Turn anyone into a wojak.
STRENGTH: 0.5-1.0
There are two models "flexible" and "strict". The strict model makes images adhere more to the mspaint style original Wojak images. The flexible model should be a little more flexibile with facial expressions and character types.
WEAKNESSES
It's not great at non-human faces by default. It might take some prompting like "Make them into a wojak. His head is a skull." or "Make them into a wojak. His face is a cartoon mouse."
The eyes tend to look off to the side, the way that most wojak images do. This can sometimes be fixed with prompting, e.g. "Make them into a wojak. He is looking at the camera."
It's not great with closed eyes, by default. Again, you can usually prompt around this, e.g. "Make them into a wojak. Her eyes are closed."
Sometimes the background will come through. Just add "Blank white background." if this happens. I haven't tested much if it works the other way if you want to keep the background.
"Why is it so big?"
I tried making a lower rank LoRA, but unfortunately it seems that Wojaks are not trained into this model at all. I had to bump it up to rank128 to get the iconic Wojak nose to come through.
TRAINING
This was trained on 207 image pairs. The prompt was always "Make them into a wojak."
The image pairs were generated from actual wojak images turned into real people using qwen-edit-2511. There was a lot of pruning of the image pairs and testing of the models that were saved at different steps. The strict model was saved at 8000 steps. The flexible model was saved at 4750 steps.
Description
FAQ
Comments (5)
How do you train edit styles? Do you have to train with a real image and a drawing in pairs? Or do you just train the art?
Pairs.
Most of the work is in creating a good dataset. If you have garbage in your dataset, you will be training garbage into the model. The hardest part is deleting things from your dataset because they are not up to par (and you don't feel like working on making new ones anymore). After that, the most time consuming thing is testing the different models. You have to run a lot of comparisons and probably set up some scripts to figure out which model is best. It took me about a solid week of work to make this LoRA.
Can you share the dataset or make a Klein version?
@vigilence I won't have time to for a few weeks, at least. I made this LoRA when I had a break from work.


















