Keywords:
From Animated to Realistic, woman, { property}, realistic transformation.Weight: 1.5 (please test), personally worked for me at 1.5
From Animated to Realistic Transformation
Description
FAQ
Comments (9)
Shit just got real
I see what you did there.
Really good lora, thank you. Is it possible for you to make a V2 without the chinese text appearing in the bottom ?
I think wan can probably do that (almost) natively. I'm getting 99.25% retention with a 27mb resize of this lora, each layer looks to be at most 2-3 rank.
Since a relatively high weight is required for the effect, might be undertrained.
interesting, I already train with 1.5k steps I think, what you suggest ? did adding more dataset helps ?
I haven't dabbled with training so my advice may not be sound, just trying to explain a phenomenon logically.
I know nothing of what kind of dataset you've used (type or quantity) so can't really say if that's the problem. Additional, more difficult to reproduce samples may challenge the model (for example a drastic transformation from like cartoon figure to a realistic cosplay). But wan's fairly capable with handling physical effects so maybe not. Perhaps it's a matter of more epochs, or the training settings.
Still, ultimately what matters in a lora is whether it gets the job done, is diverse, is stable when used in tandem with other loras.