVPO, LoL, TriAttention, StreamEdit, Forcing-KV, Dummy Forcing, MemRoPE, Astrolabe, OPSD-V, etc. The project is released under the Apache License 2.0, and each directory also includes its own copy of the license and applicable third-party notices. The technical approach is based on foundational works such as Self-Forcing, Wan2.1, Wan2.2, and MiniMax-H3.
[Limitations and information not yet provided] Official materials do not explain charging, pricing, commercial licensing methods, the form of an online product for ordinary users, the scope of platform support beyond specific minimum hardware requirements, detailed usage tutorials, or how non-technical users can get started; this content is not yet available.
