VERY IMPORTANT: FOR THIS TO WORK WITH AUTOMATIC1111, BE SURE TO ALSO DOWNLOAD THE CONFIG FILE AND PLACE IT IN YOUR MODELS FOLDER ALONGSIDE THE MODEL. This may change in the future as automatic1111 support for 2.x matures.
This is a finetune of Stable Diffusion 2.1 (as opposed to 1.4), and the results it produces are way more interesting. Click the 'i' icon on the example images of how to prompt it (negative prompt is important here). If you feel like it, please experiment and leave any interesting discoveries in the reviews.
License: I don't believe models ought to be copyrightable, so you can do whatever you want with this, including uploading and redistributing it elsewhere, finetuning it, berrymixing it, etc. I do humbly request that you share any models that you make with this one, but you're free not to.
This model was finetuned with StableTuner with a dataset of ~250 manually captioned images. Learning rate was 1e-6, 250 epochs, batch size of 24. Text learning cutoff was 75 epochs to prevent overfitting.
Note: I've accidentally discovered that this model does some amazing non-icon art. See the cityscape picture for an example of another way to prompt it.
