HandBraV2NF will produce best results if you use the natural language phrases below rather than individual tag-style words. It is primarily intended for realistic photos but I did include some Anime and cartoons in the training data so it seems to work pretty well on those too.Flux appears to be very smart about hands vs. fists and I was careful to caption accordingly, so the Lora does a fairly good job at differentiating them. I have included examples of the same prompt and seed, with the only difference being hands or fists and you can see that it followed the instruction while the rest of the image remained very consistent.Usage: Weight of 1 seems to work well most of the time. This LoRA does tend to produce enormous breasts but you can try turning the weight down slightly and/or add the preferred size to the prompt like small breasts, large breasts, huge breasts. Include the main trigger word handbra and then you can add the natural language phrases below to direct the hand positions:<ul><li>covering her nipples with her hands</li><li>covering her nipples with her fists</li><li>covering her nipples with her arms (this one is still a bit unreliable)</li></ul>I've tested the LoRA with several Flux variants in Forge, SwarmUI, and ComfyUI:<ul><li>flux1-dev-fp8</li><li>flux1-dev-fp16</li><li>flux1-schnel - you may need to play with the weight.</li><li>NF4 gave me trouble and I am still working on this one.</li></ul>These resolutions have been tested but it may work with others as well:<ul><li>832 x 1216</li><li>896 x 1152</li><li>1024 x 1024</li></ul>Other notes:Training was about 3 hours on an RTX 4090, for 4000 steps on 40 images, Learning Rate 2e-4, and very detailed natural language captions.I included the words "holding her own breasts" in every caption to help avoid ghost hands from appearing. The concept seems to have stuck so it should not be needed in the prompt, but you can include it if you see extra hands holding the breasts.NF stands for No Faces. I had issues in the original V1 that kept using faces from the training data. In V2 I made adjustments to avoid this.Have fun and I'll happily take feedback to incorporate into my next version.

Files with SHA256: c9f1dfa29cadc7876f4e04d88dae20699525be74d09a1bbf50cb9c01bf42c3ac