Trying Krea: AI Image Generator

April 17th 2025

Holy shit this program makes me want to pull my hair out!!!

I thought that out of all the professions to be replaced AI it would be graphic designers. Now? I don't think graphic designers will ever be replaced. Here's why.

I tried Krea for the first time with the objective to generate some art for my businesses' website. I thought this would be a pretty standard use-case for this new technology, and that I would be in and out in a couple of hours. After all, if Krea's homepage looks like this, full of really cool concept art, surely I can generate some simple diagrams? Right?

There's tutorials on the internet where you basically can draw very rough shapes, prompt an AI model, and have a pretty good image pop out the other end. I thought I would use a similar workflow, but go a step further.

I thought of a couple of ideas, made some collages in GIMP (Photoshop alternative but free), and I thought I could enhance them with Kera. The collages were a bunch of thrown together, transformed images downloaded from the internet. I thought that I could take the following image and "enhance" it by removing the copy-paste lines, enhancing some aspects of the lighting, extending the pins going to the silicon die, etc... Basically I wanted the AI to take the image I made and make it near-perfect.

Okay, it's not my best work, but I think you kind of get the gist right? It's supposed to be a satellite, in space, overlooking a computer chip. There's some spotlights from the satellite to the chip, highlighting certain blocks. Around the chip is a set of pins (standard way of connecting to a package), and around that is a "circuit-board".

First problem. The AI simply does not understand what any of these things are supposed to look like. It has no idea what a "silicon die" is. It can barely do proper satellites. I'd argue that it's worse at satellites than putting the correct number of fingers on a human hand. It has a limited understanding of a circuit board (not that I was going for realism in this department). It's just absolutely not prepared to generate any of these things from scratch.

What it is good at is generating humans, common scenes, stock photos, etc. If you think you can find an example of this in a Tumblr blog from 2008, then you can probably generate it.

Second problem. If the prompt it longer than a relatively short sentence, the generation "trails off" and doesn't generate anything after a certain point. Can you describe the image above using a short sentence? Honestly, I can't. If you leave anything out of the prompt, then even with the reference photo, the generation is poor.

Third problem. Krea uses "filters" to hide generation deficiencies. Without a filter the outputs kinda look like junk. As soon as you put a filter on, the generations get more consistent, look better, but also match the filter style. But it's also harder to match a prompt photo that doesn't already look or match the scenes expected by the filter! There's no filter for "cool things in space", unless you make a custom one.

And lastly, the other option is to use the composing interfaces in Krea to make your own version of the template you rigged up in Photoshop, and try enhancing that one. Again though, you'll run into the problem that if the individual objects that make up your composite are not well represented in their dataset, then you can't generate the parts of the whole.

At this point, I might as well spend hours in Photoshop and make the thing I wanted from scratch!