yadaさんのイラストまとめ

Progress / experiences so far on #hidream lokr trainings, before I forget:

vram (bf16, 1024px):
- 24G for square images single aspect ratio bs1 on dev (not sure)
- 32/48G for full, around bs1
- 80G for aspect buckets + bs8

training:
- use SimpleTuner
- use EMA (doesnt take

Progress / experiences so far on #hidream lokr trainings, before I forget:

vram (bf16, i1-full, 1024):
- 24G/48G for square images single aspect ratio bs1
- 80G for aspect buckets + bs8

training:
- use SimpleTuner
- use EMA (doesnt take vram, and it really helps)
- lion 3e-5 if

Progress / experiences so far on #hidream lokr trainings, before I forget:

vram (bf16, i1-full, 1024):
- 24G/48G for square images single aspect ratio bs1
- 80G for aspect buckets + bs8

training:
- use SimpleTuner
- use EMA (doesnt take vram, and it really helps)
- lion 3e-5 if

works really well on other v-prediction models as well! Not even illustrious-based.
(p2 for comparison) https://t.co/SAGaus1vAw https://t.co/MXKHDVyqPE

works really well on other based v-prediction models as well! Not even illustrious-based.
(p2 for comparison) https://t.co/tA5jzuOZBf https://t.co/MogujiZOsE

So...yes, #HunyuanVideo can be used as an 11B text to image generator like an undistilled Flux, but with lower base resolution and extra t compression artifact🤔
#AI #AIイラスト

So...yes, #HunyuanVideo can be used as a 11B text to image generator like an undistilled Flux, but with lower base resolution and extra t compression artifact🤔
#AI #AIイラスト

So...yes, #HunyuanVideo can be used as an 11B text to image generator like an undistilled Flux, but with lower base resolution and extra t compression artifact🤔
#AI #AIイラスト

simple