BUFF

This is a model that aims to generate a particular kind of muscular build for women: very large muscles and wide hips/waist without many veins or striations. If you want less striations, try putting flexing and veins in the negative prompt.
I recommend using AutismMix or WAI-ANI-NSFW if you use Pony models or Hassaku or Xolia if you use illustrious. These models play well with the anatomy that this LoRA aims for.

V2.2 Pony

See the V2.2 Illustrious for tagging and generation changes.
I found that Pony models require more work with tagging to get the desired images, as well as a higher LoRA weight. I recommend a default weight for Pony models of 1.4. This allows tags to make a meaningful difference without blowing out the style of the model you're using.

V2.2 Illustrious

Trained through Civit instead of Google Colab. I found that Illustrious models work fantastically with tagged datasets and don't require nearly as high of LoRA or tag weighting to get the desired result.
Muscles appear less striated, helper tags such as huge pectorals, huge thighs, and thick forearms, have greater affect, and the sizes are finally represented how they should be. Even flexing and veins tags don't completely ruin the anatomy of the muscles now. Finally, I removed thick arms as a helper tag. It seemed to reduce the affect of thick forearms and it didn't seem necessary in this version.
Enormous and gigantic sizes are larger, especially in the case of enormous. You can start to get into hyper sizes if you push the weighting of the LoRA and tag up now.
This should be the last update on the LoRA for a while. It's finally at a satisfactory spot and it plays well with other LoRAs. I'll only update to add more helper tags or if another base model proves to be even more beneficial.

V2.1
Managed to better define some more niche tags I've been working towards. The huge pectorals tag generates better, the size of gigantic muscles is closer to what I intended it to be, and the enormous muscles tag is larger, but less consistent. I also found that hands seem to be backwards when trying to generate flexing poses from behind. (It was the checkpoint that was causing this, not the LoRA). These will be remedied in the next update.
Speaking of flexing, there are a couple of new tags. single bicep pose, double biceps pose, huge biceps, thick arms, and narrow waist are all in the dataset. The poses work rather well expect for the aforementioned hand-flipping issue. Huge biceps and thick arms are more helper tags to increase the size of the arm relative to everything else, but not necessarily the forearm. Narrow waist is exactly what it sounds like and allows for more control over the body type.
Finally, I've noticed that more muscle strands were visible on average in this version, especially when generating from behind. I'll be reviewing the dataset before the next update and possibly adding another helper tag to isolate images that have very striated muscles so that can be tagged out or in.

V2.0

This version mostly completes this LoRA. There's still a few kinks to work out, like some tags (huge pectorals) needing too high of a weight to get to generate properly or some position still not generating correctly. Other than that, everything else has improved. Size tags are more well defined, tags produce better results, and the overall "quality" of the muscle generated seems to be better.

The only thing that may be odd at first is how this version responds to tag weights. Increasing the weight of a tag on this version actually produces the desired result without too much unwanted generations. Put simply, if you aren't getting what you wanted out of the LoRA, you can now play around with the tag weighting without those tags strong-arming the prompt too quickly.

V1.2

Huge pectorals tag has a more desired effect and the thick arms tag has been replaced by thick forearms. The enormous muscles tag also generates better, and from behind shots and back muscles seem better as well.

V1.1

Added size tags to the LoRA as activation tags rather than the keyword buff. They aren't perfect just yet, but they are consistent. I plan on better defining each size in the future as I find images to train on. I also fixed the thick arms tags and added pectoral tags.
Previously, thick arms would focus on the size of the arm, not the forearm. I know this is what the tag usually focuses on, but I intended for it to focus on increasing forearm size.
The pectorals tags are large pectorals, huge pectorals, and flexing pectorals. These are mostly helper tags right now, but I plan to make them fully functional by the next update.

V1.0

Managed to get this LoRA to act pretty stable. All instances of unwanted monochrome aspects or sleeves randomly appearing are gone. Pretty much all perspectives and framings seem to generate consistently now.

V0.3

Expanded and then pruned the dataset. This LoRA responds way better to SDXL than the original did on SD1.X. Seems to play with other LoRAs well and handles different angles and positions much better than the previous version. I also haven't had any monochrome problems with this LoRA.

V0.2

Added more photos to the dataset and pruned off some photos that didn't quite match the concept. This is mostly just an improvement on the original version. I have discovered that this model is obsessed with generating detached sleeves for some reason. If you encounter what looks like sleeves when there shouldn't be any, try adding monochrome and sleeves to the negative prompt.

To do:
- Support more positions
- Reduce/tag out striation
- Get enormous tag consistent again
- Get huge pecs tag even larger
- Support more clothing styles
- Increase generation detail
- Make hands generate better