[Experimental!!! Read Problems section to see limitations]If you have problems with this check out this alternative by @Gentlemanbunny : <a target="_blank" rel="ugc" href="https://civarchive.com/models/772191/blacked-bleached-whatever-underwear-read-description">https://civarchive.com/models/772191/blacked-bleached-whatever-underwear-read-description</a>Generate Images of women wearing BLEACHED underwear. Ask Flux for 'woman/character wearing BleachedWear' and maybe 'wearing underwear that says "Bleached" on it'.V3 Update:<ul><li>less nipples</li><li>can now do black or white underwear</li><li>higher lora rank =&gt; higher file size (not sure this actually helped that much)</li><li>text garbled at the edges quite often</li><li>plays nice with some loras and absolutely not with others, not sure what thats about</li></ul>Have fun!<h3 id="problems:-8toya0cmy">Problems:</h3><ol><li><s>Nipples poking through: You might have noticed the nipples staring at you, this is consistently a problem.</s><s>Solution: There is a lot of uncaptioned nipple poking in the dataset, captioning that and retraining should fix it.</s></li><li>Style Control: The LoRA overwhelmingly tends towards a 3d render faux realism. This is because the majority of the dataset is in that style.Solution: ¯\_(ツ)_/¯ availability of other images is limited, i'll see what i can do. V3 somewhat more flexible, but not that much</li><li>Text quality: It's hit or miss. Adding 'wearing underwear that says "Bleached" on it' to the prompt seems to help. Maybe try higher step counts.Solution: ¯\_(ツ)_/¯ I can try: a higher Rank LoRa, training at a higher resolution (currently 512x512), Longer Training. All of these require more GPU time, and I am not made of VRAM.</li></ol>

Files with SHA256: 6c627e077f4264f626585df8cba4a4f8dadb79abf996891f7cd61ef1cb7d316e