❤ Please post any images made with this model here for further improvement of the dataset ❤
Prompt Format:
duo selfie pose, two <men/women/things> next to each other, a <thing> in the background
Optional tags include:
<happy/shocked/excited> expressions, open mouth, pointing to the side, selfie pose
Versions:
v14 (Recommended): U-Net only; Much more flexible prompting by training with a known similar concept (duo selfie pose).
v5 (Previously Recommended): U-Net only; Improved dataset captioning for more flexible prompting. It's still kind of finicky but it's an improvement.
v8e (Experimental): Text Encoder + U-Net; Trained for fewer steps with a higher weight decay (0.02). Adjust the individual weights of the Text Encoders and U-Net if you don't get the pointing.
v4: U-Net only; Initial release, tends to produce women wearing lots of neck jewelry
Prompting for older versions:
Prompt format:
two <men/women/things> pointing at <something>, the <man/woman/thing> on the right is pointing at <something>, the <man/woman/thing>on the left is surprised
If you're having trouble getting the arm to point, try adding tags like "pointing a finger", "fingerless gloves", etc.
Description
Very experimental version that trains both Text Encoders and the U-Net. Try adjusting the weights of the Text Encoder and U-Net separately. If you don't get a closeup, increase the U-Net and Text Encoder weights.
Syntax:
<lora : model_name : text_encoder_weight : u_net_weight >
<lora;sdxl_wojakpoint_v8e-step00000800:1.4:1.5>