Tuesday 4 April 2023

Generative AI frenzy: Adobe Firefly vs. Midjourney

A firefly spaceship rendering
A spaceship on the moon... looks more like Mars, but whatever. The details on the ship are dodgy, but the overall impression is quite good. With a bit of editing, it's usable. From Firefly.

Companies are going bananas over generative AI. It's being hailed as a tectonic shift, an existential crisis, a matter of 'embrace or die'. 

Is it really all that?


Is it useless and laughably bad? 

Also no. 

A firefly psychadelic landscape
Wasn't super keen on these Firefly results; the hands are as messed up as earlier versions of Midjourney

It can be useful, it's already being embedded in services we use every day, yet it also has serious limitations (it's not 'thinking', it's just probability predicting based on a huge pool of nevertheless limited data). 

One big issue with Midjourney and Stable Diffusion are the ethical (and legal) issues, which stand in the way of wider commercial adoption. The top reasons for companies to adopt new software, after all, are efficiency and price. If it's cheaper, easier and faster, it'll be looked into. But only if it doesn't bring on all sorts of complicated legal headaches. No one wants to deal with the Legal Team. Sorry, guys.

a frog having tea rendered in firefly
The frog was meant to be serving the tea, but this is close enough. Looks pretty good!

Adobe's Firefly is designed to be used commercially from the get go: it's trained only on public domain art and Adobestock, which they have rights to. 

So how does Firefly stack up against the perfidious scrape happy Midjourney? 

Good question. 

I spent 4 months fiddling with Midjourney (before cancelling my subscription), and recently I got in on the Beta for Firefly. 

A psychadelic Firefly desert; I quite like this one.

My expectation for Firefly were (probably unreasonably) high: Adobe is the international standard for professional graphics software, and according to their press release, they're planning on incorporating Generative AI into virtually all their other products.

The interface is clean and simple It has a text input field, plus categories to help spark ideas, paired with little icons. I'm not keen on on them: they feel more limiting than helpful, and the icons are too small to get anything out of them:

The image output is... varied. Some is quite good, but much is still dodgy around the edges. The biggest issue right now is subjective: I am not finding the Firefly output as compelling as Midjourney's. 

That's one interesting US flag on the upper right

The quality difference is, perhaps, telling: many Firefly renders often look like stock art, because the tool was trained on, well, stock art. Midjourney renders have more pizazz. They are often better composed, with more impressive colour schemes and just feel... a lot more fun. 

And that fun factor is very likely derived from the living artists and the millions of copyright artwork that Midjourney (and Stable Diffusion) were trained on. 

Generative AI is only as good as the stuff you feed it. Stock in, stock out.

Maybe I'm imagining it. You decide.

I don't mean to sell Firefly short. Some images are production ready or very close to it. My views are somewhat jaded thanks to thousands of Midjourney images.

Men typing on a laptop in a firefly rendering
Stock, just slightly messed up deformed stock (first one); the others I wouldn't be able to tell without close examination, and even if then, maybe not

Overall I like these, but if you look closely the details (as is typical of AI output) don't hold up.

So while Midjourney is (to me) leagues ahead of Firefly, and Firefly will likely never be able to catch up, perhaps Adobe doesn't intend for that to ever happen anyway. Adobe's goal may just be to provide a cheap customizable stock art alternative for people to use and modify. It will undercut Getty Images and the other stock houses (if Firefly is set at a competitive price point). 

A farm in Ukraine under a magnificent sky
A farm in Ukraine... turned out amazing. Still a bit of a stock feel to it, but I couldn't tell it was AI. How much of an original stock image does it pull in? Are some essentially verbatim?

If you look at dozens of Midjourney images, you notice a smooth slickness, a plastic feel, to them after awhile. Prompt wizardry can minimize or negate this, with lens types and whatnot. Haven't noticed that with Firefly, at least not yet.

Lower value jobs will go to Generative AI, much the same way stock dominates when budget doesn't allow for original art, and they can't find overeager students to exploit. But those jobs are already largely gone anyway.

Lowest common denominator content, generic stuff companies need but don't really care about, could be generated by AI and then curated by humans. 

If the lawsuits go against artists, AI will be more of a threat to job security.

Firefly monkey chess piece
A monkey chess piece rendered in Firefly; pretty solid!

Firefly side navigation
The side nav, with the various categories and options in Firefly

No comments:

Post a Comment