Too Close to Real? or The Celebrity Face Problem

[Malice]

Hey freaks,

Quick feedback question.

I’ve been refining my base models for Harley, Wonder Woman, and Batgirl, and I wanted some outside opinions before I push things further. Do any of the faces feel too close to a recognizable real person or celebrity to you?

Harley with her makeup


Harley with no makeup


For Wonder Woman, I’m honestly pretty happy with where she’s at already, but I still want to make sure she feels original enough.



And for Batgirl: the base is an AI model from Mage, so there’s no issue there, but I’m debating freckles versus no freckles. AI freckles can sometimes be inconsistent from image to image, so I’m curious whether that actually bothers people visually, or if it makes the face feel more natural and human instead.



At the end of the day, I honestly don’t mind playing around with licensed characters for fan fiction.

What I don’t really want to mess around with, though, especially now that AI is finally getting good at it, is generating faces that look too close to celebrities.

A few months ago, character consistency with original AI faces was an absolute nightmare. If you tried making recurring characters that weren’t heavily inspired by real actresses, the results were usually all over the place.

Now it’s finally becoming possible.

It’s still a pain in the ass: a lot of models keep reverting back to Margot Robbie’s face for Harley and Gal Gadot’s for Wonder Woman, but major improvements have definitely been made on that front.

Curious to hear your thoughts.

[sneakly]

Hey freaks,

Quick feedback question.

I’ve been refining my base models for Harley, Wonder Woman, and Batgirl, and I wanted some outside opinions before I push things further. Do any of the faces feel too close to a recognizable real person or celebrity to you?

Harley with her makeup


Harley with no makeup


For Wonder Woman, I’m honestly pretty happy with where she’s at already, but I still want to make sure she feels original enough.



And for Batgirl: the base is an AI model from Mage, so there’s no issue there, but I’m debating freckles versus no freckles. AI freckles can sometimes be inconsistent from image to image, so I’m curious whether that actually bothers people visually, or if it makes the face feel more natural and human instead.



At the end of the day, I honestly don’t mind playing around with licensed characters for fan fiction.

What I don’t really want to mess around with, though, especially now that AI is finally getting good at it, is generating faces that look too close to celebrities.

A few months ago, character consistency with original AI faces was an absolute nightmare. If you tried making recurring characters that weren’t heavily inspired by real actresses, the results were usually all over the place.

Now it’s finally becoming possible.

It’s still a pain in the ass: a lot of models keep reverting back to Margot Robbie’s face for Harley and Gal Gadot’s for Wonder Woman, but major improvements have definitely been made on that front.

Curious to hear your thoughts.

I am not much with AI, but I think the problem you face is about how the machines render a fake. If you go through a lot of images of celebrities on DA or any other AI swamped platform, celebrities faces look uncannily like the source, until they are done at an odd angle, Suddenly they don't look like them at all. AI has been modeling Harley from images of Robbie in make-up. It is copying the look, without thinking, but it has learned from thousands of still images, most of them Robbie's or Gadot's WW. That is also why most of AI nudes have big fake breasts, it learned to draw women based on thousands of big fake porn titties and shaved pussies and not real boobs and a full bush. It simply is unaware their is a difference.

An interesting experiment would be to specify another celebrity as the face and see what happens. Liam Neeson as Harlrey or Julia Roberts or Zendya may give a more unique take on the character. The problem for trying to make your own is the statistical size of the population is one.

Trying to get the computer to generate a unique face is probably really hard because it will always default to what it knows and the limit of your ability to give it details. How many ways can you describe a woman's nose or chin or eyes? If you create a unique model, how well can the AI maintain that model as a reference?

Definitely not an expert....

[Opaque]

Do you train your own LoRAs?

[Malice]

Do you train your own LoRAs?

No actually, I don’t train my own LoRAs, but I use Mage so I'd be able to import one.

Have you trained your own? And if yes, does it solve the consistency issue long term?

[Malice]

Liam Neeson as Harlrey

[Opaque]

Do you train your own LoRAs?

No actually, I don’t train my own LoRAs, but I use Mage so I'd be able to import one.

Have you trained your own? And if yes, does it solve the consistency issue long term?

I have not trained my own LoRAs, but as far as consistency goes that's how you solve those issues. I couldn't see if Mage allows training or if you have to do it somewhere else first(like at Civitai), to then use on Mage.
In certain circumstances it might be overkill as AIs are getting really good at using reference images, but LoRAs gives the most consistent results. Face, costume, body type can all remain consistent. Definitely something to look into and consider. Rather than simply relying just on AI model's ability to translate reference images every time you generate.

[Malice]

I have not trained my own LoRAs, but as far as consistency goes that's how you solve those issues. I couldn't see if Mage allows training or if you have to do it somewhere else first (like at Civitai) and then use it on Mage.

In certain circumstances it might be overkill, as AIs are getting really good at using reference images, but LoRAs give the most consistent results. Face, costume, body type — all can remain consistent. Definitely something to look into and consider rather than relying solely on the AI model's ability to translate reference images every time you generate.

Thanks for the feedback honestly, I really appreciate it.

I think where I’m hesitating is that my project is stylistically all over the place by design

For example, the first image I’m sharing is inspired by Gustave Doré’s “The Fall of the Rebel Angels,” but I also have pieces inspired by romanticism, symbolism, darker styles, cinematic shots, comic-book panels, etc. Some images are very movie-like, while others lean heavily into pastiche compositions.



There's also an issue with multi-character scenes. I'm not sure how good LoRas would be for that especially.

So part of me worries that if I train LoRAs poorly (which realistically will probably happen at first), I might accidentally flatten the flexibility I currently have with references and prompting.

So as things stand right now, do you think the faces are already far enough away from recognizable celebrities, or would you personally still push the designs a bit further on that front?

Just to illustrate the range a bit more, here’s a more cinematic scene:



And this one below is more of a dream/horror sequence:



The thing is, my images are meant to illustrate a story on the platform, and the story itself is all over the place stylistically, dream sequences, combat scenes, horror elements, and… let’s just say other kinds of moments

So my concern with relying too heavily on LoRAs is that they might ultimately create as many problems as they solve.

I’m constantly jumping between different artistic directions and moods. Sometimes I even feel that approach might work against me and cost me readers because it’s less visually uniform. But at the end of the day, I’m making the kind of story I personally would want to read, alongside the images and videos that bring my vision to life as closely as possible.

Hell, even the first image I shared is still miles away from what I originally envisioned, but considering the sheer number of generations it took to finally reach something close to the atmosphere I wanted, I still consider it a success.

[jca5646]

What platform or generator are you using to render your images? Generated locally on your PC or via a web based generator, and if so, which image generator?

[Malice]

All browser based, I don’t generate locally on my PC.

I don’t really use just one. I jump between a bunch depending on what I need at the time.

Chat GPT


Grok


Krea


Mage


A combination of several




Some are better for edits, some for style consistency, some for creation of scenes, so I just swap depending on the task instead of relying on a single generator and model.