This is a LoRA version of TouchKnight's Bowser Model
https://civarchive.com/models/4960/bowser
This allows it to be used more generally as an add-on to your favourite models (the samples above were rendered using a variety of different models)
Credit goes 100% there for training - all I did was extract the LoRA diffsum from the base model.
You can either create Bowser renders in whatever style you want (from whatever base model you use) OR the way I do it - add a little bit of this LoRA to give your characters a stockier look (great for Dwarves!)
To get Bowser
Set the LoRA weight to 2 and use the "Bowser" keyword
To get a stocky look
Either;
Set the LoRA weight to 1 and use the "Bowser" keyword
Set the LoRA weight to 2 and don't use the "Bowser" keyword
The exact weights will vary based on the model you are using and how many other tokens are in your prompt.
note: If you try to run any of the example images make sure you change the Lora name, it's seems Civit changed the filename from what I uploaded
Description
FAQ
Comments (14)
Extracting a Lora from the base model appears to be a a very effective solution to trimming down our checkpoint libraries. Well done! I am going to google how to do this as I can't believe how effective an extracted Lora is. So good. This image is amazing:
https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/94e43557-e47c-423c-b5b7-4a5ded62f800/width=512
I added the above image to the png info tab in Automatic11111 and ran a render. I got a fantastic result as well. Interestingly, I did expect the same render, but didn't, likely due to a specific VAE you are using for that render?
I am definitely going to do this with all my 'style' models that are just tiny adjustments. It's better than merging because when you merge you get all the base stable diffusion stuff watering down your model.
The tool I used is here, but it's a bit finicky because it's command-line only and when you get an error it doesn't explain why (if you get an error about model sizes, I think one of the numbers has to match - IE I couldn't do this with 500 dimensions but it worked with 320 dimensions.)
https://github.com/sashaok123/LoRA-EXTRACTOR
The VAE I used is just vae-ft-mse-840000-ema-pruned, except where a model had it's own VAE. That model was Kotos AbyssProto (a NSFW model - and the thumbnails on it's page have a very specific look the creator was going for, but it's great as a general model and one of my go-to favourites even when not female porn)
I've been testing with others on Discord and for some reason I sometimes get slightly different results compared to others and we haven't figured out why. Even with the same Clip Skip and ENSD.
And it's funny after testing renders that I used to be able to get I can't now. SD is weird.
also note, the console prints "Couldn't find Lora with name bowser_LoRA320" we still get an image. Maybe this is why the renders with exact promt settings don't yeild the same image?
Are you putting it in the (stable diffusion)/models/lora folder? If you are using the <> tags in the prompt it needs to be there, not in Additional Networks. If you are using AddNet, don't use the <> tags (that's why you are getting that error).
I have found using Additional Networks gives a different results than using the <> tags that are created using the Auto1111 native LoRA. Not worse or better results, just different results.
@ShivaFang TLDR everything is great.
I am only trying to recreate that amazing render you have : ) which I can't.
"Are you putting it in the (stable diffusion)/models/lora folder?"
- yes
Render to replicate: https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/94e43557-e47c-423c-b5b7-4a5ded62f800/width=512
The trigger to this minor observation, is that I used png2info on that image and it likely used the dev version name of your LoRA
Even after making the LoRA name adjustment, I can't replicate the above image. I usually don't have a problem replicating a render if all the info and seed are there. For me the purpose of replicating an existing render is to learn and get better at navigating automatic1111 to bend it to my will. I get tweaked when I see a render I love and can't replicate it
Nice work
@damnsaint There is no 'dev name' for the lora, because it's not a trained lora. the script I used literally just outputs the file and it's the same file I used.
It's a complete mystery to me when I am getting different renders than you are. I've noticed the same when I use a cloud render service (because it's faster) instead of my computer.
Here are the command args I use when starting Automatic, maybe it's something there?
webui --skip-torch-cuda-test --precision full --no-half --autolaunch --no-half-vae --api --opt-split-attention-v1 --upcast-sampling
Try removing what's in the brackets for the lora and instead open the extra networks tab and click on the lora there to add it. (If it's not in your extra networks tab that might be the problem).
The extra networks tab is the one in the middle above the 'styles' dropdown.
I just tried to download it.. and civit changed the file name. Change the lora to 'bowserLora_1' and it should work.
@ShivaFang ahhh ...
likely it's here:
--skip-torch-cuda-test --precision full --no-half --autolaunch --no-half-vae --api --opt-split-attention-v1 --upcast-sampling
mine is:
--administrator --autolaunch --api --xformers --medvram --precision full --no-half could be the xformers flag?
This is a helpful exercise. Good suggestion. Will tool with this.
@damnsaint Oh it absolutely is the xformers flag. It's known to drastically change initial seeds as part of whatever it does to speed up renders.
@ShivaFang Wow. lessons learned today! Thank you!
@DamnSaint You probably learned this by now, but just adding for other late readers - Xformers improve speed & performance, at the cost of reproducibility. Basically don't use Xformers if generation persistence is important to you. They introduce randomness that can't be prevented.