This is a DORA file. It may not be compatible with your inference software.
This DORA was trained directly on GurilaMash-v2. It is designed to work best with that model. It will likely not work well on other checkpoints.
Bikinis for GurilaMash-v2
Trained on 1k high resolution images of women wearing bikinis. It does not always produce specific swimwear on demand, but it has a large library to pull from and usually gets in the ballpark. The design and motivation behind this DORA is to produce the highest quality photorealistic images possible with SDXL. It can also be used for SFW images not involving swimwear and still improve lighting, contrast, and colors in photorealistic images, just don't prompt for swimwear. It does a really nice job of creating high-end backyard patios with amazing water views.
To help produce the best bikini images possible, I have included the caption files from the training. Copying them into the positive prompt produces good results. As with everything BigAsp and GurilaMash, prompts should be formatted:
score_7_up, photo \(medium\), photorealistic, realistic, bikini,
That will get you where you want to go.
I'm happy to provide help in the comments, but don't complain about it's poor compatibility with other checkpoints. I've tested it on other models, it's somewhat usable on other models merged with BigAsp, but you won't achieve the quality it's capable of with the model I trained it on.
Description
32/16 DORA file trained on 1k images trained using Onetrainer, prodigy, cosine, 3 epochs.
FAQ
Comments (4)
why i feel that ghurilamash v1 sometimes is still better ? (clipskip -2)
That is the deep question. The answer is, sometimes v1 is better. Something I did just (accidentally) made V1 really kind of awesome at some things. That came at a price, however. Most lora's do not work with v1. I made v2 specifically to try to provide lora support. Instead of relying on the model to do everything at the quality level I want, I'm training loras on v2, treating it like a platform, and with model + lora i get very good results and better flexibility. Face loras, concept loras, and style loras work much better, so that's why i released this, it's not the model we deserve, but it's the one we need right now.
Nice. Any chance of a microkini specific one, like for the super tiny, dorito sized wicked weasel bikinis?
It's of interest to me for sure. I was trying to do some microkini shots yesterday but this lora is very hit and miss with them. There's very few in the dataset i'm using. To be honest, i have a huge list of lora's i'm trying to get uploaded and more to train. I will likely get to it, but, it's looking like flux will take over before I get there. We'll see.