Changelog
Track our progress as we democratize AI video generation. Every release brings us closer to matching Sora's capabilities.
v2.0.0
majorReleased on December 15, 2024
🎯Highlights
- •Reduced performance gap with OpenAI Sora from 4.52% to 0.69%
- •Support for 768px resolution
- •11B parameter model with $200K training cost
- •Both text-to-video and image-to-video in one model
✨Added
- →Support for 256px and 768px resolution video generation
- →Unified model for both T2V and I2V tasks
- →Advanced temporal consistency across frames
- →HuggingFace model hub integration
- →Google Colab quick start notebook
⚡Improved
- →Training efficiency improved by 46% compared to Sora
- →VBench score increased to 82.0% (vs Sora's 82.7%)
- →Better language understanding for complex prompts
- →Enhanced physics simulation in generated videos
- →Optimized memory usage for consumer GPUs
🐛Fixed
- →Temporal consistency issues in long videos
- →Character morphing in multi-subject scenes
- →Motion blur artifacts in fast-moving scenes
v1.2.0
minorReleased on September 20, 2024
🎯Highlights
- •VBench score: 78.2%
- •Improved training pipeline
- •Community contributions integrated
✨Added
- →Support for custom training datasets
- →Prompt templates system
- →Video quality assessment tools
⚡Improved
- →Faster inference speed (20% improvement)
- →Better handling of complex scenes
- →Reduced GPU memory requirements
🐛Fixed
- →Installation issues on Windows WSL
- →Compatibility with older CUDA versions
- →Memory leaks in long-running sessions
v1.0.0
majorReleased on June 1, 2024
🎯Highlights
- •Initial public release
- •Open-source Sora replication
- •Community-driven development model
✨Added
- →Text-to-video generation (256px)
- →DiT (Diffusion Transformer) architecture
- →Apache 2.0 license for commercial use
- →Comprehensive documentation
- →Example prompts and demos
Stay Updated
Follow our GitHub repository to get notified about new releases, features, and improvements.