Any reason for `use_adaln_lora = False` by default in `tdrussell/diffusion-pipe`?
#216
by sorryhyun - opened
Recently I've discovered that adaln could be also part of lora, which was even compatible with comfy. And actually it plays key role in the new turbo model I guess. But I'm wondering was there any reason for setting use_adaln_lora = False by default, in tdrussell/diffusion-pipe, the official source of anima lora training (at least that's how people think imo).
Doesn't diffusion-pipe also train the adaln modulation layers by default? I think only sd-scripts (and your anima lora trainer) don't train those layers by default.
Edit: Never mind, use_adaln_lora is a different thing. Though I think it's True by default https://github.com/tdrussell/diffusion-pipe/blob/a7e7decf4325c1f03e4b88b7de93640029abd011/models/cosmos_predict2.py#L125
sorryhyun changed discussion status to closed