Note: there's a <a rel="ugc" href="https://civarchive.com/models/90261/l666-lora">LoRA version </a>of this model which generates better faces. Further development of this model will mainly focus on the LoRA for now.This model generates pictures of women (mainly Asian and Caucasian) wearing beautiful lingerie. Note that the "L" is upper case so not to be confused for a 1.This model supports most Danbooru tags related to people, women, fashion, lingerie, etc.Recommended tags:<ul><li>Asian or European. If you specify Asian, you'll get Asian girls. If you specify European, you'll get white girls. If you don't specify either, you might get either or mixed-race girls.</li><li>bra, panties, underwear, etc: that's the main thing this model is trained to produce.</li><li>good aesthetic, good quality, good blur, good noise, good contrast - Put any or all of these in your prompt if you want to the image to look better. Note: I've programmatically labeled images this way, <a target="_blank" rel="ugc" href="https://cafeai.notion.site/WD-1-5-Beta-Release-Notes-967d3a5ece054d07bb02cba02e8199b7">similar to how WD1.5 was trained</a>.</li><li>pntyad: Some of the images were trained on lingerie/panty ads which may result in higher quality images but with a greater possibility of watermarks, text, signatures, and other artifacts appearing. Highly recommended that you add the appropriate negative prompt.</li><li>(Negative Prompt): poor aesthetic, poor quality, poor blur, poor noise, poor contrast -Put any of these in your negative prompt if your images aren't looking as great.</li><li>(Negative Prompt): <a target="_blank" rel="ugc" href="https://huggingface.co/yesyeahvh/bad-hands-5/tree/main">bad-hands-5</a> - this textual inversion seems to work very well when you're not getting great hands.</li><li>(Negative Prompt): <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/Nerfgun3/bad_prompt/tree/main">bad_prompt</a> - similarly, this textual inversion works great when you're not getting properly formed limbs.</li><li>(Negative Prompt): text, logo, signature, username, watermark - A lot of images in the training set contained logos and watermarks so you'll want these in the negative prompt.</li></ul>Note that this model seems to work better with upright (512x768) and square images rather than horizontal ones.

L666

This model was trained using EveryDream on about 78K images of women wearing various types of clothing, mostly lingerie. The "666" comes from this model being trained in batches of 666 images, with a conditional dropout of 0.0666, a learning rate of 6.66e-6, and 6 full epochs. Though this may not be the most optimal set of training parameters, it certainly turned out better than what I expected.

Files with SHA256: 46b09d2e3bf9c019d04df205d874d5171fa00a8ddebd65085c373e15ff696006