A unified framework and targeted changes to parameterization, architecture, and objectives stabilize continuous-time consistency models, achieving FID scores of 2.06 on CIFAR-10 and 1.88 on ImageNet 512x512 with two-step sampling at 1.5B scale.
For diffusion models, we generally need to focus on the intermediate time steps since both the clean data and pure noise cannot provide precise training signals
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models
A unified framework and targeted changes to parameterization, architecture, and objectives stabilize continuous-time consistency models, achieving FID scores of 2.06 on CIFAR-10 and 1.88 on ImageNet 512x512 with two-step sampling at 1.5B scale.