OneTrainer now supports DoRA and layer presets (like in Lycoris)! They also added new optimizers and other stuff
Picrel - I trained Lora for 330 steps, used sigmoid timestep distribution, noising -10.0, bias 0.25.
New TIGER optimizer was 20-30% faster than prodigy
>>82111050Yeah, current version is pretty bad.
Maybe i'll make a new one in september. Though i don't think it's necessary. All holomem should have loras already, and models like holodayo already know all popular vtubers
