Dogeza is gacha as fuck, so I baked one for a consistent facedown dogeza pose where you have to worry less about body horror. Not intended for face-up usage, although you might be able to finagle it.
Dataset cleaned of trembling and speech_bubble, so it's unlikely you'll have to neg those.
Use:
dogezaLarge breasts are difficult to prompt for unfortunately.
Dataset has strong bias towards nudity. Will contain trace amounts of random objects strewn across the floor as those aren't tagged properly/cleaned.
Description
FAQ
Comments (4)
Very nice! ❤️
Hey,
I really like your model! :D
I'm trying to train a dogeza model as well, because I like that pose and for learning purposes. But im struggling. The bodies generated are always weirdly deformed and its struggling with faces, because it tries to draw them normally 'looking at viewer' style.
Do you have some tips how to train complex poses like this?
How do you select your images for the dataset and how many did you gather?
How did you tag them for training and what settings did you use?
I would be glad if you could share some knowledge :D
https://files.catbox.moe/x58jxc.toml for the settings I use with EasyScripts.
Other than that, it's just consistent composition and angles when selecting images. I used around 65 images in the dataset. I strongly prefer having more, at least in the 100-150 range, but there weren't any others that fit the quality criteria. I cleaned sfx and dialogue using LamaCleaner.
For tagging, any of the WD taggers are good. While good tagging and tag consolidation does make a difference, I'm usually too lazy to do anything of the sort.
At best I'll do a half-hearted attempt at removing some redundant tags. For example, if you're doing a bikini lora where the design is consistent, you can remove tags like striped bikini/blue bikini/microbikini and just have a single bikini tag because they're all describing the same thing.
Dataset quality and consistency has the largest impact on the outcome of the lora, no amount of tagging will save a bad dataset with inconsistent poses and angles.
Thanks for your reply!
What's the purpose of setting scale_weight_norms to 5?
And what is betas = "0.9,0.99" and gamma = 0.85 doing?
I'm training a Locon at the moment. I have like 50 images for different angles e.g. from side, from behind, diagonal and 'normal'. They are seperated into different folders.
Also I did not use reps. I used more epochs. My thought was, that it shouldn't learn the specifics of the images, just the general pose. And I bumped the learning rate a little to 0.0002 so it ignores the details.
I'm training for 9k steps with a dim of 128, 128 alpha and conv 16, alpha 16.
Maybe it's going to be better this time xD