Normalizing Trajectory Models
A few-step diffusion generative model framework that models each step of the reverse process as a conditional normalizing flow. The approach maintains exact likelihood computation while enabling high-quality generation in few steps, combining shallow invertible blocks with deep parallel network structures.
Timeline (last 90 days)
Normalizing Trajectory Models (NTM) model each step in the reverse process as an expressive conditional normalizing flow, rather than Gaussian denoising
NTM can be optimized under exact likelihood training, preserving the exact likelihood computation property that existing few-step methods abandon
NTM adopts a hierarchical strategy in its architecture: each step internally uses shallow invertible blocks, combined with a deep parallel structure
NTM circumvents the scalability bottleneck of normalizing flows in global high-dimensional modeling by restricting the flow to be shallow and operating only within individual diffusion steps
NTM is still an early-stage research result, and whether it can achieve sample quality comparable to mainstream diffusion models on high-dimensional generation tasks such as image generation remains to be verified experimentally
All Facts (5)
Normalizing Trajectory Models (NTM) model each step in the reverse process as an expressive conditional normalizing flow, rather than Gaussian denoising
50%UnverifiedNTM can be optimized under exact likelihood training, preserving the exact likelihood computation property that existing few-step methods abandon
50%UnverifiedNTM adopts a hierarchical strategy in its architecture: each step internally uses shallow invertible blocks, combined with a deep parallel structure
50%UnverifiedNTM circumvents the scalability bottleneck of normalizing flows in global high-dimensional modeling by restricting the flow to be shallow and operating only within individual diffusion steps
50%UnverifiedNTM is still an early-stage research result, and whether it can achieve sample quality comparable to mainstream diffusion models on high-dimensional generation tasks such as image generation remains to be verified experimentally
50%