I've created a LoRA called "NSFW Diversity" that's focused on bringing genuine diversity and realism into NSFW image generation. Honestly, I started off just wanting flux to behave better—to stop giving me those weird, unnatural faces and limited looks. So I did a ton of experimenting and research, tweaking settings, and fine-tuning. Turns out, by working to create truly diverse NSFW concepts, fixing the faces came naturally as a bonus.Right now, it's still in alpha. I messed up the initial training by adding new concepts during the workflow (lesson learned), so it's a bit rough around the edges—especially when it comes to full nudity, which it doesn’t quite handle well yet. But it already adds a bunch of different NSFW concepts, which is what I was aiming for at this stage.It blends amazingly with other realism-focused LoRAs, so feel free to mix, match, and experiment to your heart's content. This is definitely a "play around and find out" kind of tool—see what works best for your creative process!If you end up liking the lora and want to see more improvements faster, consider supporting me here:https://ko-fi.com/V7V21CHE9DYour support genuinely means a lot. Compute costs add up quickly, and there's only so much I can invest into tuning each month. Every bit helps keep this going! And yes, i let gpt rewrite my initial description, i suck at these. Due to me beeing an idiot and deciding to add concepts mid training, the model might sometimes mix concepts in weird ways, especially when you use the word "butt". It works, but sometimes you need to do an extra gen. It works best with Euler + beta and 32 steps, but sgm_uniform + dpmpp_2m also works very well. You can combine it with the flux realism lora, although my end goal is to create a high quality standalone lora, the realism lora can give an extra kick to realism if you want.

Files with SHA256: 6ecd7ee550078d86b2f042ceb38369ab231a692652fda1920226977ffd9014fa