SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

LONGLIVE - An Interactive Long Video Generation Framework Introduced by NVIDIA and Others

Author: AI小集 · Original source: AI工具集

What Is LONGLIVE

LONGLIVE is a real-time interactive longvideo generationframework jointly introduced by NVIDIA and other leading institutions. The framework addresses the dual bottlenecks of efficiency and quality in long video generation through a frame-level autoregressive (AR) model, combined with the KV-recache mechanism, streaming long video fine-tuning, and short-window attention + frame sink technology. LONGLIVE can generate high-quality videos up to 240 seconds long at 20.7 FPS on a single H100 GPU, supporting real-time prompt switching and dynamic adjustment, opening up new creative possibilities for fields such as creativity, education, and film and television, and representing a key step in advancing AI video generation from a "toy" to a "productivity tool."

LONGLIVE

LONGLIVE's Main Features

  • Real-time interaction: Supports users in inputting streaming prompts in real time during video generation, dynamically adjusting video content, and guiding the narrative or changing the style.
  • Long video generation: The model can generate high-quality videos several minutes long, supporting complex narratives and scene development.
  • Efficient inference: Achieves a real-time speed of 20.7 FPS on a single NVIDIA H100 GPU, supports video generation up to 240 seconds long, while maintaining high fidelity and temporal continuity.
  • High-quality generation: Ensures the visual coherence and semantic consistency of generated videos through innovative technology, maintaining smooth transitions when prompts are switched frequently.
  • Low deployment cost: Supports INT8 quantized inference, further reducing model size and deployment cost with almost no performance loss.

LONGLIVE's Technical Principles

  • KV-recache mechanism: When switching prompts, the key-value (KV) cache state is "refreshed" by recomputation, clearing residual information from the old prompt while retaining visual and motion cues, ensuring smooth transitions in the scene and precise execution of new instructions. The recache operation is integrated into training, enabling the model to learn during training how to transition smoothly after prompt switching.
  • Streaming Long Tuning: Addresses the quality degradation problem of autoregressive (AR) models in long video generation, simulating the inference process through "rolling extension" and reducing inconsistency between training and inference. Local supervision and gradient detachment are used to avoid out-of-memory (OOM) issues caused by backpropagation through long sequences, ensuring reliable supervision from the teacher model.
  • Short-window attention + Frame Sink: Based on short-window attention, the attention range is limited to a local window, significantly reducing computational complexity and memory requirements. A frame sink mechanism is introduced, restoring long-range consistency by retaining global anchors (such as the first frame block of the video), while maintaining the efficiency of the short window.

LONGLIVE's Project Addresses

  • GitHub repository：https://github.com/NVlabs/LongLive
  • HuggingFace model library：https://huggingface.co/Efficient-Large-Model/LongLive-1.3B
  • arXiv technical paper：https://arxiv.org/pdf/2509.22622

LONGLIVE's Application Scenarios

  • Creative video production: Creators can adjust video content and style in real time, quickly generating long videos that meet creative needs, improving creative efficiency and flexibility.
  • Educational content generation: Teachers can generate teaching videos in real time according to teaching progress, dynamically inserting knowledge points or cases, enhancing the interactivity and interest of teaching.
  • Film and television production: Directors and screenwriters can preview different scenes and narrative paths in real time before filming, quickly adjust scripts and shooting plans, and reduce production costs.
  • Advertising creation: Advertising teams can generate advertising videos in real time according to client needs, quickly adjust creative directions, and improve the relevance and appeal of advertisements.
  • Game development: Developers can generate game cutscenes or dynamic backgrounds in real time, adjust content in real time according to the game storyline, and enhance players' sense of immersion.