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
v14 @ 4500 steps
The training was interrupted so I don't have the exact config :/
But it seems to work well with AdaptiveGuider with 2 steps skipped and CFG 2 (I did not train with CFG)
Results look a lot like fingers but still progress.
FAQ
Comments (18)
Please put this lora on Tensor.art
I just reuploaded it over there but I'm not sure if they allow you to use Flux loras yet :(
@markury I'll check to see if they allow loras and if not hopefully they do soon
@markury Couldn't find it on Tensor, do you have the link.
@markury Thanks but something is wrong with it. It says this for some reason, (The model deployment is abnormal, please re-upload/contact customer service).
@Gentileschi Try this version https://tensor.art/models/761303487725311854/AndroFlux-v15
@markury Well that message didn't pop up this time so I'll try to use it tomorrow. Thanks.
Do you caption all those images that they are actually photos? It seems like the model loses the ability to draw other styles, when this lora is applied. Also, does it work only with Clip, no T5 (or whichever Flux uses by default)?
No Flux LoRAs train either of the text encoders, Clip-L or T5xxl, so you can't use the normal LoRA loader. You'll have to use LoraLoaderModelOnly, which bypasses the text encoder pipeline. This is a photorealistic LoRA because of the way Flux learns, as it inherits style and artifacts before concepts. Multiple captions were used per image so it's hard to say which include the photo tag. It's also trained for much much much longer than other LoRAs so you really can't tone down the realism unfortunately.
@markury I'd suggest to ensure that the style (photo, dslr, movie, cinematic - whatever) is properly captured to minimize the effect of losing other styles. Otherwise, if it sees "dick" only on photos that are not even captured as photos, it will "think" that it needs to trigger that specific style when it sees "dick".
Amazing improvement comparing the current lora with the one from 2 days ago. Very nice job!
>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.
The Bulge Discord? There really is a discord for everything, haha
Amazing work, I need to try using 512p instead of 1024p for the training data. What kind of learning rate are we talking about?
Thanks and good luck for future training!
Haha yeah that's my discord. It's based on the gay bar called "The Bulge" in Parks and Recreation. And thank you for the kind words. Currently my LR is 3e-4 but it does make the model blow up after a few thousand steps, so I might go lower.
@markury I'm currently also trying out 512p at batch size 8 with a huge dataset of 50k nsfw images, and currently running at 6e-5. In my test runs over 1e-4 it either burns up your model or it never really converges and fluctuates between two local minimas it seem.... 6e-5 is slow as fuck tho....
@pyros_sd_models What gpu and trainer are you using? May the gods have mercy on your soul if you're using Ostris or XFlux.
Nice, you're making huge steps. It's already quite well working, I'm getting some really nice results. Flaccid ones are quite hard to archive tho but still, .. even some 'interactions' are possible and looking good.
Hi, could you tell which interactions and prompts are working for you please? Ian having trouble to achieve any
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.