Are you tired of people walking around with big stupid faces?
Do you secretly wish for smaller faces on your fellow man?
Don't fret – Kirkfaced is here! Now for FLUX!
Trained on portraits I manually photoshopped to have a less obscene face size. Biased towards headshot portraits and can struggle to produce sideways facing people, and is a bit stubborn when it comes to create people in action. I haven't managed to get any notable effect on animals. Will fix it for potential future release.
Triggerword: smallfaced
Use "with giant fat head and a tiny face" in your prompt if you want to help it along.
Description
FAQ
Comments (12)
Wow, how could I have been so wrong all this time? I'll never generate a big face again...
Simply put, they're just too vulgar!
what happens if you use this lora with negative slider? Does face become too big? :D
And btw, how many images was it trained on? (Perhaps you could say a few things about the training, it will help with the inference)
Trained of 17 images total. 9 unique images, with 2 different crops and captions for 8 of them.
All of them are celebrity 1/4-headshots, from chest and up. Plan is to do a second training later, with a more diverse dataset, but photoshopping small faces in profile was a pain :)
Training was done already at epoch 3 (out of 12), which kept it somewhat flexible and still produced desired effect – but that means you might need to crank up the LoRA weight quite a bit (1.6 - 1.8) if you want more humorous photos. (but this of course force it more to be a headshot)
Captions were basic, like "smallfaced, A wearing B and C" and "smallfaced portrait of smiling A woman, wearing B and C".
Prompts like "Tiny face, big head, large forehead, large chin" etc all help with the concept
Let me know if there is anything else.
Been mainly using "TurboVisionXL", "AetherVerse Lightning XL" and "Proteus-Rundiffusion with clip" for these.
Oh, and regarding negative value. Yes, you do get some funny results, people looking like they are in pain :) But not sure if it's 100% consistent, haven't done that much testing on the negative :)
@WizardWhitebeard yeah thats awesome thanks! 17 images only! Wow thats incredible. Any specific sampler?
@WizardWhitebeard the negative is more likely sad -> emo -> gollum depending on the neg strength value. some can be mitigated in prompt.
@yofoton174609 just the one best suiter for each model. Tend to default to DPM++ SDE Karras + Hires-Fix
@yofoton174609 yeah, if it's something SDXL already has a grasp of you can get away with using really few images. My most downloaded LoRA – the pulp comic book one - was trained on a single image(!)
Unless I already have a large dataset, I train a LoRA with the images I do have, and use that + CN / inpainting to create more images that can be used for training. Sometimes you end up with something that works ok right from the start.
In celebration/solidarity w/ your post here I've re-published a Charlie Kirk LoRA on my channel from a while ago. Zooooom
Hey, liberals:
checkmate!