CogOmniControl

Model Overview

CogOmniControl is a undisclosed-parameter model developed by Academic/Research. Released on 2026-05-19.

---

📊 Quick Specs

---

✨ Key Features

• Reasoning-Driven Controllable Video Generation: Decouples creative intent understanding from video synthesis
• CogVLM Intent Cognition: Specialized VLM trained on authentic anime production data to turn sparse sketches into dense reasoning
• CogOmniDiT Generation: Unified Diffusion Transformer supporting multi-condition in-context video synthesis
• Reinforcement Learning Alignment: Aligns DiT generation paths with VLM reasoning outputs using RL
• Closed-Loop Best-of-N Harness: Uses planned evaluators to select optimal generated video candidates
• Professional Production Workflow Focus: Designed for storyboard sketches, clay renders, and complex multi-modal controls

---

🔗 Resources

• Hugging Face: CogOmniControl
• GitHub: Repository
• Paper: arXiv
• Website: Project Page

---

📜 License & Access

Other/Custom — See repository for specific license details.