NVlabs/LongLive

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

2563 stars · 248 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

LongLive is an NVIDIA research framework providing parallel training and inference infrastructure for real-time long video generation, using NVFP4/FP8 quantization, KV-cache optimizations, and multi-shot autoregressive techniques. It achieves real-time frame rates (45.7 FPS) for generating minute-long and interactive videos with diffusion models like Wan2.2 and SANA-Video.

Use cases

• generate long videos in real time from text prompts
• train autoregressive video generation models with parallelism
• run FP4/FP8 quantized video diffusion inference on GPUs
• generate interactive videos that respond to prompts mid-stream
• distill video diffusion models for faster inference
• compress KV cache for ultra-long video generation

When to choose

• you need real-time or interactive long video generation on NVIDIA GPUs
• you want to train or fine-tune video diffusion models with efficient parallel infrastructure
• you need quantized (NVFP4/FP8) inference for video generation throughput

When to avoid

• you need simple one-off video generation without GPU infrastructure
• you work outside the PyTorch/NVIDIA CUDA ecosystem
• you need production video editing rather than generative video synthesis

Facets

framework · maturity active

video-processing llm-inference llm-training gpu-computing machine-learning deep-learning gpu-computing artificial-intelligence python video-generation diffusion-models nvfp4 quantization real-time-inference kv-cache parallelism nvidia video gpu linux docker

Member repositories

For agents

markdown · JSON · MCP: product_card(name="NVlabs/LongLive")