Goku

lastmod 2025-02-14created 2025-02-141 edits

Saiyan-World/goku: Video Generation Foundation Models: https://saiyan-world.github.io/goku/

ByteDance's flow-based video generation model

About

• Goku is a new family of image and video generation models based on the "Rectified Flow Transformer", aiming for industry-leading performance
• Applies various optimizations such as "data curation, model design, and flow formulation" for high-quality visual generation
• Key features
  • Precise image and video generation through high-quality data curation
  • Uses the Rectified Flow technique to strengthen interaction between image and video tokens
  • Delivers excellent performance in image and video generation
• Supported generation tasks
  • Text → video generation
  • Image → video generation
  • Text → image generation
• Performance benchmarks: high scores on major benchmarks
  • 0.76 (GenEval - text → image generation)
  • 83.65 (DPG-Bench - text → image generation)
  • 84.85 (VBench - text → video generation)
• Goku-T2V scored 84.85 in the VBench performance comparison, ranking 2nd as of October 7, 2024
  • Outperforms several commercial text-to-video models (AnimateDiff-V2, OpenSora, Gen-3, Kling, etc.)

See also

• Deep learning
• GAN
• Gen-2

Favorite site

• Github - Saiyan-World/goku