Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion

Explore a comprehensive conference talk on Consistency Trajectory Models (CTM) and their application in improving score-based diffusion model sampling. Delve into the innovative approach that generalizes Consistency Models and score-based models, allowing for flexible traversal along the Probability Flow Ordinary Differential Equation in diffusion processes. Learn how CTM achieves state-of-the-art performance in single-step diffusion model sampling for image generation tasks. Discover the advantages of CTM, including its ability to combine adversarial training with denoising score matching loss, and its versatility in accommodating various diffusion model inference techniques. Gain insights into new sampling schemes, both deterministic and stochastic, that leverage CTM's capabilities. The talk covers background information, the CTM concept, performance improvements, multi-step generation, and concludes with a Q&A session.

- Intro + Background
- Consistency Trajectory Model
- Student Beats Teacher
- Multi-Step Generation
- Conclusion
- Q&A