Akira Toriyama style 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 ☕
Triggers with toriyama akira. I suggest weight above 0.5.
I've uploaded the 60 epoch and 40 epoch version. Both can work depending on the model and the desired strength.
Also I've found it gives usually better results with CLIP skip 2 (except the first image, maybe that was lucky).
This is trained on Anything v4.5, so use that or AbyssOrangeMix2 if you want accurate results. Other models may still work.
A big difference from my other LoRA's is that this one basically requires highres fix at least at 1.5-2x size.
Keep in mind this is LoRA, so you need to use it as such. It's not a standalone model and it's not a TI.
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
Yeah I know the first pic was made with v1, but it was a bit lucky and so I still wanna use it as preview for the page, as this version is the front one. I managed to get a similar one (thing that I wasn't able to do with v1, despite being the one that generated it)
FAQ
Comments (13)
That's a really great LoRA, if you don't mind me asking, since I don't know too much about LoRAs, how much time did it take to finish this LoRA?
Do you know a good tutorial for creating LoRAs?
you mean the training time? I don't remember. Less than an hour on a colab for sure. There are various tutorials now, but the only thing that matters is the dataset.
@Lykon Not the training time, the complete time from start to finish, setting up the datasets etc.
@vynenave926 didn't count. I can tell you the longest part is making the examples
@Lykon Last question if you don't mind. Since this is a style, how many datasets did you have? Because for the LoRA training you can use multiple datasets as it seems and I was wondering if for a style , you would use multiple datasets or one big one?
@vynenave926 multiple data sets are for multiple different concepts. If the style is the only concept, there's no need for more than one data set. And afaik, you also don't tag anything besides your desired trigger word.
Don't work with another character LoRA model that is not the Goku one. Please fix it.
can you post an example? This lora is pretty old and trained using the precursor method. Help me understand if there is actually a problem, so I might consider re-training it.
@Lykon
You see, I tried to generate images of Asuka using this LoRA next to yours, but it seems that the style of the LoRA overlays yours, and it never looks at all like Toriyama's artstyle after completing the hires fix. I'm using Anything V4 as the checkpoint model.
https://civitai.com/models/20058/asuka-langley-souryuushikinami-evangelion
PD: And I would also be very grateful if you could re-work it so that it would not depend so much on the hires fix, and it would be easier to get better results directly with SD Upscale. My GPU has 4 GB of VRAM, so I practically have to pray that my local copy doesn't crash when I finish generating an image with hires fix enabled at 1.5x.
@Blake_Gat I didn't check, but from what you're describing, that model got style bleeding. You can try to lower the weight of the character model or find a less overfitted character lora. Have you tried with some of my character loras to see if the problem is really the style one?
Thank you so much! ^__^
Toriyama Akira R.I.P
Hi, just wondering - which is the "most trained"? "_3" has only 20 epochs while v1 has 60? Shouldn't they be named the other way around? Are the names correct?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.