@sabi123456 Oddly, one checkpoint was making a 1978 version of lois lane. Another was making a version of her I do not know. But they were all brown hair and with the same hair style. Oddly, none had bangs like the 78 version tho and instead had their hair parted in the front. No clue what it's trained on but was very surprised to find loislane also prompted the same thing but around 60% strength of lois lane as the prompt.

And I also got superhero like additions on JLU-LL on one checkpoint but not others. It could have been a coincidence but I didn't have that result using JLULL on that checkpoint thus why I landed on that one. Made me wonder if somehow JLU was trained into one of them and the hyphen was separating it to prompt somehow. It shouldn't work that way but there is A LOT of stupidity in the core construction of SD so I won't underestimate it. Still dumbfounded they didn't include subgroups for prompts when things like color bleed across prompts without isolating it.

@sabi123456 That data was all in regards to NO LORA/TI. It's already trained into checkpoints by default for the most part. I downloaded and examined your LORA since then and I can see there is an issue in general.

You trained this like a concept and not a person. You gave it no trigger and yet somehow ended up with the partial trigger of "loislane woman" which isn't included on this page as the actual trigger. Every single image's caption text file was "loislane woman, lois lane, dc, cartoon, animated, illustration".

You need to run BLIP or such to caption your training data to decide what her clothing is prompted as and then try the prompt choices from that captioning in SD just to test that it works and didn't pick the wrong shirt style for example. You also need to prompt the background content. Afterwards, you'll end up having to edit every text file to look something closer to... (Just an example as idk what your training images look like to be more exact)

JLULL, cartoon, sky, city, buildings, car, blazer, skirt, pearl earrings, gloves, purse,

The idea is to create a trigger that doesn't exist in checkpoints as a prompt. This prevents what I spoke about in earlier comments thus JLULL at the start of every text file. The prompts after that which you see above are things that SD will remove from training into the LORA itself. Anything left over (not prompted) is then trained into the LORA. No hair, eye, face, body, breast, gender, and so forth. You can add certain things if you wish for it not to learn them (like long hair) but it's best to train them by not prompting and allow the user to adjust prompts to enforce or remove these things imo. To each their own on that last part though as it'd really depend on what you're making most times and how SD handles it.

Note: I've seen some users prefer to use a single line of prompt such as "JLULL, a cartoon woman in a city with buildings with a purse wearing a blazer skirt gloves with pearl earrings" but I've no idea how well that does or does not effect training and it's too messy for me to want to ever use but I have used it as "a cartoon woman in the city wearing a blazer and skirt, gloves, pearl earrings, buildings, sky" and I noticed no real difference between just single prompts or a single long prompt.

Also, you should avoid double prompting like you did with "cartoon, animated, illustration". I'm not sure which one is best tbh but I have used cartoon with great effect when I'm unsure of which to use though I've not tested it enough to know if one is better than another.

Granted, as I said before, SD is not made very well and training is gambling really. If they were smart, then LORA training would be done with detection of objects/things in the image and generating a mask of each item that you then caption and select to train or remove it thus giving captions to what you train and so forth to increase training data farther and more precise. As it is, the method I spoke of above is how you are supposed to train but sometimes it can be a total crap shoot on if it works how you want or not. But your training method was for concept and not a person and thus it's not very versatility.

-

I used the 100 version as it's the first. As far as using it as is, it pulls the art style and her outfit too heavily. To get control of the style you need a weight of 0.2 and to get control of her outfit you need a weight of around 0.4 but that pulls the art style too heavily. And you can make it work for the most part at 0.2 but it requires extra prompting like black hair. It is actually possible at this point (low weights) the actual prompt (not trigger) of lois lane is making this work better at those low weights. Without any actual trigger, it's impossible to lower the weight of the trigger and test farther.

It can still be pretty good using at a weight of 0.3 and "loislane woman, black hair" in pos & "blazer jacket, skirt, brown hair," in negative as needed depending on the outfit choice. Give it a highres.fix of 0.44 on Swin2SR_Realworld~~ and it cleans up the face and style a bit. But LORA aren't supposed to be this difficult. If you don't want to try altering all that to remake it, those are the settings I've used to make it more useful. I'll upload those two images so you can see how/when it does work.

Just a heads up if you miss the actual edit on the 3rd image... Lois Lane with and without the lora look nearly the same appearance wise and this LORA works a lot like an art style LORA. I couldn't prompt out the art style at higher weights no matter how hard I pushed it. And her outfit pushed into the one you trained on no matter how I prompted either. Oddly, she often came without a skirt and wore a thong without any prompting. Hence the bikini which at higher weights was a blazer jacket bikini.

Steps are/were better at higher numbers with this sampler thus is why I use(d) it at 150. I only had it work better at a value below 75 on a single LORA/checkpoint combo before and I do tests of 55/75/100/125/150 before picking which to use, this was best for me. You're using euler a I believe and that works better at lower values for me in case that seems wrong, just a different sampler and what it likes/dislikes.

Let me know if any information is missing but I'm pretty sure this helps show the training data wasn't setup extremely well unless you ONLY want to make a cartoon version of her with that outfit controlling her look.

And I tried other checkpoints, just this LORA took so much control over the look that even realistic checkpoints return the same results as western cartoon styled ones.