Gab AI Image Generation in Style

«Men of all shapes and sizes, ages and creeds, and states of marital or relationship bliss enjoy, every now and then, the sight of a woman with no clothes on. It’s just as well we do, you know, otherwise there’d be no new little earthlings, would there? If you want to call that oppression or sexism or the commodification of the female body then go right ahead, but don’t expect me to talk to you at dinner parties. I prefer to call it sexual attraction, but then I’m a sad fuck who spends half his life in front of computer, so what the hell do I know?»
Michael Marshall Smith, «More Tomorrow»

Playing around with Gab AI (https://gab.ai or gab.ai/start/gabby to try it out) — it is impressive. To my eyes, it can imitate Renaissance-Michelangelo styles and much more.

«To my eyes», because I am not an art expert. The closest I come is to photography, when I notice … weaknesses in composition or lighting that escape others. But in the same vein, I would not notice when a painting style if «off». I only see the general direction, not the crucial details. I guess Gelett Burgess put it well with «I don’t know anything about art, but I know what I like.»

BTW, I don’t believe in trigger warnings, but it would be an asshole move not to point out that some of the following images contain nudity. It is probably not something to read at work or any public place.

So, let’s have a look.

Renaissance

First Renaissance (did already show a few in this posting), and yeah, I went a bit DC’s «Poison Ivy» here. And BTW, you can provide quite a specific hair color and style — there is more than just «red» (e.g., Light Red/Copper, Auburn, True Red, Deep Red/Burgundy, or Red Brown).

If you look at the image above, you see a common problem with AI generated images. You probably notice the small mistake in her left eye (from her POV), but can you count her number of fingers on her right hand on the number of fingers on yours? It’s these «tiny» details that are incredibly frustrating.

And of course, there are other variants possible that «Poison Ivy», e.g.:

Yeah, it’s hit or miss, not only with artifacts but also with style. The following is not exactly Renaissance, but still nice.

In case you are wondering about the strawberries — I provided «strawberry blonde» as hair color. I guess it crept into other aspects of the painting — or it was just an accident.

 

Luis Royo Style

Looking at the styles, if Michelangelo works, why not something contemporary like Luis Royo. An incredible painter — just search for him online, you will find a lot of impressive paintings. For example, on his official page: https://www.luisroyo.com

Let’s see whether it works, if you specify an image with the intro:

(masterpiece:1.5), (by Luis Royo:1.6), (painting:1.2), (bestquality:1.2), (highlydetailed:1.2),

and then some other information, e.g., that it is one person, clothes, hair color, etc.

Results, among others:

Some deviations from his style, but close. Others are a bit further off, but still beautifully done:

Yeah, I know, but look at her right hand.

Impressive.

 

Other Styles

Overall, Gab AI needs to have enough training data to replicate a style. I do not know with which data it was trained and when it had enough. From the results — ignoring the huge variance in the produced images and going for the best cases — and judged by a layperson: Michelangelo, yes, Luis Royo, at least partly. More modern graphic artists like Stjepan Šejić (e.g., «Sunstone» graphic novel) perhaps partly.

With «(by Stjepan Šejić:1.6), (Sunstone graphic novel)» and a few choice keywords you end up in the general direction. Although the style is still different.

(And yeah, the proportions are off. Artifacts … meh.)

 

Painting Real People

Finally, how about real people? No a fan of stardom, but shortly before trying it out, I heard a story about Taylor Swift at the Super Bowl (interesting marketing strategy). So let’s try it out:

(masterpiece:1.5), (by John William Waterhouse:1.6), (Taylor Swift:1.5), (painting:1.2), (bestquality:1.2), (highlydetailed:1.2), ultra-detailed, warm, (1girl), (detailedeyes:1.2), (lighting:natural), (render:high), {shot:Portrait:1.4}, (today:1.0).

Less John William Waterhouse (clothes yes, but background completely different style), but recognizable as Taylor Swift. Apparently it has enough training data to create a somewhat realistic painting of her (again, as a layperson, thankfully).

 

Conclusion

Overall, Gab AI is impressive. The downsides are that it is hard to get the painting exactly how you want it to be, a correction function is severely missing. And it is hit or miss. Not only with image artifacts, also whether the image shows what you did imagine. But hey, if image generation works you can bypass the limits of your drawing skills and show what you see in your mind. Gavriel Salomon’s «Supplantation» on steroids.

One interesting issue is the use of this tool for … errr, revealing paintings. The Swift image above can be seen as a homage (or something). But you could, for example, combine

(masterpiece:1.5), (painting:1.2), (bestquality:1.2), (highlydetailed:1.2), ultra-detailed

with

(Taylor Swift:1.5)

and

(partiallydressed:1.2), (semitransparent:1.2), (naked:1.3), (revealing:1.3)

and end somewhere way different. I mean, who needs leaks if you can go for deep fakes? Hmmm.

So far, Gab AI has some limits and will stop some of these images (more on that in another posting), but yeah, that is quite the explosive issue here.

But anyway, happy generating.