Creates a sequence of a character losing their clothes. It's not focused on the professional art of stripping although I'm making sure it's well trained on poles and poses. Basically this lora does two things:<ol><li>Keeps the subject looking the same between poses. Tags like hair color, eye color, etc. will apply to all images of the subject. Articles of clothing should also keep their consistency.</li><li>Targets other tags like which clothing they are wearing, expression, pose, etc at different views. You can't really control where they go but due to training it gets pretty good at guessing.</li></ol>The higher the strength, the better it gets at these, but at the cost of image quality. I like the look at about 0.8 strength; above here the cost in quality gets more apparent, and below here the lack of consistency is noticeable.<h3 id="v4-update!-khs9rwun7">V4 Update!</h3>Replaced a lot of the training images to get things even crisper looking.<h3 id="tips-7pjv6r1nn">Tips</h3><ul><li>Due to the random nature of diffusion, some seeds will never lead to a "solution" with a logical sequence. Image-to-image variation can help if you can get close to what you want but need to fix details.</li><li>For a seed that works, the lora strength basically affects how it forces consistency between faces, clothing, hair color, etc. Higher strength gives better consistency but will usually degrade image quality.</li><li>More steps also seems to help consistency of sequence a little, at least with the Euler A sampler. It seems having more steps (like 30-40) and lower strength is a path to better image quality. However strength has much more effect than steps and it's probably a better use of GPU time to try two seeds at 0.8 and half the steps.</li><li>If you get an image you like, you can try keeping the seed+prompt and lower the lora strength to see how it goes. The "less strength + more steps" trick makes more sense for fine tuning an image than for spamming seeds.</li></ul>

Stripping / Undressing Sequence

Trained on a much bigger set with more variety, hopefully we're getting there.

Files with SHA256: 012cff4cd8c2f6c437e3b3d46531515c1a55597c1f79943ba176e47304c70b09