Paper 2509.22622
LongLive: Real-time Interactive Long Video Generation
- Published
- Sep 2025
- Research lab
- NVIDIA
- Citations
- 189
- GitHub
- Not linked
01 In brief
Summary
LONGLIVE is a frame-level autoregressive (AR) framework for real-time, interactive long video generation, addressing efficiency and quality challenges in diffusion and AR models.
It introduces KV-recache to refresh cached states with new prompts for smooth, adherent prompt switches; streaming long tuning to enable train-long-test-long alignment; and short-window attention with a frame-level attention sink (frame sink) to preserve long-range consistency while accelerating inference.
LONGLIVE fine-tunes a 1.3B-parameter model in 32 GPU-days, achieving 20.7 FPS on a single NVIDIA H100, supporting up to 240-second videos, and maintaining strong VBench scores.
It also supports INT8 quantization with minimal quality loss.
Experiments show LONGLIVE outperforms baselines like SkyReels-V2 and Self-Forcing in interactive long-video quality and speed, with user studies confirming its effectiveness.
The framework is built on Wan2.1-T2V-1.3B and is open-sourced at https://github.com/NVlabs/LongLive.
Key limitations include dependence on the base model's capacity and lack of curated real-video data, bounding absolute quality.
Future work may incorporate supervised data to overcome this ceiling.
The paper includes detailed appendices on training, quantization, and qualitative showcases.
Overall, LONGLIVE demonstrates a practical solution for real-time interactive long video generation, balancing efficiency and quality.
Its contributions are significant for creative and cinematic applications.
The method's generalizability to other AR models is noted, with implementation on SANA-Video showing further acceleration.
The framework's design choices are validated through ablations, confirming the importance of long-video tuning for efficient inference strategies.
The paper concludes that LONGLIVE's approach is essential for advancing interactive video generation.
The code and model are publicly available for reproducibility.
The…
02 From the paper
Abstract
We present LongLive, a frame-level autoregressive (AR) framework for real-time and interactive long video generation. Long video generation presents challenges in both efficiency and quality. Diffusion and Diffusion-Forcing models can produce high-quality videos but suffer from low efficiency due to bidirectional attention. Causal attention AR models support KV caching for faster inference, but often degrade in quality on long videos due to memory challenges during long-video training. In addition, beyond static prompt-based generation, interactive capabilities, such as streaming prompt inputs, are critical for dynamic content creation, enabling users to guide narratives in real time. This interactive requirement significantly increases complexity, especially in ensuring visual consistency and semantic coherence during prompt transitions. To address these challenges, LongLive adopts a causal, frame-level AR design that integrates a KV-recache mechanism that refreshes cached states with new prompts for smooth, adherent switches; streaming long tuning to enable long video training and to align training and inference (train-long-test-long); and short window attention paired with a frame-level attention sink, shorten as frame sink, preserving long-range consistency while enabling faster generation. With these key designs, LongLive fine-tunes a 1.3B-parameter short-clip model to minute-long generation in just 32 GPU-days. At inference, LongLive sustains 20.7 FPS on a single NVIDIA H100, achieves strong performance on VBench in both short and long videos. LongLive supports up to 240-second videos on a single H100 GPU. LongLive further supports INT8-quantized inference with only marginal quality loss.