metadata
license: openrail
Reddiffusion is a fine-tuned SD-768 model trained on some of the top art shared on Reddit, training consisted of 60 epochs with aspect ratio bucketing with a resolution of 896.
Dataset followed the format of medium-subject-token, an example prompt would be:
a painting of a bird, best of reddit, by Wadim Kashin
This model isn't meant to be a radically different model, it aims to add details to the 768 model and extend its art capabilities while improving higher resolution generation and using different aspect ratios.