A simple trained model a muscular male photos, the image is labeled using Joy Captioning so it was trained on a somewhat long text. This model might generated bad anatomy sometime, but the text capability is still preserved (well, mostly)
This model aim to add concept of male genitalia that seems to be missing from the base model completely. But I am not sure if they works with a thinner man
CFG scale range between is 2.5 - 4.0, with 20-25 diffusion step is okay, but more seems to give plastic skin. Tested using ComfyUI with non-quantized flux, fp16 t5_xxl, and "Flux (C64 Lora)" from openart.ai
Experimentation Notes
Using ComfyUI Inpire Pack > Lora Loader (Block Weight) with only last layer mask
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1, is the only real layer that encode the concept of penis. This will preserve almost all the flux capability, and decouple the muscle-bear body type
Description
Sample trained model, so don't expect much. The trained image mostly has only one subject, so if there is more, the generated image going to get wonky
FAQ
Comments (3)
Although it lacks in photorealism a bit and does tend to push toward bigger guys, I can confirm that it will do slimmer guys with some massaging and that the penises are probably the best I've seen in flux so far--though still not perfect.
Kudos to you for creating this! It gives me hope that we'll see more LoRAs and flux checkpoints with good, diverse men and their anatomy.
The genitals took a while before it was usable (~2k steps), but my current dataset has somewhat limited pose, and gear towards the bigger guy, mostly because I don't have the dataset on the skinnier one. It's also not as photo-realistic, but I think post-processing from Pony Realism could alleviate that. At least for now, it should be a good-enough lora for the missing concept
Also I'm testing the model myself as well. Here are some observation, generating at lower resolution (512) seems to be more stable than generate large image (1024) directly. And even though the genitals had been form, it occasionally got doubling / twinning which I don't know how to address that yet
See experiment section
Using ComfyUI Inpire Pack > Lora Loader (Block Weight) with only last layer mask 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1, is the only real layer that encode the concept of penis. This will preserve almost all the flux capability, and decouple the muscle-bear body type
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.