Are all artificial intelligences equal?

AI image generators have been around for a while. But recently they’ve improved dramatically and now should you need them, there’s an abundance of AI-powered generating tools on the internet. But how good are they really? CW&’s Innovation Taskforce put three different AI image generators to the test.

As you may have discovered, or been told, it’s all about the prompt when it comes to AI. So each generator - Midjourney, Runway and Stable Diffusion - was given the same artistic challenge. Read on for the results and our verdict.

Prompt 1: A dark forest. Moss seen up close with tiny blue mushrooms. Photorealistic. Light shines in beams through the canopy.

All excellent propositions, however we prefer Midjourney for the level of photorealism, as much as it can be given the fantastic nature of the scene described in the prompt! Runway’s render feels a little less polished, while Stable Diffusion has a definite cartoonish feel. Overall, the brief has been respected by all three.

Prompt 2: Grogu made of crystal. An environment made of furry hills in pastel tones. Trees are made of iridescent soap bubbles. Two suns in the sky at sunrise. 3D render unreal engine.

None of these feature Grogu, the popular character from Disney’s The Mandalorian. Maybe AI generators aren’t up to speed on the latest shows or maybe this specific fictional creature hasn’t had enough time to leave its mark on the AI zeitgeist. The backgrounds produced are all interesting in their own ways, the prompt was quite wild, so it’s interesting to see how far AI will go to interpret the brief. We still have a preference for Midjourney’s ability to produce really impressive, high-end visuals. Stable Diffusion is sticking to its anime look and feel - good for them!

Prompt 3: Blade runner sci-fi style. Moody rainy city at night. Orange and blue accents neon signs. Street food stall.

Solid results all round! Some generators are better than others with text, but overall it would appear that a pop-culture reference from the 80s has left enough of an imprint on the fabric of the internet to inspire a rich tapestry of visual outcomes (metaphor alert!).

And then there was the time we really, REALLY needed an image depicting a corn field with no perspective, just a front side view of the crops without furrows and all 3 generators really let us down. Thanks for nothing AI.

Prompt 4: Corn field, no rows, no dirt, no soil, flat frontal view, no perspective.

Well that’s neither here nor there. And why the sad black and white palette Runway? Just dreadful..

The verdict:

In short, AI image generators excel in the aesthetics of computer-generated imagery, akin to what software might produce. While some topics bring a wide variety of visual propositions, others - particularly the prompts based on pop culture references - tend to create versions of the same reference. Discrepancies exist in the outcomes and styles for each tool, but AI generators tend to share the same struggles and lack personality.

If you want infallible reliability, need a sliver of emotion, or sometimes just simply what you asked for, we think you’re better off talking to a human. When it comes to talent, AI undeniably has some, but for anything with a soul, nothing beats the flair of living and breathing artists.

There is an indisputable convenience aspect to these tools, not to mention the exhilarating God-like thrill of creating something from ‘scratch’ in mere seconds. As the technology exists, and it’s good, it would seem silly not to leverage it. However, AI will only ever be as good as the prompts it’s given and the content it can muster from the deep dark abyss of the internet’s seemingly endless catalogue of images.

Finally, it’s worth keeping in mind that anything AI-generated was ‘borrowed’ on some level. Determining whether the inspiration is part of the public domain or the work of a little-known artisan is challenging. Until regulations catch-up, we should remember that we bear a responsibility to use this technology with caution and give credit where it’s due.

Previous Next