As many people pointed out before me, trained Loras are super weak in style, regardless of the steps, learning rate, or number of songs in the dataset. You can overtrain a Lora with 10k epochs, and it will be like putting a normal Lora at 0.3. Do we need some kind of special settings or a different workflow than the one used by RyanOntheInside?
As many people pointed out before me, trained Loras are super weak in style, regardless of the steps, learning rate, or number of songs in the dataset. You can overtrain a Lora with 10k epochs, and it will be like putting a normal Lora at 0.3. Do we need some kind of special settings or a different workflow than the one used by RyanOntheInside?