Sheertex LoRAThis LoRA focuses on (semi-)transparent fabrics like mesh, lace, chiffon or tulle. It is trained on how they interact with light and the underlying body. It enhances the realistic behavior of sheer materials by improving diffusion, fabric density, edge definition, and visibility beneath the cloth. The LoRA is trained on lingerie and nightwear styles including sheer bras, mesh bodysuits, lace camisoles, chiffon nightdresses, and semi-transparent skirts and slips, in colors ranging from neutral beige and white to black, navy, and soft pastels.You get pretty good nipples with the default flux1-dev model; they are not perfect but still pretty realistic and I didn't want to overshoot the training goal. You can also pair it with NSFW models to get slightly better nipples but you have to take care of the weights to prevent the nipples sticking through the fabric (read further down for recommendations.)Information about the showcase pictures (all were done on default flux1-dev):<ul><li>The first 10 images were generated based on the captions of the dataset</li><li>Images 11-15 were generated with an open prompt to let flux decide</li><li>Images 16-20 are generated with the "mysticXXX" LoRA while mysticXXX hat 0.4 strength and sheertex 0.9</li></ul>If you use nipple LoRAs you have to lower them waaaaayyy down to something around 0.2 or 0.3 or else they will stick through the fabric. Avoid prompting for nipples or it will happen too. Pure nipple LoRAs have a high signal strength and will override sheertex if you go too high with the strength.Works with character LoRAs. I am using (SFW) character LoRAs at 0.95 strength, a NSFW LoRA (mysticXXX or a nipple LoRA) at 0.2-0.4 and sheertext with around 0.8.The LoRA was trained on the default flux1-dev model.Dataset and training data:<ul><li>173 images of different models and outfits</li><li>8 steps per image</li><li>10 epochs</li><li>cosine scheduler with 0.2 warmup and 0.8 decay</li><li>Captioning was done with joycap-batch</li><li>Buckets with mostly 1024x1536 images, sometimes 960x1536</li></ul>Polish<ul><li>24 images of close-up nipple shots through the fabric</li><li>8 steps per image</li><li>10 epochs</li><li>cosine scheduler with a very low learning rate</li><li>Buckets with 1024x1536 images</li></ul>As always: I did not polish any of the images in the showcase, they all came out like this.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.Q: Is that a single outfit or what...? A: No. I don't train single outfits because that makes not much sense to me. Instead, I teach flux the characteristics of different fabrics or garments (how it looks, how it falls, how reacts with light, what is visible through it and so on). This gives freedom and creativity in the prompt as long as the correct type of fabric is used.

Files with SHA256: 2835ee1875c1645172c0540356b5ad62943a79d565ab441450fdbbe4a2a86c58