Sure I can do that if you like. I've just posted it there also.

I just tried to generate using my model I uploaded to Tensor art and it won't let me because it won't allow the use of any nsfw tags... looking at their policy publishing and generating nsfw is strictly prohibited. So seems very likely this site you've requested I upload to will not tolerate my model...

Not gonna lie, I've spent the last hour trying to figure out the simple task of just uploading a model to tensor art and the amount of hoops it's making me jump through, it says model won't be visible to others without cover image, but whenever I generate or try to upload cover image nothing happens. Tensor art is just really not it. I've tried to upload model maybe it's super slow and it'll appear later maybe it won't appear ever idk at this point the site you've recommended I certainly wouldn't recommend to anyone. idk why anyone would use that site over civit. from my pov it is very frustrating to use that site.

Thanks for your suggestion, I know the one you're talking about, I'm pretty lazy but if I find time I'll try give it a go. Or at least I might try make this model a little more flexible to more easily use to create ponyboys. Maybe training the ponyboy and ponygirl tags into one model? I'll try experiment with it at some point.

@Dazrock Yeah I play around with what tags to include and what tags not just to see if it improves or makes worse. Usually the tagging is a little trial and error. I also just train the model a few times and take snapshots at different Epocs and compare samples from each model side by side to get a sense of which version to go with, each having being better or worse in different ways.

I agree that theoretically you shouldn't tag anything you intend to train as part of the concept in general. But nothing beats trial and error. I think because I found riding had an association with horses I wanted to try and disconnect that association by including the tag, to avoid the girls from becoming too horse like. By including riding as the tag it makes the training sort of less likely to confuse the girls for horses. But yeah I think with riding it was a tweak to tag the riding tag in some images, I think particularly images where the ponygirls were larger to avoid ai mistaking them for horses. Something like that anyway it was a little while ago now.

It could also of just happened to have been one of the runs where I was mostly experimenting and didn't bother editing the tags that much and was more focusing on selecting what images to use and what parameters to use keeping everything else the same, and didn't redo afterwards with tag adjustments because lazy.

@Dazrock  But yes in general you should not use tags associated with the thing you want to train. The reason I included riding is because I wanted to reduce association to horses and riding is associated to horses. So this is more an exception to the rule. Trial and error is best approach it's always difficult to understand how the ai associates things until you test it out. It's also entirely possible that the better results could've just been down to the training seed. But that was the logic I had when doing the batch of model training which I ultimately selected over the other versions I'd done.

@Dazrock Ultimately this is all purely me trying to make sense of the results of my trial and error, hard to differentiate between what is superstitious and not. I think also it could well be completely different what tags are best to include or exclude depending on the base model and how the combinations of associations on the base model work.