Infinite Alien Dildo Generator

If you like weird dildos, this is the LoRA for you! While this is primarily made around the concept of tentacle dildos, due to the vast variety of strange dildo forms it can generate, it is more like a general fantasy alien dildo generator. It can be somewhat unstable, and some checkpoints/LoRA/prompting may not play well with it without some settings tweaks, but it can create some very good and interesting results. Prompting for attributes is one of its biggest strong suits. Size, color, multi-color, texture, transparency, translucency, glow, it can do it all. Only your imagination is the limit!

_________

Recommended strength: 0.5 to 1.0

Trigger word: TentacleDildo

_________

Training & selection insights: This one was very difficult to train and select for, but this specific version is the best middle ground between quality, variety, and strength range. I can't remember exactly, but it probably took sifting through dozens of training runs and over 100 versions, with thousands of image generations. Out of those, only 3 stood out as final candidates, which were then meticulously compared against each other. Still I feel like this version isn't quite as perfect as I would want it to be, but at the same time it also exceeds my expectations in other ways.

Its only real drawback is that its keeper rate isn't as high as I would like, but that may also be a big ask in this circumstance, given the nature of the LoRA itself. I typically shoot for a keeper rate of 80% or higher, but due to the inherent variety of this LoRA, its keeper rate can fluctuate. I would say anecdotally, it probably has a keeper rate of at least 50%, and depending on the prompt and settings, can be in the 80-90% range. But it's a balancing act, and given some of the particularly interesting results one can get out of it, I think that all in all, it gains a lot more in that regard than it loses in its fluctuating keeper rates.

Captioning insights & prompting tips: As with all of my LoRA, this is captioned with JoyCaption, not just a simple trigger word. I think that aids the capabilities of this LoRA greatly, and it was pretty much essential for proper training. When you're dealing with training on soo many forms with soo many different textures and colors, captioning those details really helps. Prompting tips will be pretty hard for this one due to the sheer variety of possibilities. This was not trained on any furry stuff, so you won't exclusively find any of that here, but it probably can be prompted for manually, and I'd say this LoRA can be used as a helper when used in conjunction with ControlNets or other LoRA.

As some general tips from my own observations; You can get some body horror with this model, but it's not exactly the model doing it, because there were no people trained on, so all of those details come from the checkpoint itself. Even despite not being trained on people, it can still of course be used for insertions. if the image is looking a little noisy or banding in the background like a low quality GIF, then lower the strength a little, change what checkpoint or LoRA you're using with it, and/or adjust the prompt. Even decimal places in the LoRA strength can make a difference.

I can't account for how everyone prompts, I generally try a few types of prompting strategies, along with various settings on/off, and combining with other LoRA and checkpoints, so I test things extensively before I actually release them. But I can foresee some people having issues not getting good results if they prompt a certain way or rely on specific checkpoints or settings. All I can say is keep experimenting and you'll be amazed by the results once you dial things in.

Since this model is soo reactive, the longer the prompt the more instability/strangeness you will likely get, so short prompts are better for coherency, but long prompts can also benefit it as well, depending on how strange you want things to get. Depending on the checkpoint, you may not even need to use the full trigger word, something as simple as "dildo", or no trigger word at all could still produce desired results, and can produce different varieties, since the captioning also recognized the object itself as a "dildo". But these concepts are different across checkpoints, so the results will vary.