What This LoRA Does
ReVulva is aimed at bringing a degree of realism to Qwen Image 2.1’s concept of the female reproductive anatomy. If you’ve ever generated on the base model without any LoRAs, you’ve likely witnessed the horrifying alien lobster monster that Qwen diffuses in place of the proper female form. These extreme weights were difficult to overcome, but I think this first attempt does a decent job of correcting Qwen’s overblown (and sometimes terrifying) interpretation.
I trained this LoRA on a dataset of 40 images (publicly available via Google Image search) with any likenesses and personally identifiable information stripped away before training. No information regarding sex acts or sexual positioning was included, as I only intended this to act as an anatomical supplement. It performs well across a variety of contexts, subjects and settings, and Qwen appears to intuit sexual context fairly well without being explicitly told to do so.
My dataset captions use a mixture of highly detailed anatomical terms and slang language, which Qwen interprets interchangeably. My aim was to inform the model about the specifics of the female vulvar anatomy and how the various parts relate to one another, though the model can still make mistakes at times. For instance, it can get confused about the location and appearance of the clitoris—which, ironically, is a struggle that men everywhere have grappled with for generations.
ReVulva functions similarly with text-to-image and image-edit workflows, and prompts identically between applications. I’ve noticed no difference in quality or understanding between the two.
Samplers, Schedulers & Prompting
I’ve tested it on a range of samplers, and it performs similarly to the base model, in most cases. I personally prefer either euler or er_sde, paired with the beta57 scheduler, but it will work on whatever you prefer—I’ve also tried a few res samplers, which handle things adequately too.
The LoRA was trained slow and long (about 3,600 steps for this version) partly in an attempt to avoid overfitting and style drift, and to allow the model to absorb the finer details. Some differences in composition, subject pose, background, and color/lighting are inevitable, but many of these hindrances are easily overcome with positive or negative prompting. For instance, since the model was trained on a dataset of highly realistic images, it may tend toward more realistic lighting or a more natural color palette. You can counteract this to some degree by positive prompting things like “high contrast” or “vivid colors.”
As I said, the training effects on style and composition in my testing aren’t super apparent, but if they annoy you, you can certainly prompt however you like and the model appears to compensate. I’d highly recommend that you check out my prompting in the example images, though it’s not required for the LoRA to function correctly.
Strength and Settings
Most of my testing was done at a LoRA strength of 1.0, so you can keep it there for most general applications. I’d recommend a range between 0.8-1.2 if using it as a standalone. Though I haven’t done extensive experimentation with other LoRAs, I’d recommend peeling back the strength to 0.4-0.6 if you’re using in tandem with another LoRA that modifies aspects of the female anatomy—particularly the pelvic region.
My best results—though I tend to use verbose prompting—were achieved with a CFG scale of 3.0-4.5 and a number of samples around 30-40, depending on the level of detail in the specific prompt/scene. My workflows are pretty basic though, so you might discover a different combination works better for you.
Again, since my LoRA weights are essentially trying to overpower the transformer’s overblown and scary impression of female anatomy, raising or lowering the strength of the LoRA can have a pronounced effect, as the base model’s weights are either allowed to leak through a bit more or suppressed further.
Parting Notes
ReVulva is my first attempt at LoRA training, and the process was super laborious. I went through 3-4 dozen full training runs (on consumer hardware) before I attained a result that I wasn’t completely dissatisfied with. Even so, there’s always room for improvement, and I may push this toward a full v1.0 release if there’s enough interest or a generally positive reception.
This LoRA was trained on a fairly diverse sampling of subjects with widely varied body types and anatomy. While I hope that choice helped inform the model on the many variations in the female form, it may have served to confuse it to some degree. As always, if you encounter any issues, re-roll the seed and see what happens. Adjusting the number of samples may solve some problems too.
While I’ve only sought to fill a gap in Qwen Image 2.1’s knowledge of the human form, I’m not promoting or encouraging any specific use case for my LoRA by uploading it, and I claim no responsibility for how you choose to employ the results of my work. As always, please be responsible with your generations, and thank you for downloading!