Summarized by AI based on all the captioning .txt files:This LoRA is trained on a wide range of partially removed and casually open clothing focused on adult female models, emphasizing outfits where garments are lifted, unbuttoned, draped, or unzipped to reveal the body while maintaining a natural, photographic style. It covers open robes, velvet capes, loose knit sweaters, cardigans, plaid shirts, denim vests, leather jackets, trench coats, and fur-lined coats, as well as lingerie elements like lace bras, bralettes, panties, bodysuits, and mesh or satin robes layered over the nude body. Many images feature everyday garments worn in an undressed or in-between state—off-shoulder blouses, lifted turtlenecks, open bathrobes, unbuttoned shirts, and skirts or shorts pulled up or aside—often with bare breasts visible and, in some cases, full nudity beneath the top layer. Different models were used with different breast sizes. Materials include velvet, silk, satin, lace, mesh, cotton knits, denim, leather, faux fur, and wool, with colors ranging from earthy neutrals and cozy plaids to bold reds, deep greens, and blacks. The dataset also frequently incorporates accessories such as scarves, shawls, blankets, cardigans, hats, and glasses, and occasionally high heels or boots, situating the models in domestic interiors, rustic outdoor locations, or stylized lifestyle scenes. Overall, the style ranges from intimate casual undressing (sweaters, robes, shirts, and blankets slipping open) to more stylized outfits like leather jackets with short skirts or open coats, consistently highlighting the contrast between fabric texture and exposed skin in a sensual but photographic, non-cartoon aesthetic.DISCLAIMER (Do read this!): This is NOT a breasts/nipple LoRA! This LoRA focuses solely on "how does clothing look if it's partially removed?" because I found that flux - even the NSFW models - have issues with that, similar to the "full glass of wine" issue. Maybe you will be able to generate good nipples with that, then it's fine but there are NOT enough steps to have perfect nipples trained. I HIGHLY advice you to use other LoRAs along like mysticXXX (What I did in the showcase), a NSFW model (Just make sure that it's close to the original flux1-dev model, for example the NSFW MASTER FLUX) or use inpaint after generation.The LoRA was trained on the default flux1-dev model for maximum compatibility.Dataset and training data:<ul><li>436 images with around 150-160 different outfits (maximum of 3 pictures per outfit)</li><li>2 steps per image</li><li>10 epochs</li><li>buckets with mostly 832x1280 px</li><li>cosine scheduler with 0.2 warmup and 0.8 decay</li><li>Captioning was done with joycap-batch</li><li>alpha == dim: 32</li><li>Trained on a 5090</li></ul>If you use it with other loras, for example a character lora, try to lower the weight a little - I had good results at around 0.6 - 0.8.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.<h3 id="faq">FAQ</h3>Q: Nipples look terrible! WTF?A: You probably didn't read the part that says "Do read this!". The captioning focused on how clothing and garments look when half-dressed or open. Besides that, nipples need a lot more steps.Q: 2 steps per image and only 10 epochs? So every image was seen only 20 times?A: Yes because flux already knows all the garments and materials. The only thing flux had to learn was "how does it look when it's partially removed? What body parts are visible? How do different fabrics behave?". Plus there were 436 images in the dataset, so flux learned that concept pretty quickly.Q: Why is this LoRA so big?A: Because I have used network_dim and network_alpha == 32 for best qualityQ: I am running out of memory when using this LoRA!A: You're using SwarmUI or an A1111 fork, that's common due to how they handle TE1 weights. Use ComfyUI or lower the TE1 weights of this LoRA.

Files with SHA256: 03b57a22f7636dc30de370d34c15b0de9bee43b813e5b948b48db5e3f07cc2dc