The albedo dress (also known as the twitter dress, the chain dress, or just That Dress) is a backless halterneck fishtail evening dress with a trident halter straps, cutouts over the hips, and gilt chains in order to accentuate the hip cutouts. It is known as the Albedo Dress because the version that went viral in late 2023/early 2024 was made by dubious online retailers to be a legally distinct version of a dress warn by Albedo from the light novel series Overlord. While 'chain dress' and 'fishtail dress' are a bit more generic, they also resulted in more dataset clashes (dresses made of chainmail or giving non-piscine girls fish tails), so 'albedo dress' seemed to work most reliably with the fewest disambiguation clashes.This LoRA is designed for PonyDiffusion version 6. To use it, use the normal PDXL v6 prefix, then some variant of the following phrase:score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, black albedo dress with gold chainsFeel free to play with the colors in the tag. The two dress colors with the best results are the ones that were actually for sale: "black albedo dress with gold chains" and "red albedo dress with gold chains". For dress color, the colors white, blue, grey, and purple work reliably. For chain color, gold, silver, and white work reliably. It is possible to request 'no' chains as a color to just have a trident halterneck fishtail evening dress with hip cutouts, but it tends to remove the hip cutouts entirely when it works.While the albedo dress that was trending on Twitter featured red or black flowers in the bustline, only a small percentage of images in the dataset actually had the bustline flowers - more images have flowers in the subject's hand, hair, or in the background! You can try with the "flower" tag, but depicting it with the flower will probably require careful inpainting.Strength 0.9 seems to give the best results. Strength 1.0 may give vertical line artifacts in the background. Strength 0.8 and lower often gives generic fishtail dresses.LoRA created with 1230 images over 10 epochs. The trainer crashed so, so many times. While some 3D and catalog images were present in the training set, this LoRA is not suitable for creating images of 3D or photo style, and is heavily biased towards solo characters in simple, abstract or gradient backgrounds.
