NUS-HPC-AI-Lab / SpeeD

SpeeD: A Closer Look at Time Steps is Worthy of Triple Speed-Up for Diffusion Model Training
Apache License 2.0
162 stars 5 forks source link

Some questions on cifar-10 #6

Open DocDode opened 2 months ago

DocDode commented 2 months ago

How do you evaluate the cifar-10 dataset ? & How many samples do you use for evaluating cifar-10

BDeMo commented 1 month ago

Cifar-10 is used to evaluate the conditional generation, as shown in appendix A.4 in our paper. We follows the EDM eval. setting, which use 50K generated image and all available real images to compute FID. Plz refer to EDM's repo for more details.