I’m practicing more and more with the new FLUX model and something bothers me: I find that the faces of “random” girls are still very stereotypical like in fashion magazines, especially when it comes to lips which are always oversized. It looks like the girls have had botox injections for ten years.
I tried with prompts containing terms like “random girl, thin lips”,, and in the negative “oversized lips, botox, surgery, implants, silicon, fake”, but nothing works, I always get a clone of “emily ratajkowski”.
Has anyone ever managed to make a generation with these thin lips like Kirsten Dunst’s?
By the way, does anyone know why every generation, even without style, always adds “natural beauty” in the prompt? I think it distorts the action.
It’s all about you learning how to tell the AI what you want. I’ve started using the prompt [brain] button on many of the generators (usually to the right of the prompt, below the [dice] button), essentially the AI-Text-plugin at work in the background. It creates a prompt from a few keywords you put in. I’ve learned that (English not being my native language) it simply does a better job than I will ever be able to.
Using this generator, no-style, 6 images, with the [brain] button I filled in:
Portrait photo of a very casual Caucasian female, plain Jane, unassuming, fairly triangular face with wider than normal cheek bones, somewhat pointy chin, regular shaped thin lips, blonde medium length straight bob-cut hair, little inquisitive frown on her forehead, positively Kirsten Dunst’s younger sister.
And it generated this:
A very casual and unassuming Caucasian female with a touch of resemblance to a younger Kirsten Dunst. She has a fairly triangular face with distinctively wider than normal cheekbones that add a unique character to her visage. Her somewhat pointy chin is softened by the gentle curve of her jawline, and her regular shaped thin lips are parted slightly in a way that hints at a subdued smile or a contemplative thought. Her blonde hair is cut into a medium length bob, the straight strands grazing her chin and framing her face in a way that accentuates her delicate features. The most engaging element of her expression is the little inquisitive frown that furrows her forehead, suggesting a curiosity about the world around her. Her attire is simple and comfortable, blending into the background of the photo and allowing her natural beauty to take center stage. Her eyes, though not explicitly described in the prompt, one can’t help but imagine, hold a spark of intelligence and a gentle warmth that would be quite captivating. The overall effect is one of understated elegance and approachability, leaving the viewer with the sense that she is someone who is both easy to talk to and full of hidden depths.
Without making any changes, you will get ‘normal’ lips (next to the dreaded supermodel lips, btw). Casual Photo works too, while Cinematic and Professional Photo will bring back the supermodel lips.
It’s all a matter of fiddling and tweaking the prompt. Either by changing the [brain] input, or making slight changes to what it generated.
And I just noticed that Cursed Photo will yield a perfect Plain Jane
As for the lips, the only solution I found so far is to use a specific photographer’s name, and when the plugin tries to copy it, usually it helps with the lips. But it’s not always the case. Sometimes and more frequently recently, a new face is generated with normal lips, which gives me hope in the future it learns to do it. As for the prompt question, it depends on the generator you’re using. If you’re using your own, you have full control over the prompt. I suggest that you edit the one you’re using, find the “natural beauty” line and delete it, then save it. That’ll create a new generator specifically for you.
Are you able to share the generator you are using? Some generators, even no style is selected, would add additional text on the prompts. Might also depend on the prompt you are using. Some words have more weight, or the model is really just trained on images with said features.