A concept depicting women with extreme muscles.
XL version: There is a trigger word but the LoRA seems to be working even without it. Sample prompts along the lines of "RAW portrait photo of extrmusc woman presenting muscles". If it feels too strong at full weight, try to lower it to more reasonable levels.
Flux version: Trained with the same dataset, it gives good result at full weight but you can go higher for more extremes.
Description
FAQ
Comments (14)
Huh. Would you consider making a less extreme version? Something along the lines of competitive athletes, seems like that might have more use in general. Still buff but not cloud levels xD
If you post a bounty, ideally also including images you think are the best suited, I can try that. This is not my niche thing anyway, just tried to follow the bounty request :)
It's too strong, changes the face and style of generation a lot and ads artifacts.
Even in your generations you can see the artifacts messing the style and face.
Guess that should be easy to compensate with lowering the LoRA weight though. I admit I didn't test that but no reason why it would not work if full weight is too strong. If you can, try that and let me know if it helps - I can add a note about it to description anyway and also look if there is no better epoch to use. The face problem is probably related to not using ADetailer at all....
@galloguille Yeah I tested it a bit now and you are actually right, guess my eyes are not serving well... But lowering the weight to about 0.5-0.7 seems to be giving much better results. Thanks for the feedback, appreciated :)
@dolirama126 I tried lowering the weight but then it doesn't do much. But hope there will be more lora's like this.
Works like a charm, I merged it into a checkpoint on a whim and did some further merging of it, and they came out excellently. Strengths were all at about 0.3-0.4 for the LoRA (after merging).
Thanks for the feedback, really helps with adjusting things so they work for everyone in the end :)
I think all your input data comes frome one photoshoot, it adds unnecessary pixel noise and other artifacts, even the same face.
No, the photos were from diverse sources and the same face is there simply because I only used "woman" in the prompt for sample images. Pixel noise and artifacts are there indeed, but I attribute it to unnecessary resizing of smaller images - should have probably left it to bucketing to deal with. But it is what it is - thanks for the feedback :)
@dolirama126 Well, if you use bucketing, then the data from smaller images doesn't go to the main bucket. That's why I try to use at least 1024x1024 input images for my own XL training. About the same face - I get my generated faces morphed into the specific one which comes from your LoRa. Either I need to use smaller weight than 1.0 with the LoRa or you could get more randomized input, or perhaps it would help if you add more captioning/tagging for the facial details? Whatever the training tool doesn't know or understand, it goes into the concept. Good luck!
how much image/dataset you're using for this lora? it's pretty strong even with 0.5 and keep giving me same face
Hmm, I don't remember that exactly but it was slightly less than 50 images. I'm not sure about that "same face" thing - if you look at the gallery below the model, there is pretty nice variety of faces already, some even clearly at much higher strengths than 0.5.