ReVulva is a LoRA for Qwen Image 2.1 that’s meant to supplement the model’s deeper understanding of the female human anatomy as a concept, allowing for verbose and complex prompting while leaving its underlying comprehension of style and detail largely untouched. It was trained on a highly varied dataset of different subjects with completely different features, in an effort to show the model what a vulva can look like, so it doesn’t just copy/paste a single examples. Check out the attached images for a look at how to prompt with this LoRA.

—TLDR—

Strength: 0.8-1.2 (leave at 1.0 in most cases; ≤ 0.8 if combining with other LoRAs)
Samples: 30-40
CFG Scale: 3.0-5.0
Sampler: Your Choice(any that work with QI 2.1)
Scheduler: Your Choice(any that work with QI 2.1)

—Changes in v1.04—

• More detail and greater anatomical understanding

• Expanded dataset with complex, handcrafted captions (no AI employed for this part)

• Increased from 16 to 32 Linear/Alpha ranks (larger size; more weights)

• Longer training (~180 epochs/5,760 steps)

• Less style/detail drift from base Qwen

• Better prompt adherence

—Keywords Used in Training—

vulva, pussy, labia majora, labia minora, clitoral hood, clitoris, mons pubis, vaginal opening, anus, asshole, anal indentation, perineum, pubic hair, and many more...


—Do’s and Don’t’s—

❌ Don't prompt vagina
✅ Do prompt vulva or pussy

Why? Qwen knows that vagina refers to the internal anatomy only. If it tries to show you this on your subject, the results will likely be less than spectacular (and sometimes horrifying).

❌ Don't prompt anus or asshole
✅ Do prompt anal indentation

Why? Same reason as above. The model will try to pull the literal, internal anatomy outward, making it prominent to satisfy your prompting. Pretty sure this is a Qwen thing—my dataset captions actively tried to counteract this.

—Tips for Optimal Use—

• Responds well to complex prompting— be as descriptive as you like. Qwen’s VLM understands a lot more than just the trained keywords.

• Be specific (or not)— if you just prompt a picture of a woman’s vulva or a shot between a woman’s legs, the LoRA will still trigger, but you’ll get Qwen’s interpretation of “neutral” anatomical features based on my training (i.e. typically recessed or indistinct parts). This can actually be useful in some cases where the anatomy is not the primary focus.

• Reseed frequently— since the dataset is varied, re-rolling the seed can give you different feature combinations.

• more samples + higher resolution = more detail | higher CFG + specific language = greater adherence to your prompt— play around with this until you strike a balance that works for you. I’ve discovered that the right combination is highly subjective to the scene/subject that you’re trying to generate.

—Known Issues—

• Might occasionally go awry and produce something that makes you want to bleach your eyeballs—but hey, Qwen even struggles to get fingers and toes right a lot of the time.

• The weights favor certain aesthetics over others. I made my training data as well-rounded and detailed as I could, but Qwen came away with its own preferences. Very specific prompting can sometimes help push it outside the box.

• May not always depict all parts when it should (e.g. clitoris is missing). If this happens, and you are really attached to a particular seed, just mention clitoris or clitoral hood in your prompt to remind the model to include it, or positively prompt accurate vulvar anatomy or something similar. Alternately, you can negative prompt inaccurate vulvar anatomy, though the effect is usually a bit weaker.

• May infrequently produce an upside down vulva when subject is in an inverted position (e.g. bending over, lying prone). You can correct this by removing the word from your prompt altogether (e.g. a shot from behind of a woman's vulva turns into a shot from behind of a woman or a shot from behind of a woman's ass/buttocks—the LoRA weights should still come through) or by re-rolling the seed.

—Training Methodology/Rationale—

This LoRA was trained using Ostris AI-Toolkit, using the native BF16 version of Qwen Image 2.1 (though it performs similarly using the INT8 quant for generation). My approach was to let the model learn at a moderate pace while only allowing it to make changes to its weights in small increments. The result was a very long, gradual training run that I hope helped deliver a deeper understanding of the concepts I was trying to impart.

The end result was more or less satisfactory, though I didn’t pass along quite as much depth as I’d hoped for. I’m training on consumer hardware, so realistically, without a blank check to a data center API that can handle a much-expanded dataset, higher-resolution images, and longer training runs, I’ve pretty much capped out in any practical sense with my current setup.

My dataset images and captions were limited to anatomy only. No information about sexual positions, sex acts, or sexual language (aside from light slang) was included—not that I have a problem with any of those things, but that information just wasn't the focus. No individual likenesses were included in the training data.

I'm not ruling out a future update, though I don't have a firm timeline right now. I may take another few runs at this LoRA once my GPU has had a chance to cool down...

As always, enjoy and please generate responsibly. If you've read this far, thanks for downloading!