【Demos】LongLive compared with Sora2 in Interactive Long Video Generation

Right after Sora2 was released, we ran an interesting comparison with LongLive, our open-source model released just two days earlier, focusing on long video generation. Since Sora2 can only generate 10s clips, we used GPT-5 for prompt engineering, enriching the inputs with background & context to improve coherence. This experiment is essentially “a powerful closed-source video model + GPT-5 prompt engineering” vs. “a 1.3B open-source model designed for long videos.” 👉 Note: this comparison video is shown at 2× speed. 🎥 Sora2 shines in visual fidelity, camera motion, and physics simulation—far beyond LongLive-1.3B. We assume Sora2 has much larger scale & parameters, and we look forward to its technical report. ⚠️ With Sora2’s length limit, even with GPT-5 assistance, abrupt inconsistencies still occur between clips (e.g., in Video 2, Paul’s hair & mountain background suddenly change). ⚡ Meanwhile, LongLive-1.3B is faster. Sora2 takes ~3 minutes to generate 10s, while LongLive-1.3B runs at 20.7 FPS real-time on a single H100. ℹ️ LongLive-1.3B was fine-tuned with our long-video extension strategy on top of Wan-2.1-1.3B (a strong short-video model), achieving real-time ultra-long video generation. LongLive is fully open-sourced, including training & inference code and model weights. 📦 Code: https://github.com/NVlabs/LongLive 📄 Paper: https://arxiv.org/abs/2509.22622 🧩 Model: https://huggingface.co/Efficient-Large-Model/LongLive-1.3B 🌐 Demo: https://nvlabs.github.io/LongLive