Age Progression [Concept] {Pony}版本v1.0 (ID: 924381) 综合资源合集综合资源合集

Image posts are highly appreciated! You don't even need to leave a review. I just love to see what people are creating with my stuff.

WARNING AND FULL DISCLOSURE: This LoRA is finicky and controlling it requires a good bit of guesswork, as well as some/a lot of trial-and-error in some cases. I highly advise both reading this description and fully looking at the metadata on the example images (click the ⓘ at the bottom-right of any of the images, then click the copy button beside the "remix" box, and paste that into Notepad) to see the prompts used, the image dimensions, and the number of sampling and hi-res steps used. I'd advise copying my usage to start unless you're confident in your abilities.

Produces an image of a single character at several ages/growth stages/body styles at a left to right slant. The main triggerword is ageprogression.

Beyond this the subtrigger words are a mix-and-matching of your choice of young/teenager/adult and flat chest/small-medium-large-huge breasts. Using these subtriggers are largely required or else it'll likely spit out several versions of the character looking exactly the same.

There is no consistent way to control how many "characters/ages/forms" show up in an image. Best practice is that if you want more to show up, make the image wider or shorter. If you want less, make it thinner or taller. Aside from this, you can play seed roulette until you get a few images with the number of forms you want and then reuse those seeds to tweak your results, as the number of forms should generally stick around as long as your prompt changes aren't drastic.

"Why isn't this LoRA available for use with CivitAI generation services?"

I checked before deciding on the subtrigger prompts by checking that the tags young and teenager were allowed on CivitAI by making sure images using those prompts were allowed. I found the examples were in the thousands, so that I assumed they were safe. But given that my sample images had to pass a moderator check before my LoRA could finish publishing, and that a user informed me that those tags can't be used with CivitAI's generators, it seems I was mistaken on that. So between the difficulty of someone getting the intended results with half of the subtriggers being unusable, and that as I mentioned earlier this LoRA requires a lot of trial-and-error, I've decided it's better for users to not waste their buzz/time trying to get results from this. If you want to use this LoRA you'll need to either find a generation service that allows those prompts, or, if it's within your means, the much better option of generating locally.

Donations accepted through my Linktree if you wanna support me.

Information irrelevant to LoRA usage beyond this point

Issues/ranting:

This LoRA attempt was number... I lost count between the months of coming back to it. 15? 20? 25? There were even two attempts made after before I went back to this one. This attempt was trained solely on images with 2 to 3 "characters/ages/forms". Trained using the triggers 2form and 3form. These DO NOT WORK as far as I'm concerned. You are free to try and use them, but my testing lead to failure and I write them off. The LoRA/Pony/Stable Diffusion/whichever also ignores the fact that all the training data had only 2-3 forms and decided it's completely acceptable to generate images with 4, 5, and even 6 forms. In a previous attempt I even created two separate LoRAs for this. One trained entirely on 2 form images, the other on 3 form images. Even with the training data separated neither one could consistently only generate only 2 or 3 form images. I am absolutely at the point of giving up on any understanding of how the AI is processing (or not processing) this concept. Unless SD3 is somehow amazing at comprehending complex concepts, please don't ask me to update this. My sanity can't handle going through this again.

Fun stuff:

Due to the absolutely appalling lack of good quality AP imagery across the internet, I went and generated my own images and stitched them together to create a bunch of images to throw into my training data. Between tossing off the bad quality images I found around the net and the fact that I spent an unholy number of hours/days making sure all my training data images had PERFECT hands, this LoRA (at least by my testing and creating of sample images and personal bias) actually spits out images with decent hands fairly often. As a bonus, I uploaded the training data I created over on Pixiv (contains some NSFW images).