Photoshop AI: Where are we really today?

Photoshop AI: Where are we really today?

There's been a lot of talk in recent months about the new Artificial Intelligence in Photoshop. Especially the topic of Generative Fill and the wondrous things you can do with it. I took an unbiased look at it and tried it out, perhaps a bit more rationally than many YouTubers out there.

 

It's an incredibly exciting topic. I don't usually do image compositions where I blend different scenes together in Photoshop. But every now and then it would be useful to add a detail.

Whether this is reprehensible or a fake is not my point in this article. I don't want to ask any moral questions, I just want to look at the technical capability of the Photoshop 2024 version from my very personal point of view.

After all, many Photoshop apostles have been so hyped about the new Generative Fill features on YouTube over the past few months that it's fair to be a bit skeptical. Here's my experience.

Let's begin

I don't think we need to talk about the cover photo at all. The white top and shorts that the AI has put on Dominika look terrible. Oh yeah, I didn't choose extra bad examples here, but the best results that Photoshop helped generate. So these are supposed to be white shorts. Let's move on quickly.

It is certainly very difficult to adjust perspectives, proportions and colors to an existing image. A mammoth task that Photoshop also handles very quickly.

Lighting is okay, but unfortunately, the golf bag doesn't look the way I would imagine it. More like a foreign object. Like a sports bag with kids' golf gear.

And the golf club itself, well, a lot has been said about crippled hands and AI anyway.

Please do not be scared!

Speaking of crippled... I asked to place a female golfer in a blue costume behind Dominika. The result is more like something out of a horror movie. Three robot arms, one unmotivated leg and is that supposed to be a face?

I really don't want to make fun of AI. Please make up your own mind.

A more pleasing example

Now let's move on to a pleasant example. I wanted to place a golf player in the background. At first, AI generated only sitting players with clubs. A new sport, so to speak. Then I decided on the senior from behind. No face, that's safer.

He looks a bit artificial though, doesn't he?

The cap on Dominika's head looks a bit dull, but at least the light, the proportions and the way it is put on are right.

The cropped golf cart, on the left side of the image, looks like something out of a toy store. I have had 12 versions generated and this is by far the best. I am not convinced.

I like the castle in the background though. Well done, Photoshop! Perhaps the elbow should not have been cut off so squarely for this. But I would like to overlook that for once. Color, brightness and blur are really well done.

When AI refuses to work

 

And again a golf kart in the background in this photo. Well, not so great either. As a client, I wouldn't accept that. In this tiny version of the photo, it's sort of ok.

But now comes the real problem: I wanted to give Dominika a golf club in her hand, but only messages came up that this would be forbidden.

Generative fill result was removed due to violation of guidelines.

Is it related to nudity or bare skin in general? Does AI detect anything too erotic in the hand movement? I don't know. It can't be because of the word Golf Club, because it was generated in the other photo. This means that there is some kind of censorship going on here, which is really annoying.

When AI patronizes, I don't like it at all.

Invisible Golf Club

Let's overdo it

It's impressive what can be done. Sometimes, the generation is totally off and doesn't understand simple things like tree and places a tent instead. Also adding realistic looking sunglasses on Dominika was impossible. I gave up after 30 generations.

And yet, it's addictive to play with it. So, I spent an hour working on this image. You can clearly see the problems:

  1. The landing plane feels immediately fake.
  2. Both golf players are too contrasty. And why is one wearing a full body suit with a hood?
  3. The golf carts still look odd (after 30 times trying).
  4. A squirrel was there in reality, but I was too slow to photograph it. It definitely looked different from this one. But at least, it's cute.
  5. The clothes on the ground look a bit weird, too. It was forbidden to place underwear on the grass. Only saying.

 

But I totally believe the golf ball was lying in the grass, don't you?

And what if I told you there was no sand bunker and no lake...

What is my conclusion?

Generative Fill is an impressive thing. Most generated items look unrealistic as of today, let's be honest. But for small things, or when luck happens, the results are indeed surprisingly good. And this is just the beginning of AI. How good do you think it will be in 5 years from now?

I'm not sure it's a good idea by Adobe to introduce an additional credit payment system for using the tool. Especially when the results are still very modest and especially low-res.

Navigate