This LoRA was trained on still images of bizarrely distorted faces. It responds to the word caricature and descriptions of exaggerated facial characteristics.
An example of a prompt that will draw out all the exaggerated features is: caricature of a person with exaggerated features including a long face, enormous bulging eyes, big drooping ears, a bulbous nose, gigantic lips, and an expressive face.
Z-Image Turbo:
All sample images were created with 9 steps using res_multistep/simple in Comfyui. Doesn't seem to work on animals as powerfully as the Wan version does.
Wan:
The largest portrait orientation I could go without ridiculous mutations was 768x1024. Hands are usually borked, unless you do some kind of hires-fix. Also, each gallery image contains a ComfyUI workflow where you can see the settings used.
Description
FAQ
Comments (15)
This is why I'm STILL in therapy 🤪
Too Many GOOFS!!!!
Hahahaha, this one is so funny.
Love it! Well done 🙂
Really extraordinary lora. Continue your amazing work! I am having so much fun using it.
Thanks, glad you're having fun with it! And thanks for posting what you've made with it - I love seeing what other people think up :D
So coool!
Excellent! Thank you.
This works really well! From scary-evil faces to cute-dweeby style. But I didn't read the description well, and launched right into making test vids. And it worked fine! It worked as well as most other LoRA I've tried so far.
This was very encouraging because I'm considering playing around with Wan 2.2 training and I didn't want to have to animate everything just to train on something that's an image style, material or even an object. I honestly don't have any specific movements I want to train right now. Eventually I can probably come up with some weird morphy things in Cinema4D, but I'm lacking most of the basic LoRA infrastructure I have with Flux. Wan 2.2 is a fast and capable T2I model. It just needs more LoRA that aren't boobs, girl's faces, sex positions, etc. 😉
I really loved your goofballs lora. Will there be a Zimage Turbo/base version too?
Yes! This is one of my favorites. I trained a Turbo version, but never posted it because it didn't quite take the same way. I'll train it on Base in the next few days and see if it works. Based on what I've seen so far, I'm optimistic :)
@bblink787 Oh god that would be great. Now that I only use Wan 2.2 my results with your LORA in T2V aren´t good anymore. For T2I it´s okay but it takes too long to generate. Your LORA is my alltime favourite and I haven´t found another one like it. I believe in you! Thanks for your excellent work!
@Hasenbein I think I like the Z-Image version better. I merged two LoRAs I trained with different settings and it's more versatile than either one on its own. I hope it works for you!
@bblink787 What a nice surprise! I have problems with my internet connection. The download takes ages. I fear I will only be able to try it out tomorrow evening. Thanks a lot in advance!