This LoRA was trained on multiple different cutout bodysuits, most of them inspired by the amazing designer Adam Saaks. Part of this is already in my darkgothicstuff LoRA, but I decided to create a separate LoRA for that and increase the dataset massively.

How to use:

The captions were done in natural language, this means you can prompt how the cutouts should look like, what is exposed and so on. For example use phrases like:

For halter style: e.g. shoulder straps, double shoulder straps, off-the-shoulder, strapless and so on.

For cutouts: geometric cutouts, web-like pattern, curved cutouts, extensive cutouts, triangular, angular, connecting straps and so on.

For cutout location and exposure: chest, waist, hips, midriff, sides, thighs, calves, abdomen and so on.

For the neckline: round neckline, scoop neckline, deep scoop, V-neckline, halter neckline, off-the-shoulder neckline, high neckline, and so on.

Always use the trigger cutout_bodysuit. You can also use the following advanced trigger words in addition:

holder type:

cutout_halter_neck

cutout_off_shoulder

cutout_shoulder_strap

cutout_strapless

sleeve design:

cutout_sleeveless

cutout_short_sleeve

cutout_long_sleeve

leg design:

cutout_high_cut

cutout_full_length (if you want "normal" legs, use full-length legs in the prompt. If you want cutouts here as well, use full-length cutout legs)

There is also a special trigger cutout_cropped that generates tops instead of bodysuits. However, this one is a little underrepresented since it was not the goal of this LoRA.

This is, just like my other clothing LoRAs, not a "single outfit LoRA". Instead, I trained how fabric works, falls, looks and behaves and where the cutouts are places and how much skin is exposed.

Would love if you post your creations here with it.

The LoRA was trained on the default flux1-dev model for maximum compatibility.

Dataset and training data:

• 182 images

• 7 steps per image

• 8 epochs

• bucks with 1152px max size, no auto scaling

• cosine scheduler with 0.15 warmup and 0.8 decay

• Captioning was done with OpenAI

• alpha == dim: 32

If you use it with other loras, for example a character lora, try to lower the weight - I had good results at around 0.5.

As always: I did not polish any of the images in the showcase, they all came out like this. I did 2 images per prompt and picked the better looking one.