← Back
OpenAI
OpenAI Upgrades Sora 2 with Higher Resolution, Longer Videos, and Character References
OpenAI API · apifeaturemodelsdk · developers.openai.com ↗

Key Capability Updates

OpenAI has released an updated Sora 2 prompting guide documenting several significant improvements to the video generation API:

  • Character References: Upload a character once and reuse it across multiple videos with consistent appearance and identity
  • Higher Resolution: Generate videos in up to 1920×1080 or 1080×1920 resolution (available on sora-2-pro)
  • Extended Duration: Maximum video length increased from 12 seconds to 20 seconds
  • Video Extension: Extend existing videos using the full initial clip as context, not just the final frame
  • Batch API Support: Asynchronous video generation jobs for larger production workflows

API Parameters and Configuration

The guide clarifies how to use explicit API parameters to control video output:

  • model: Choose between sora-2 or sora-2-pro
  • size: Supported resolutions vary by model (sora-2 supports 720×1280 and 1280×720; sora-2-pro adds 1024×1792, 1792×1024, 1080×1920, and 1920×1080)
  • seconds: Set duration explicitly (4, 8, 12, 16, or 20 seconds)
  • characters: Reference up to two uploaded characters in a single generation

Prompting Best Practices

The guide emphasizes that successful prompts treat the model like a cinematographer receiving a storyboard brief. Key recommendations include:

  • Balance specificity with creative freedom: Detailed prompts provide control and consistency, while lighter prompts allow the model surprising creative variations
  • Describe shots cinematically: Include camera framing, depth of field, action beats, lighting, and color palette
  • Use multiple iterations: The same prompt generates different results—this enables exploration and refinement
  • Keep sequences organized: When describing multiple shots, maintain distinct camera setups, subject actions, and lighting for each

The guide includes concrete examples and emphasizes that prompting is iterative—small adjustments to camera, lighting, or action dramatically shift outcomes. Developers should expect to collaborate with the model rather than expecting deterministic results.