>>64195055i originally trained with clip 1, because fluffyrock/easyfluff uses 1.
for this version i simply loaded weights, changed base model to nai, and trained for a short time. it's probably too late to switch to clip skip 2 now.
>>64194892yeah, vpred/zsnr seems good, but this can still work without it.
i tried taking Easyfluff and finetuning it with no v-pred and zsnr - it starts outputting normal images pretty quickly.
picrel is ef+hll after finetuning without v-prediction for 2k steps. it seems to work and should get better with more training.
it could become an interesting model for merging
