MASC - Realistic Masculine Hunks and Men
Photorealistic, diverse men for Flux, with a focus on detailed faces and close-ups. A collaboration between Divergent AI and Gen Hunks.
Base model: Flux.1 Dev (gallery made with fp8)
Trigger words: none. Trained on descriptive natural-language captions.
Strength: 1.0. If you see anatomy issues, add steps before lowering strength.
Steps: 20 for good quality, 40–50 for maximum realism
Resolution: tested at 1024px in various aspect ratios
Prompting: describe it like a photographer would. Cover the subject (pose, expression, clothing), the environment (place, time, weather), the camera (shot type, lens, depth of field), the atmosphere (light, palette, mood) and the finish (for example "cinematic film still").
Example: An ultra-photorealistic, cinematic film still of a handsome man in his mid-30s. He is wearing a tailored navy blue linen suit with an open collar white shirt. He stands confidently on a modern minimalist rooftop terrace, looking slightly off-camera with a relaxed yet assured expression. The background features a vibrant city skyline at sunset, rendered with a shallow depth of field typical of an 85mm lens. The scene is bathed in the warm glow of golden hour light, casting long, soft shadows.
Dataset: 300+ curated editorial, social and artistic semi-clothed portraits at 512/768/1024. No frontal nudity in the dataset. It also works for surreal and artistic scenes.
Description
Full overhaul of the v0.3 prototype. Multi-resolution training (512/768/1024) gives more detail and more consistent faces.
FAQ
Comments (2)
Why is this LoRA so immense and what does it do?
TLDR:
I went with a rank 32, alpha 32 LoRA because of the database size, what my research informed me to try, and for fidelity.
I'm open to trying a rank 16 to see the downsides, if any.
My LoRAs feel more like full fine-tunes due to their capability outside of their training data, producing stunning and realistic imagery. A full style LoRa.
I'm training an NSFW model and would love to chat about LoRA training sometime!
Hey Rudy, thank for your comment man!
Due to the database size and for what I was going for. I wanted to go for a LoRA rank and alpha. Also due to some research I had done. Also for fidelity reasons.
I'm still learning and more than willing to try do a rank 16 and see what the downsides are (for what I'm going for) and in terms of outputs. I've trained on it before, but I've trained so many hours (and failures) since then. This model on rank 32, alpha 32 is just what I landed on that I liked enough to put out. Really figuring it out as I go.
I find it a very capable model, but I can understand for those with VRAM constraints it but be unnecessary. Also seeing what you've achieved with what I assume is rank 16 alpha 16, seems it could be capable enough.
Although I do find my LoRA are highly capable outside of their training data, they feel more akin to a full fine-tune as opposed to basic LoRA. Been able to generate some stunning surreal imagery and other stuff outside of it's dataset and it still generates it with great realism, creativity and prompt adherence, most times. I'm actually currently trying to train a NSFW model. Mind if I hit you up in messages to chat about LoRA training sometime? Sorry I'll TLDR my wordblah.