<h3 id="reignsoles-v2.0-(anima-preview-2):">ReignSoles V2.0 (Anima Preview 2):</h3>This is an upgrade in almost every sense of the word. Images in the dataset were further selected for and refined, more images were included in the dataset, improved captioning, improved natural captions, etc. It was trained on Anima Preview 2 rather than the original Anima Preview. Furthermore, the LoRA was trained using tag dropout to ensure even simple prompts could generate quality images. Similar to 1.0, each image was manually captioned with a natural language description along with Danbooru style tags. This Lora should also be capable of realism to a limited degree, as it was trained on a subset of real life images of feet as well, each again personally tagged and manually captioned. Finally, I lowered the size of the LoRA to increase it's compatibility with other Anima-based models aside from the original Anima Preview. Lastly, I added in the images from my other illustrous LoRA into the dataset. This allows it to generate feet in the 'foot pussy' position as well to a high degree of accuracy.The tag "perfect feet" was used on the images in training which had feet I subjectively liked the most. And the tag "detailed soles" was used when the sole of the foot had visible depth and wrinkles. The tags "real feet", "real soles", and "realistic" were added to every real life image, so use those tags exclusively when prompting for realism.<h3 id="reignsoles-v1.0-(anima-preview):">ReignSoles V1.0 (Anima Preview):</h3>This LoRA enhances the quality of the soles of women's feet for use with Anima in ComfyUI. It was trained on 109 high quality hand-selected images of women's feet from Danbooru. Said images were all manually captioned with a natural language description along with the accompanied Danbooru tags.Using tags such as "perfect feet" or "detailed soles," you can expect to increase the quality and realism of the generated feet. In addition, it enhances text to image accuracy when describing the position and pose of the feet relative to the other parts of the body / image.

ReignSolesAnima

Trained for a total of 75 epochs on 109 images. Each image was captioned with a mix of Danbooru tags and natural language. Rating Explicit, Safe, etc were discarded. The best images were captioned additionally with "masterpiece, best quality". Rank 64, Alpha 32, LoRA Dropout Rate 0.008, Learning Rate 1e-5. Image watermarks + japanese text in the dataset were manually removed. The zip file Civit Ai marks as a 'dataset' is actually just the full training settings I used in conjunction with SD-Scripts to train this LoRA, in case anyone else wanted to play around with said settings.

Files with SHA256: 035bf84436cb441cff16c2df63d87236f2b2906680a550bf4304e98aaa5d83f0