LongLive: Real-time Interactive Long Video Generation

arXiv 25.9 2025 35.5 method

TLDR

LongLive is a real-time interactive long video generation framework using frame-level autoregressive design with KV-recache and streaming long tuning.

Reasoning

The paper presents a novel framework for long video generation that addresses efficiency and interactivity, achieving real-time performance. However, the abstract lacks quantitative quality comparisons and detailed evaluation metrics.

Read-first score

Read-first score 35.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 0.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=github

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Topical relevance 42%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 143.

Keyword Scores

world model

world simulator

generative world model

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Frame-level autoregressive (AR) design for real-time interactive long video generation
KV-recache mechanism to refresh cached states with new prompts for smooth and adherent prompt switches
Streaming long tuning to enable long video training and align training and inference (train-long-test-long)
Short window attention paired with a frame-level attention sink (frame sink) to preserve long-range consistency while enabling faster generation

Methodology

LongLive adopts a causal, frame-level autoregressive design that integrates a KV-recache mechanism, streaming long tuning, and short window attention with a frame-level attention sink. It fine-tunes a 1.3B-parameter short-clip model to minute-long generation in just 32 GPU-days.

Key Results

LongLive sustains 20.7 FPS on a single NVIDIA H100, achieves strong performance on VBench for both short and long videos, supports up to 240-second videos on a single H100 GPU, and further supports INT8-quantized inference with only marginal quality loss.

Resources

Paper Code