( . Y . Y . )
Description
FAQ
Comments (13)
three boobs.. nice.
now quad boob, multiple penis
希望出双排但多排的巨乳
Aside from the LoRA itself, those are very good prompts, thank you!
Thanks for work. Your best.
Your training results are really good. Can I ask about your process? Number of clips or images/sizes/steps etc.
trained on diffusion-pipe (last time we tried ai-toolkit it had terrible results)
rank 32, lr 1e-4, ~1500 steps (we actually trained more, ~2000 steps, but 1500 seemed to be the best epoch that captured to concept without overtraining. It's important to compare and test your epochs!)
~140 images at 512px, 2/3 of which clothed and 1/3d nude, and showing a variety of huge/large/medium/small breast sizes on a variety of ethnicities/skin tones
and then another 25 images at 1024px of nipple close-ups just because Minimax needs some help on that front (each concept/model is unique and may need different adjustments)
no videos (because minimax already understands what breasts are, it just needs to learn they come in three)
@2_as_1 That's neat. Thanks for sharing. It seems like Minimax is friendlier to images than Wan and doesn't seem to "lock-up" when you train on them. Interesting.
@boobkake22 it's still going to be dependent on the concept you're trying to train. A genitalia lora is likely going to need a much bigger and more varied dataset including videos to fully train the concept, since minimax doesn't have a great out-of-the-box understanding of it.
@2_as_1 sorry to barge in like that, but the lora size actually more resembles rank 16? Rank 32 should be around 600 MB. Anyways, fantastic results!
Edit: maybe I'm mixing things up, I should really sleep more 😅
@cupra Interesting ... it's definitely rank 32.
Some trainers (not diffusion-pipe) use AdaLN (whatever that is) which increases file size. Or at least that's what Claude tells us... so take it with a grain of salt🤔🤔
@2_as_1 ah, that makes sense. Thanks for the clarification!
@2_as_1 AdaLN is something H3 uses itself for modality conditioning. Some trainers target these extra layers to fully train in motion and audio, but guess if you're training on images it doesn't matter at all.