Ryukijano/CatCon-Controlnet-WD-1-5-b2R

Experimental proof of concept made for the Huggingface JAX/Diffusers community sprint

Demo available here [My teammate's demo is available here] (https://huggingface.co/spaces/Cognomen/CatCon-Controlnet-WD-1-5-b2)

This is a controlnet for the Stable Diffusion checkpoint Waifu Diffusion 1.5 beta 2 which aims to guide image generation by conditioning outputs with patches of images from a common category of the training target examples. The current checkpoint has been trained for approx. 100k steps on a filtered subset of Danbooru 2021 using artists as the conditioned category with the aim of learning robust style transfer from an image example.

Prompt : Prompt: "a chibi of hatsune miku".

Conditioning image: An image from the class of style we want to generate.

Target:

Major limitations:

• The current checkpoint was trained on 768x768 crops without aspect ratio checkpointing. Loss in coherence for non-square aspect ratios can be expected.
• The training dataset is extremely noisy and used without filtering stylistic outliers from within each category, so performance may be less than ideal. A more diverse dataset with a larger variety of styles and categories would likely have better performance.
• The Waifu Diffusion base model is a hybrid anime/photography model, and can unpredictably jump between those modalities.
• As styling is sensitive to divergences in model checkpoints, the capabilities of this controlnet are not expected to predictably apply to other SD 2.X checkpoints.

Waifu Diffusion 1.5 beta 2 is licensed under Fair AI Public License 1.0-SD. This controlnet imposes no restrictions beyond the MIT license, but it cannot be used independently of a base model.