With a carefully selected noise schedule, diffusion models trained with as few as 32 latent states, or composed from single-state models, match 1,000-state training and converge 4-6x faster.
NVIDIA NVLink Interconnect
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Disentanglement in T-space for Faster and Distributed Training of Diffusion Models with Fewer Latent-states
With a carefully selected noise schedule, diffusion models trained with as few as 32 latent states, or composed from single-state models, match 1,000-state training and converge 4-6x faster.