Carter Cruise
⭐If you like these models and want more, please rate and comment!
⭐After a new release gets 10+ thumbs up, more will be posted.
⭐Special requests available at [email protected]
Description
FAQ
Comments (5)
I realy appreciate your effort it seems like you are basicly trying to do every Adult film star. Which takes alot of time and effort so thank you for putting that energy into it =) I have one little bit of feedback for you if I may. Whilst I applaud your effort and realy like what you are doing I do think this a an example of quantity over quality, some of them are great, but I do notice the more and faster you roll out these lora's they are starting to degrade a bit in quality, for this lora it does look like Carter Cruise but if you spend a little bit more time on one lora instead of focusing on uploading one after the other I think they will improve greatly. Like I said I realy like what you are doing so don't take this the negative way and definetly keep it up =)
I've been using the same settings and techniques for the training prompts. I think the largest differences with some might be due to the source images more than anything else. As I find better datasets I'll keep refining the training of ones that don't match as well.
@SurferKate What settings do you use? and what learning method Cosine or constant? might be able to give you some tips I have been experimenting a lot lately with different settings. Must say though I have not found the one solution perfect settings that work for each model. seems each data set requires a different approach what also is a big thing to take into account is the person I am training already known to SDXL if the answer is yes it will be easier. if you need help send me a pm on discord which is the same name as my user name here.
@MakeThemComeAliveAIArt Prodigy, learning rate 1, cosine. usually image sets of ~15-40. try to use good quality without any obstructions to face, no hats, etc. general variety of poses/perspectives
then I'll run it for ~100 epochs in batches of 1-4, whatever comes out to ~1,500-2,000 steps total.
all against base xl model, using Blip2 for auto captioning to start, then replacing all "a woman" with "[namehere]" and correcting any mistakes, adding more detail for background, shot, pose, etc.
Would love to know better settings to test with. These are the first I've found that are consistent enough to use each time.
@SurferKate For known people cosine is good, for unknown my personal findings are that constant is better and a bit higher learn rate, 3 or 4. but at constant if you find the image gets over exposed or deformed you might want to lower learn rate a bit or image repeats