WORK IN PROGRESS, but capable of giving some pretty nice results.IMPORTANT: LORAs work differently from normal models. To use them, install this extension, then use WITH Elysium Anime v3:<a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-webui-additional-networks">https://github.com/kohya-ss/sd-webui-additional-networks</a>You can grab elysium anime v3 HERE:<a target="_blank" rel="ugc" href="https://huggingface.co/hesw23168/SD-Elysium-Model/tree/main">https://huggingface.co/hesw23168/SD-Elysium-Model/tree/main</a>As the name implies, this is a LoRA that draws heavily from Steins;Gate illustrator Huke's stylings. The strange, photorealistic cities, the splotchy watercolors, the near-blinding brights, it's all here. Trained with 150+ images. Use with Elysium Anime v3. The safetensors version works identically to the ckpt.TO DO: -Add a prompt frequency list-Refine training data to prevent having to prompt wrangle-Add daru :,( and everyone besides the main 3 I guess-Make this description better when not tiredTraining data is here, since it's been requested! Had a data loss incident shortly after training, so not everything is here yet, but I'm working on recovering stuff. Here's the main chunk of data I DO have, which are 104/109 images for the style:<a target="_blank" rel="ugc" href="https://mega.nz/folder/pHcTjKZQ#p0iagKQhr_CDm-gUWcAp4A">https://mega.nz/folder/pHcTjKZQ#p0iagKQhr_CDm-gUWcAp4A</a>LORA training for the style was 8 epochs, 5 repeats. Also used were three twenty image folders of Mayuri, Okabe and Kurisu, repeated twice each. It wasn't enough to change output too meaningfully, but I have a feeling if you experiment with more repeats on similar folders, you might get better results.

Files with SHA256: 2f6e14d44b4bb2bcc3d9833deee9d946bf2fd98a11943ed2088c3d3fca3e17c5