A lora for people on all fours. Particularly suited for landscape images.
Recommended weight: 0.8
To assist lora, use prompts such as "side view", "on floor"
Description
A lora for people on all fours. Particularly suited for landscape images.
Recommended weight: 0.8
For longer prompts, assist lora with prompts such as "on floor", "side view" etc.
FAQ
Comments (5)
You're doing great work Barry. May I ask how many images you use in your datasets for these poses?
Thanks Gurila! For non pony models I find you need quite a few.. I would recommend at least 200+ for more complex poses likes these. This lora is still far from perfect and it used 200.
@JustBarry dammit, i was afraid you were gonna say that. I'm cooking a lora right now with about 600, and it's soooo slowwww. But i guess it's better than burning gpu time for a bad end-product. Thanks man. Happy baking.
@gurilagardnr I think poses tend to be one of the hardest things other than artstyle to mimic lora wise, you need a lot of different angles and people to get it reliable! Could I ask what the lora you're making is for?
@JustBarry Well, that's a deep question. I've got a dataset of about 10k images broken down by style and pose. I don't have the hardware to finetune SDXL, so my strategy right now is to do individual loras and see if I can't merge them into a model. I've got plenty of experience with 1.5, but as it takes so long to do anything on SDXL on my hardware, i'm always looking for ways to cut corners without cutting quality, and without having to endlessly retrain due to training parameter mishaps. I've got all the major poses one would expect from erotic and pornographic photography. My test run, because it's easy to train, is just women sitting. Legs spread,legs together, knees together, standard stuff.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.