All of them are frogs 🐸(it's one of the labels in the CIFAR10 dataset), because why not :), although maybe a horse could have been a nice choice too.
This time we have done things 4 ways, since we have CFG(Classifier free guidance) and EMA(Exponential Moving Average)
Note: In the CFG paper the authors mentioned heavy saturation when CFG levels were too high(i.e = 3), we are able to reproduce that same effect in the above picture.
Clearly the results with EMA are much better generally, however CFG helps with more stable looking images :)
- Landscapes dataset (for unconditional model)
- CIFAR-10 Resized (for conditional model)