Loading post…
AppleProblem & solution
Normalizing Trajectory Models
Summary
Diffusion models decompose sampling into many small Gaussian denoising steps, an assumption that breaks down when generation is compressed to a few coarse transitions. Existing few-step methods address this through distillation, consistency training, or adversarial objectives, but sacrifice the likelihood framework. Normalizing Trajectory Models (NTM) models each reverse step as an expressive conditional normalizing flow with exact likelihood training. Architecturally, NTM combines shallow invertible blocks within each step with a deep parallel predictor across the trajectory, forming an end-to-end network trainable from scratch or initializable from pretrained flow-matching models. On text-to-image benchmarks, NTM matches or outperforms strong image generation baselines in just four sampling steps while uniquely retaining exact likelihood over the generative trajectory.
Context
Diffusion models decompose sampling into many small Gaussian denoising steps, which breaks down when generation is compressed to a few coarse transitions. Existing few-step methods use distillation, consistency training, or adversarial objectives but sacrifice the likelihood framework.
Approach / What changed
NTM models each reverse diffusion step as a conditional normalizing flow with exact likelihood training. It combines shallow invertible blocks per step with a deep parallel predictor across the trajectory, forming an end-to-end network trainable from scratch or initializable from pretrained flow-matching models. Self-distillation uses a lightweight denoiser trained on the model’s own score function to produce high-quality samples in four steps.
Takeaways
NTM models each reverse diffusion step as a conditional normalizing flow, enabling exact likelihood training unlike distillation or adversarial methods.
The architecture combines shallow invertible blocks per step with a deep parallel predictor across the trajectory, forming an end-to-end network.
Self-distillation trains a lightweight denoiser on the model’s own score function, producing high-quality samples in four steps.
FAQ
How does NTM model each reverse step?
NTM models each reverse step as a conditional normalizing flow with exact likelihood training, combining shallow invertible blocks per step with a deep parallel predictor across the trajectory.
How does self-distillation work in NTM?
Self-distillation trains a lightweight denoiser on the score function induced by the model itself, enabling high-quality samples in just four steps.
How does NTM compare to other few-step methods?
NTM matches or outperforms strong image generation baselines in just four sampling steps while uniquely retaining exact likelihood over the generative trajectory.
Source and freshness
- Source
- machinelearning.apple.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.