GitHub - vita-epfl/Stable-Video-Infinity: [ICLR 26] Stable Video Infinity: Infinite-Length Video Generation with Error Recycling
2 months ago
- #AI
- #machine-learning
- #video-generation
- Stable Video Infinity (SVI) enables infinite-length video generation with high temporal consistency and controllable storylines.
- SVI-2.0 Pro is now available on the Poe platform, accessible via chat interface or API.
- Community contributions include tutorials and workflows for SVI, enhancing its usability and features.
- Key features of SVI include open-source availability, support for diverse generation tasks, and efficiency with LoRA adapters.
- Upcoming releases include Wan 2.2 Animate SVI, promising improved performance for infinite-length generation.
- SVI's design mimics a director's workflow, maintaining bidirectionality within clips for quality assurance.
- Installation and setup instructions are provided for different SVI models and workflows.
- Training and inference scripts are available for custom data, supporting various SVI applications like talking heads and dance animations.
- SVI's Error-Recycling Fine-Tuning encourages the model to correct its own errors, improving video quality over time.
- The project acknowledges contributions from related works and encourages citation of their research paper.