https://www.jasonfletcher.info/vjloops/index.html#internal-r...
I remember wanting to do a Pokemon finetune (before such things were commonplace) and the dataset cleanup was the worst part. Stability issues and lack of decent multi-GPU setup being a close second.
StyleGAN2/3 is quite sensitive to fine-tuning models with a dataset that contains lots of variation, such as the Pokemon dataset. But it will converge much better with a dataset of just Pikachu images in tons of different poses. It seemingly wants to find the most common pattern in the dataset and interpolate within that space. From there nailing down the ideal value for the gamma attribute is difficult, but luckily can be reduced every 1000 to 2000kimg until the model no longer trains favorably. Currently my typical gamma strategy is: 80 > 10 > 5 > 2