Photographers are periodically under threat as each wave of technology renders various specialisms obsolete.
I was told when I went freelance there was no future because of “digital”. While that’s not been the case, it did kill news (with the help of the other horsemen: the internet, stock photography*, and pestilence).
But certainly the relentless advance of Photoshop, the iPhone, technical automation and instant communication can be punishing to an industry, and frustratingly so when combined with the inherent lack of understanding that goes along with its mass market audience.
And now AI.
The fear is not that it will do the creatives’ job for them - clearly this is nonsense. If there is one, it’s in the same Faustian pact threatened by the (currently bland) utterances of ChatGPT: anything which can be automated - and automated well - will, at first, free up creatives’ time and energy. But in exchange, and quickly after, there will be less need for those creatives. To put it another way, not everything I do in the course of a project is specialised, difficult or skilled labour. And when that’s taken out of my hands, there’s imbalance.
One-eyed Fletch, taken just now to illustrate what a good boy he is.
But let’s go back a bit. I can’t publish what I said when I tested my Canon R6 for the first time two years ago, but let’s just say its eye-tracking technology was game-changing. Out of the box, my first test photos were of one of my cats, Fletch, sitting six feet away in front of a glowing fire in an otherwise dark room. And every single frame was sharp. My previous camera would never have achieved this. I should add he only has one eye, and the camera found it. (I don’t even bother checking sharpness any more.)
And I won’t bore you with recent advances in Photoshop and Lightroom, but will just say that many things which may have taken a couple of (boring) minutes just a couple of years ago can now be done in a few seconds, thanks to AI.
But as I’ve suggested, easier for me means easier for everyone. So at least in some areas, what I can bring to the table in terms of skills and experience is gradually reduced, as we all level up. Part of my time from every job is spent assessing and selecting images. There are now apps that can check sharpness, composition and blinks, and do this for me. I know a dozen ways to mask out hair, but that hard-earned knowledge is less and less useful with every incremental update to Photoshop’s “refine hair” tool. And shooting with (my current overuse of) a very shallow depth of field with my expensive 50mm lens - unthinkable before my (also expensive) R6 - is more and more convincingly achieved with my iPhone’s “portrait mode”.
None of this is new of course. Technology improves. But just as with the text-based AI’s, we’re way beyond autocorrect here. Not for the first time, I couldn’t tell if a photographer was joking or not when he posted on my forum a few weeks ago that he was considering ditching his gear and using an iPhone. So aside from the usual concerns about the mixed blessings of hardware / software updates (which improve results) and automation (which speeds up post-production), we’re now reaching a very different point, where you can create work with minimal input or understanding. You don’t even need to to own a camera.
So there’s that.
I based most of the prompts on my current headshot
But what about the results themselves? How good are the images that the AI’s can generate? Why should I worry about creeping AI indirectly affecting my livelihood, if it can just smash through the front door by actually making images, and doing a better job at the same time**?
Let’s consider portraits, and corporate portraiture, in particular (interestingly, the latter is an area which has massively grown because of digital, since every company needs a website, and often an “About Us” page). I’m interested to see if AI touches on my commissioned work directly: going back to the original “nonsense” concern, if we can just describe a person and get an image, then why use a photographer at all? Will we get to the day where, say, HR could ask for temporary access to a staff member’s Facebook or phone photos, pull out recent images, and use them to generate professional-looking headshots in the house style in a matter of minutes?
I had a go with DALL-E, midjourney and Stable Diffusion. Using a source image of myself, taken by me, I used various prompts including “corporate portrait”, “professional”, “headshot”, “in the style of Alex Rumford”, and “photorealistic” to generate new images. Would they resemble me accurately?
No. Not even slightly.
Which was a relief.
The more available images there are online, the better the results (for instance Beyoncé), but currently - when uploading - midjourney (which seemed to be the best tool for this) allows only two. And combining two images didn’t change things much.
Secondary images (left by the brilliant @docubyte)
Even playing around with the sliders and prompts the results were, at best, approximations for a better version of me. Results were slightly cartoonish, “Americanised” (presumably because most source material is from the US?), and almost always better-looking.
There’s been a lot written about AI bias, but it’s interesting to see results are akin to a Snapchat filter. It has the disappointing effect of feeling less descriptive (“this is what I think you would look like”) than it does prescriptive (“this is how you should look, ugly”). It’s depressing enough thinking about the negative effects of existing in-phone editing software which makes noses smaller and eyes larger, skin smoother and lips fuller.
I’m only guessing, but this “beauty ideal” in AI would presumably be from the influence of the more photogenic members of society (actors, models, perhaps even stock image models) whose appearance would make up a large percentage of the millions of source material portraits, and so influence the output.
The first set, which had a decent variation, had minimal prompts.
The faces on #1 and #3 could pass as photographs, but the shirts and collars look drawn); #2 and #4 have slightly unusual cheekbones. #4 has a glint in the eye, which is interesting.
Looking at #3, it’s slightly Pixar-cartoonish (and are the eyes quite right?)
This time with a jacket. #4 looks the most realistic. Again, none look like me.
Wider shots, with the subject being smaller, could be more forgiving with facial features (#3).
This set was generated with no word prompts, just two source images. They’re consistent, but, alas, consistently not close to the originals.
There’s a lot of talk about bias in AI: a key prompt here was “friendly” and the results are decidedly more feminine.
Again without the “corporate” prompt, there’s a lot more variation in this set.
So it’s a way off, yet. And while there are options to further refine / create variations and possibly improve results, none of these originals was close enough for this to be worth trying. And to be clear: the primary goal here is photographic realism. If you like, there are of course plenty of options for interesting filters or styles one could apply to your LinkedIn portrait which aren’t photography at all (I recently saw a really effective set on a website (presumably architects or graphic designers) which had a clear Julian Opie look to them). But if an image is meant to be realistic, it has to look exactly like you.
You’ll note that I’ve smuggled in the assumption that a corporate portrait’s main purpose is merely to describe appearance. Which it isn’t. That’s a passport or the badge ID you’re thinking of. A portrait - yes, even the humble corporate headshot - needs to say something about the subject. Actually (in theory, at least) the mood / expression in a plain shot on a white background therefore has to count for more than a full-length environmental portrait (where clothes, surroundings and lighting help do the work for a more unique and interesting shot). That’s to say, the more easily something can be copied and regenerated, the more bland it would have to become. But I could be overstating this, and the market wouldn’t care: a free image which takes minimal effort and minutes to produce and remains pretty neutral will usually be preferable to a far more costly photograph which ‘says’ something.
And what about environmental portraiture? The below examples are extrapolated from the headshot, and are mostly awful (thankfully).
With just a headshot, the prompt here was “half-length, wearing a grey long-sleeved t-shirt, standing in a modern office.” #3 isn’t too bad, but the others aren’t anywhere near realistic enough to be photos, nor are they stylised enough to be anything else.
Stability AI (in DreamStudio) has the option of using data from (whilst keeping) an existing image to extend it. In another attempt (not shown here) I had it work on the source image at the same time, but it immediately looked less like me. #4 is the only one which nearly works. Perhaps a couple more iterations might something passable.
Starting with just a headshot, this is from DALL-E and took a couple of minutes to build. The description was, “Man in a long-sleeved grey t-shirt in front of a plain office background.” A few minutes further in photoshop and this could be passable.
Perhaps it’s just not what AI is good at. While so much of the concept art is truly brilliant, and some of it realistic, my first impressions are that it’s not directly a threat to portrait photography.
But I’ll check back again in six months.
*Stock photography is a zombie. It’s dead, yet it continues to feed by killing off potential commissions.
**I’m told that the effect of AI is already felt directly, or is soon to be, in photographic areas including automotive, fashion e-commerce, interiors and still life / products. I do a lot of portraits, and even if they can’t be done by AI, market forces mean that if other genres’ photographers’ work is reduced, it makes sense for them to move in on my patch (the positive term is “diversifying”…) in the same way that PR photographers had to move into weddings when the ex-press photographers joined their ranks, en masse, a decade ago. Leading to the question: where to go next? What genres will be safe tomorrow in an industry entirely based around technology?)