AndroFlux is a LoRA model made for FLUX.1 dev by Black Forest Labs, because there’s a glaring lack of accurate male anatomy in AI models. I’m trying to fix that by making sure AI can depict male features more accurately, balancing out the overwhelming female-centric content you see everywhere.
Community and Support
Join The Bulge Discord at https://thebulge.xyz if you’re into gay AI art—it’s a great place to share and learn. If you want to support the project, whether with feedback or donations, it’s much appreciated.
Current Status
The model isn’t perfect, so don’t expect flawless results.
Purpose
AndroFlux is all about making sure AI models can represent male anatomy—especially penises—accurately. With that said, it also has a general photographic aesthetic and focuses on twinks so it is still useful for other SFW purposes. It’s a step towards more balanced AI-generated art.
Technical Details
AndroFlux was built and tested primarily on FLUX.1 dev, but it should also be compatible with the Schnell base models, including the quantized fp8 version. These are distilled models, making training tricky and affecting the LoRA’s performance, requiring a lot of optimization.
Training Process
Resolution Journey: I started training at 1024px, but the model wasn’t converging, so I dropped to 512px. The quality wasn’t great, so I ended up at 768px. This resolution offers a good balance between detail and training time.
Captioning: Captioning was crucial in training. Contributions came from the AndroAtlas dataset, part of the AndroFlow project. Ephemeron01 from The Bulge Discord assisted with captions using models like llava-hf/llava-v1.6-mistral-7b-hf and a custom-trained PaliGemma VLM. Specific prompts like "man, penis" and the activator "AndroFlux" were also used in later versions.
Acknowledgments: Huge thanks to bghira and Terminus Research Group for their support, including providing an A100 for some of the training runs.
Usage Tips
For best results, use prompts that specifically mention male frontal nudity, such as "penis."
Recommended settings:
Resolution: 1024x1024 or 832x1216
Flux Guidance: 3.5
LoRA Weight: Start at 1.0, reduce to 0.5 if artifacts appear
CFG: 2.0 (1.0 will be faster though)
Suggested Prompts
A nude man in his {late teens | early twenties | early thirties}
His penis is {erect | flaccid}
His penis is {large | small | average}-sized
Version History
See "About this version" on the right for version-specific updates.
Description
You will see nightmare fuel and overly long dicks. You have been warned.
Full retrain, works best when used in in combination with v15:
v17 at strength 1.0
v15 at strength 0.5
Use a CFG (not FluxGuidance) of 1.0 (faster) or 2.0 to increase contrast/coherency and introduce negative prompting, YMMV
This was the last run on the A100, so updates might be a little slower from here on out. In the meantime, I'm working on making the dataset more accurate for the erect/flaccid distinction.
Same as last time still applies: You have to cherry pick them and it doesn't seem to listen to "flaccid" or "erect" but the generation still benefits from including one of them in your prompt.
FAQ
Comments (10)
This is amazing. Works so well already!
Hey! Saw in the update that you were looking into improving the dataset so I thought i'd give some possible pointers I've discovered through my own training:
1. Quality is more important that quantity with Flux datasets. Having 50 high quality images with a large variety for the model to train off seems to produce better results than 1000 similar images of mediocre quality
2. Giving variety to the training prompts seems to produce better results. Rather than using a repeated trigger phrase like "his penis is [x]", having a variety of prompts like "he has a [x] penis", "his [x] penis being visible", "showcasing his penis, which is very [x]" and so on seems to produce far better results. You can modify your existing prompts with a simple python script that replaces instances in your existing prompt files with a list of these sorts of alternatives at random to achieve this without too much effort
3. Regularization images seem to be incredibly important for Flux to better train the model on your concept and prevent bleeding/overfitting, which your model seems to be doing since it will add penises to nude women unprompted. It should also allow the model to generate penises using styles separate from the training data, which could be important since I've seen people mention that your model seems to force a realistic style. From what I've seen, regularization images should consist of a variety of different types of the class you're training (since you are training people with penises, regularization images should primarily consist of people of both sexes, various races, ages, hair colors, clothing, image styles, poses, etc.). These regularization images should make up 20-50% of your dataset.
Since regularization images should consist of images and prompts that are as close to the original model as possible, and because captioning can be quite tedious, I've actually made a workflow for ComfyUI that will generate 1024x1024 images using Flux and save a corresponding .txt file with the prompt specified alongside it. The workflow also contains a collection of regularization images I've already generated, so you can just drag and drop them into your existing dataset folder without issue. It can be found on my profile or by searching "Regularization Image Workflow & Collection for FLUX".
Thank you so much for the detailed tips! I will definitely be making changes to the captions to make them more flexible.
Regularization images are already in the dataset, but to the point of it adding penises to women, that's not exactly an undesired behavior. If you were around for the SD1.5 penis scene then you'll be familiar with the T-girls LoRA, which had a similar approach.
Ultimately it's not meant to be a porn model where there are advanced interactions between men and women in case that's what you're trying to do. It add penises, so if you don't want a penis, you should obviously not use the LoRA. I like to think of it as akin to what something like AndroGel does (to some extent, not entirely). Hence the name, AndroFlux. Penises are for everyone :)
I'll give your workflow a try though and see if we can make this model even better! I genuinely appreciate your efforts to help me get this right :)
@markury oh yeah, definitely not suggesting that the ability to add penises to women should be removed; it's more so that the LoRA seems to try to add male genitalia to any nude character unprompted.
As an example, if I load up 2 LoRAs, one for male nudity and one for female nudity, and I give a prompt like "the man on the right has a big penis, whilst the girl on the left has a pussy," it doesn't seem like FLUX will attempt to give the male character female genitalia, but it will give the female character male genitalia, which seems like it could limit the viability of your model for a lot of people going forward. Just something to be aware of!
@TheGreatOne321 Ah I see what you mean! That's a valid concern. It's important to note that a lot of LoRAs right now are being trained with Ostris' AI Toolkit which had a huge bug (which Ostris knew about but just ignored anyone who brought it up... rude!) So steer clear of those when mixing with my LoRA, from both a technical and ethical aspect. Also avoid LoRAs that introduce guidance scale, as I worked really hard not to break the distillation on the model and using a LoRA that doesn't alongside it will cause all sorts of problems. Just more things to be aware of that cause issues when mixing LoRAs with Flux.
But yeah, I'll see if I can work on that issue. More regularization will probably help as you mentioned. Most of the regularization I do now is for asses and body type so we don't get dickbutts or twinks every time. Cheers again for being so helpful and active, your kindness goes a long way :)
Mah dude - I can't use Flux yet but I absolutely love what you're doing here. It's some quality work. Always impressed by your stuff!
markury is on Tensor.art to and you can use his flux lora and his bananadiffusion base model on Tensor.
Do you have any guidance on prompting for everyday body types? I'm having trouble getting things outside of twink, twunk, and muscled. I'm still getting used to prompting in Flux and I know it skews toward magazine-quality images so could be me or model.
Thanks overall! This project is nice and got me into Flux finally.
Describing them as fat, chubby, or fat and chubby, out of shape, beer bellied, etc. has been giving me more regular looking men.
Flux is really good at following descriptions so if you want a specific body type, just describe the details to get rid of what you don't like.
Using the realism lora also helps a lot. And specifying a specific age or age range.
I also wanted to write add "potbelly" to your prompts :D No seriously, it works. Some ideas to add to your prompt:
potbelly/beer belly
receding hairline
Average muscle tone
Man boobs
Slight paunch
Thin legs
patchy facial hair
Short stature
Dad bod
poor posture
Wide waist
Skin blemishes / age spots
...
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.