I wanted to test out training character LORAs which aren't present in the dataset using my normal methodology with Illustrious-XL, so here's an Ellen LORA.There are three versions:001 uses only screenshots of asset ports to MMD and Koikatsu. Trained to see how well this method works on Illustrious.002 replaces many of the Koikatsu screenshots with a few images of the most on model fan art I could find. Trained to see how bad the first one is.010 is trained using a workaround for base_weights. Around 90% of the training images are screenshots.<s>Both models were trained using a style LORA and SD-Script's base_weights functionality, but my attempts at training one that works well for this purpose haven't been going very well. As a result I think that 001 has a lot of style burn in. Using both LORAs in conjunction at lowered weights might produce results that are relatively style agnostic while still remaining accurate.</s>As it turns out base_weights doesn't work correctly in the versions of SD-Scripts that I've been using to train Illustrious LORAs. As a result versions 001 and 002 were actually trained against the base model instead of one merged with a style LORA.

Files with SHA256: adaf79db640abae6f40e0c78bd6ce78fad00e7671c728250505967d70a5bbc93