Link to training programs:
Initially created an unconditional UNET model that denoises a 128x128 rgb image to create an image of a tattoo design
Trained on a dataset of 4.2k tattoo designs
After 10 epochs of training:
After 30 epochs of training
After 50 epochs of training
After 50 epochs there were some images that looked like tattoos but some images did not. So I decided to start over this time training a conditional unet model that receives CLIP embeddings of the prompt and to use a Variational Autoencoder to compress the images so I could train diffusion of higher 512x512 resolution images.
Trained on a labeled dataset of 4.2k tattoo designs
After 1 epoch:
Prompts:
a dragon on a white background a fiery skull a skull a face a snake and a skull
After 10 epochs:
After 20 epochs:
After 30 epochs:
After 40 epochs:
a dragon on a white background a fiery skull a skull a face a snake and a skull