Hello! This LORA is responsible for making a certain position quite reliably.
Since we all know that hands and feet are not really SDs forte, the training data was set up to alleviate that problem.
As a result - you should be getting nice results pretty much every time.
The trigger is the keyphrase "pstng position", for example: "woman in pstng position" is enough to generate some nice content.
This LORA works on various models. Hovewer I have not tested yet on anime ones, but seeing how flexible it is (skin color, etc) I see no reason why it would not work.
There are two LORAs, with lower and higher precision (differs in size), but the training was all the same.
PS. I have a DB model of this as well, if someone is interested - just let me know and I can make it available if needed!
Description
FAQ
Comments (13)
By the way, you should take a peek at Controlnet.
Yo bro, can u explain a bit how you train a lora for 1 specific position?
Yo, sure!
I have attached the traning data so you can check what was used exactly.
Then I trained a dreambooth using this class prompt "photo of a woman" and this class instance "photo of woman in pstng position".
Once I got a checkpoint I extracted LORA using kohya_ss LORA extractor.
PS. In the training data I made sure to select images where the hands themselves are not visible yet the composition is still good (no cutouts) and this training was successful as well.
You are doing humanity a great service! Thank you. Works really well - especially combined with ControlNet (Depth, Canny or NormalMap).
Why is this file a 'pt' file yet its labelled here as a LoRA?
what do you mean? it is a LORA file, what file extension did you expect? :)
@danGorst Because .safetensors is now the standard for both regular models and LORA
People would be much less reticent to download a .safetensors file
@syl sure, i can upload the safetensors version, i think kohya can convert it
However he said that this wasn't LORA and I was confused
Cant load this lora :(. I got this error: Bad Lora layer name: lora_unet_down_blocks ...
The webui Easy Diffusion cannot used this LORA, can you upload a safetensors version?