Thank you all for 800 followers. This was one of the earliest LoRas I ever tried training and working with. I'm a bit of an experimenter, which is why there ended up being so many versions of this LoRa. Because of this, I tend to get quite sloppy.
Character is meant to be depicted as 18 years old and older.
Illustrious v3: Hopefully the last version of this LoRa to be made. Properly trained on Illustrious v1.0 on PixAI to work best with actual illustrious models and checkpoints. This was trained on the most images out of all the LoRas. However, many of the images captured her with her "hair up" which is why the hair might stick up sometimes or come off as very messy. Probably a big mistake on my part.
Illustrious v2 (now currently NoobAI): This version was trained on base model NoobAI-XL on PixAI. I didn't know what the difference between NoobAI-XL and Illustrious was at the time but the images looked very accurate on PixAI. I was unaware of the effect they had on this site and wondered why most images came out so "anime looking" here on CivitAI. The base model has now been properly switched to NoobAI.
Illustrious v1: Was actually trained on Pony Diffusion v6 but I just switched the base model to Illustrious here on CivitAI just to see what the result was. Literally the exact same as Pony v2 but with the base model switched.
Pony v2: Trained on more screenshots than on Pony v1 due to my dissatisfaction with Pony v1.
Pony v1: I was still experimenting with what kind of screenshots the AI would be able to train and what not. Still very early on in my AI LoRa development. I remember being unsatisfied with how it turned out.
Description
tanks to nanashianon for calling me out for being a fukin n00b
FAQ
Comments (3)
Seems to work mostly fine, though my advice:
1: Using emo as a tag seems to be spawning some extraneous accessories. These can generally be clamped down on by adding "Print shirt, wristband, choker, ripped jeans, earrings" to negatives (this one can be done without retraining and is just adding it to the description). I'd recommend avoiding using a general, ambiguous tag that will (at best) be redundant with other tags in data set (all the images that would be tagged "emo" would be tagged with the trigger anyways) in the future (some very general tags like "1girl" are often going to be on all your dataset but they can possibly be removed from the prompt in certain situations).
2: I'd recommend a less ambiguous trigger. Using triggers like "Candace Flynn" will likely run into issues eventually between the base model already knowing a character to some degree (and may already be causing some collision), the multiple common words being read as independent tags, and (rarely) Civitai's onsite generator finding a real person's name in it and thus refusing to do anything with it. Using something entire artificial like "CanFl_EmV" will greatly reduce the risk.
3: I think it's a bit undertagged. Not a critical issue with a relatively simple outfit, but the more outfit parts you can nail down to a tag, the better the learning. The big one is that that (from the pictures I can find of the episode, haven't seen it) she's actually wearing glasses, but it has been mislearned them as eye makeup instead due to lack of tagging them as glasses. I think "glasses, t-shirt, white belt, belt buckle" are tags you could have used.
4: Are there any pictures of her in this outfit from behind without her eyes visible? Seems overly biased toward including the eyes if shown from the rear. May be show accurate.
Tanks for the advice. It's true. I didn't initially put in the "glasses" trigger because I didn't actually trust the generator was going to deal with it very well, ESPECIALLY since she is mostly seen in this outfit with one eye covered. I thought that was going to confuse the generator and force open her covered eye. But it seems to work fine. Still, I put in as optional because I personally like the eyeliner look rather than the glasses.
@frozenfapper69856 It won't be a problem as long as you're consistent with the way the glasses are tagged (don't use "black-framed glasses" in one image and just "glasses" the next, make them all "black-framed glasses") and have a decent number of pictures where they're visible with some showing they're clearly glasses. My peliminary data on footwear (that I really need to finish up and publish) shows ~7ish is the number needing for relatively complex footwear to be consistently rendered correctly, and simple stuff only needs ~4 images in training data. Simple glasses shouldn't take that many to get right (though including more pics is always fine as long as they don't bring down the average quality for no reason), especially when there's a more specific tag (black-framed glasses) with high usage.