This is a finetune model based on <a target="_blank" rel="ugc" href="https://civarchive.com/models/24383/grapefruit-hentai-model">grapefruit</a> checkpoint.I used around 1k images from gelbooru by artist Muririn and Kobuichi for finetuning. Initially, I wanted to just train a Lora or standard dreambooth, but neither really achieve anything even remotely useful. If anyone have tips for finetune SD models, I am all ears.About the examples:The first one (hot spring one) was created with simple prompts, I selected this one because it also genereated text windows just like a galgame. But this can be negated with negative prompts.The second one (Ningguang) used <a target="_blank" rel="ugc" href="https://civarchive.com/models/8546/ningguang">this </a>Lora. I tested a little bit, for all loras I tested, the image will be over saturated if I don't lower the lora weight. I simply dialed the weight down to 0.5. Maybe there are better solutions.The third one used the same prompt with <a target="_blank" rel="ugc" href="https://civarchive.com/images/329636?modelVersionId=29179&amp;prioritizedUserIds=39650&amp;period=AllTime&amp;sort=Most+Reactions&amp;limit=20">grapefruit intro image.</a>I didn't really play around with the parameters to explore further. These are just some random samples that look okay-ish. Overall, I found the finetune model is more susceptible to distorted images such as missing limbs (hot spring example). I assume some merging with more stable models or more prompts are needed to mitigate these issues.Lastly, trigger words yuzu$style is not a must, but it does emphasize the style a little more.

YUZUFinetune

This model didn't train text encoder. I.e. U-net only finetuning.10 epochs with only 4 repeats for image. 25% caption dropout.

Files with SHA256: 434c2b4b02d013700a482c93bed49d3392c46ff0194936aadde4e3ca4140e913