Tifa Lockhart <All Outfits> LoRA
Making models can be expensive. Do you like what I do? Consider supporting me on Patreon 🅿️ or feel free to buy me a coffee ☕
Kind of an update/remake to my Tifa Lockhart TI.
Tifa is basically my favourite female fictional character ever, so she was of course one of my first TIs, and wasn't super good.
At some point I noticed nobody had yet made a LoRA of her because there are already so many decent TIs, so I decided to make one, and being this a LoRA, I could fit ALL Tifa's outfits. This one triggers with tifa lockhart but can also accept game/movie names to specify which Tifa (e.g: final fantasy vii advent children for her AC outfit + descriptive tags) and dress tags for the official Remake outfits (e.g: tifa lockhart's refined dress). Check pics on Danbooru for all the tags, or the examples I provided. It also gets very easily her cowgirl outfit and, with a bit of failed attempts, her green dress from the flashback (kind of). Of course it can do her OG FF7 outfit. I used 0.6 weight.
For the pics I used AOM2 and Any4.5, and for one I even pulled out my Tetsuya Nomura LoRA and my Yoshitaka Amano LoRA.
How to use LoRA's in auto1111:
Update webui (use
git pulllike here or redownload it)Copy the file to
stable-diffusion-webui/models/loraSelect your LoRA like in this video
Make sure to change the weight (by default it's
:1which is usually too high)
Description
FAQ
Comments (19)
Oh my god, yes!! I can't wait to use this model! So, so excited. Forgive me if this is a noobish question but: how does one train LORA? Is this basically based on all images uploaded to danbooru? Or a private tifa collection? ^.^
That really depends on what you're trying to achieve
@lykon Does that mean both is technically possible? Very interesting.
Can you share the versions of some of the models you used, specifically for pics where her face is in the direct light?
I wrote the ones I used in description, they're easy to find
what is AOM2 ? And where did you source Anyv4.5 ? Many have not come from trustworthy sources.
@lykon what is AOM2 ? And where did you source Anyv4.5 ? Many have not come from trustworthy sources.
got it so this was abyssorangemix https://huggingface.co/WarriorMama777/OrangeMixs thanks
@muuuuuu yeah sorry I didn't reply earlier. They're AbyssOrangeMix2 and Anything V4.5
"Make sure to change the weight (by default it's :1 which is usually too high)"
I'm not sure who told you this, but if you have to lower the weight of your LORA, that means you did not train it properly.
I understand that a lot of LORA on here require that the weight be lowered, but that is only because people here would rather be "first" than spend more time doing it correctly.
weight ending up different than 1 is an effect of descendant training instead of precursor. Still it's just a scalar value. The fact that you can use the lora on multiple styles/outfits and even realistic models without style bleeding is proof that there is no overfitting. That rentry guide is ancient and oversimplifies overfitting, which is an entirely different concept.
I've proven it already by making countless offset loras and by having examples of this one with different style.
Anyway I'll be making an offset version of this just to show you :)
see this example in the meantime: https://civitai.com/images/765774?modelVersionId=68700&prioritizedUserIds=53515&period=AllTime&sort=Most+Reactions&limit=20
@Lykon So, a LORA is a lower dimension representation of a Dreambooth/finetune. In those cases, you aren't able to adjust the weight. A LORA that requires lower weight to use properly analogous to a Dreambooth that is fried. You wouldn't upload a fried finetune, because there would be no way to remove that effect. You're taking advantage of the fact that LORA weight can be changed as means to avoid properly training.
I get it though, first is better than correct because you can make more money that way, especially when it comes to popular characters. I know how this works.
The laugh emoji makes it clear that you're dismissive of criticism.
I totally understand though. "Training models is expensive" for you. I've been stuck on the same LORA for about 7 hours, and resisting the urge to upload it in it's broken state is getting harder and harder to resist.
I get it.
@Lykon I bought you 5 coffees so that you can hopefully buy better hardware and figure out how to do this properly in the future. Good luck
@poisenbery dude I've been studying AI for 10 years. Seriously. You just read a rentry page and you think you got knowledge on this.
You have no idea what overfitting is. If you could simply fix that with a scalar value, then you solved the biggest problem in AI right now.
best advice I could give you is to try new things. You use anythingv3 and you quote rentry's guide. You seem stuck in 2022 for some reason, things have changed, we learned new things. I'd advise you to try making flexible loras instead of 1-word ones. That's basically a TI, no point in having a lora like that.
done, uploaded offset version usable at weight 1 with proper prompting. Magically fixed "overbaking".
Should be proof that what you read on the Internet is not always true.
@Lykon if i may ask a question, does that mean that offset anime loras perform better with realistic model, than a non-offset? would it still be recommended to tweak the weight blocks when using it for a realistic model?
@SomebodyOnceToldMeh nope, offset just means I moved the sweet spot. To have a lora work on both anime and realistic and on any style you need them to be not style overfitted. Offset won't help you with that.
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.