We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Sep 30·3 min read
Sora 2 System Card
TL;DR
Prior video generation models struggled with physics accuracy, visual sharpness, audio synchronization, and creative control. OpenAI built Sora 2, a unified video and audio generation model that improves physics simulation, increases visual fidelity, synchronizes generated audio with video, and expands stylistic range.
✦ Why It Matters
Engineers can now generate synchronized video-audio content with better physics and creative control for applications requiring realistic, steerable media generation.
Key Takeaways
- Sora 2 unifies video and audio generation, enabling synchronized multimodal output.
- Physics simulation improvements allow more realistic object motion and interactions.
- Enhanced steerability gives users finer control over generation parameters and style.
- Expanded stylistic range supports diverse creative outputs beyond prior model capabilities.
Full Summary
Video generation models have historically faced challenges in producing physically plausible motion, maintaining sharp visual detail, synchronizing audio with video content, and offering fine-grained creative control over output style. Sora 2 is OpenAI's next-generation generative model designed to address these limitations through architectural and training improvements.
The model combines video and audio generation into a single system, enabling tighter synchronization between modalities. Key improvements include more accurate physics simulation (objects behave according to real-world rules), enhanced visual sharpness and realism, synchronized audio output that matches video timing and content, improved steerability (user control over generation parameters), and expanded stylistic range allowing diverse creative outputs.
While specific quantitative metrics are not detailed in the announcement, the system represents a significant step forward in multimodal generative capabilities.