Normalizing Trajectory Models
9 hours ago
- #generative models
- #image synthesis
- #normalizing flows
- Introduces Normalizing Trajectory Models (NTM), which model each reverse step as a conditional normalizing flow with exact likelihood training.
- NTM combines shallow invertible blocks within each step with a deep parallel predictor across the trajectory, forming an end-to-end network.
- It can be trained from scratch or initialized from pretrained flow-matching models.
- NTM's exact trajectory likelihood enables self-distillation, allowing a lightweight denoiser to produce high-quality samples in just four steps.
- On text-to-image benchmarks, NTM matches or outperforms strong image generation baselines in four steps while retaining exact likelihood over the generative trajectory.