VPred models are trained with V-Parameterization so they will only work on VPred models like NOOB-AI-XL-vpred or Obsession, etc.
This is a DoRA so make sure to have your stuff somewhat updated
Remember to check out the artist!!! Without their art, these LoRAs wouldn't exist!!!
Version history:
NAI-XL EPS v1
Same dataset as before but different training settings
Overall, it's good but can be too strong at times so try lowering weight to 0.8 or maybe even less if it makes weird artifacts or just bakes the image
NAI-XL VPred v1
Same dataset as before with slightly more tags than before and trained for VPred Illustrious-XL models
v2.0
Now it's a DoRA trained on a new dataset
Seems to be better overall but can't use the specific prompts at the end of description (those are for v0.1)
v1.0
Was better than v0.1 but not good enough to upload
v0.1
Dumped 1000buzz with a half-finished dataset just to see how well PonyXL works (it was my first PonyXL LoRA attempt)
Worked surprisingly well for the bad dataset, but changed the seeds too much and was inconsistent.
This style:
Large asses, wide hips, futas packing serious meat
For Pony versions:
Weights:
Overall, you can leave it at 1.0 but sometimes you have to change it (have had times where it would only work only below 0.6 and above 1.4)
This LoRA likes to fail if you give high weights for some tags that other style LoRAs don't fail on. It generally starts failing once a tag is over the weight of around 1.3–1.4
Prompting (for v0.1):
I recommend starting with: score_9, score_8_up, score_7_up, source_anime
There is no activation tag
Other tags:
Wide hips
The LoRA by itself creates wide hips so in most cases you won't need the tag
Penis sizes
Size can be controlled by large penis and huge penis
Simply adding penis in most cases will give it a random size
There should be decent results with flaccid and erection
Same thing for testicles
Horse penises
Use equine penis
Urethral
Since the artist also has plenty of these kinds of artworks, I tried to add it here as well (for now works sometimes)
Urethral insertion
use this tag for object insertion like tentacles (at least it kind of worked with this tag)
Urethral penetration
Use this for penis in penis action. I actually haven't tested if this even works
You can also try adding x-ray along this tag (might work, haven't tested)
pov also might work with this tag
Urethra tag can generate a gaping urethra in some cases
Jadf-tan
It is possible to create this character if you use:
futanari, small (or medium) breasts, black hair, blunt bangs
It kind of figures out the rest usually
Description
Just an unfinished version that performed better than expected, so I decided to upload it.
FAQ
Comments (4)
Hey, first of all thanks for taking the time to do the LoRA, I hope you can finish it. On the other hand, I wanted to ask you if, when tagging in CivitAi, you put tags like "score_9", "score_8_up", etc. Since I would like to try to train some LoRA but I don't know the correct way to tagging to train in Pony. If you could give me some advice I would appreciate it.
All tags (even the auto-tagged ones) started with "score 9, score 8 up, score 7 up, score 6 up, source anime, rating questionable (or explicit)". The tagging software, when I applied these tags in front of all, removed the underscores. I don't know if that matters. I'm playing around myself with PonyXL since there is very mixed information online... This was literally the first try with random parameters on the website trainer with an unfinished dataset. As much as I hate to say it, both me and you will just have to try multiple times and see what settings and tags work better... I'm still only used to SD1.5 LoRA training. After I finish this LoRA (might take until the middle of summer though) I can update you what worked and what didn't... might create a spreadsheet or something...
@SkagenM Thanks for replying, I get the point. I will try to train a LoRA and see what happens, I have the dataset so I would just have to tag it. Although, for my part, the CivitAi auto-tagger, when tagging does not start with "score 9, score 8 up, etc..." in fact, it does not even place it in the prompt. I will try to do it manually.
If I manage to make progress or find a configuration that gives me good results I will let you know.
@moonless_ For tagging I like to use sd-tagtool because there is an option to modify tags for all images at once but it deletes all underscores (though with notepad++ you can add them back). For automatic tagging I use an extension in A1111 that can process a whole folder at once, then I just manually go through them and delete wrong tags and add missing ones... The auto-tagger on this website doesn't add them (I don't think so at least...). Though at the same time I'm not sure if you even need them to begin with. For this LoRA I tried to remove score_9 from sketch images and just added the tag "sketch" so it kind of understands the more scuffed look... I don't know if that worked. For now, can't do quick training on the website since I spent all the buzz I had (around 1000 for 11 epochs with 483 images).
If you want to update me later on your progress, you can just DM directly through the website's DM system. Good luck with your training :))



