We’re over half a year into the age of AI, and its abilities and limitations for both text and image generation are fairly well-known. However, the available AI platforms have had a number of improvements over the past months, and have become markedly better. We are slowly but surely getting to the point where generative image AIs know what hands should look like.
But do they know what science looks like? Are they a reasonable replacement for stock images? Those are the meaningful questions if they are going to be useful for the purposes of life science marketing. We set to answer them.
A Few Notes Before I Start Comparing Things
Being able to create images which are reasonably accurate representations is the bare minimum for the utility of AI in replacing stock imagery. Once we move past that, the main questions are those of price, time, and uniqueness.
AI tools are inexpensive compared with stock imagery. A mid-tier stock imagery site such as iStock or ShutterStock will charge roughly $10 per image if paid with credits or anywhere from $7 to roughly a quarter per image if you purchase a monthly subscription. Of course, if you want something extremely high-quality, images from Getty Images or a specialized science stock photo provider like Science Photo Library or ScienceSource can easily cost many hundreds of dollars per image. In comparison, Midjourney’s pro plan, which is $60 / month, gives you 30 hours of compute time. Each prompt will provide you with 4 images and generally takes around 30 seconds. You could, in theory, acquire 8 images per minute, meaning each costs 0.4 cents. (In practice, with the current generation of AI image generation tools, you are unlikely to get images which match your vision on the first try.) Dall-E’s pricing is even simpler: each prompt is one credit, also provides 4 images, and credits cost $0.13 each. Stable Diffusion is still free.
Having used stock image sites extensively, and having spent some time playing around with the current AI offerings for purposes other than business, it’s not clear to me which is more convenient and takes less time. Sometimes you’ll get lucky and get a good AI image the first try, but you could say the same about stock image sites. Where AI eliminates the need to go through pages and pages of stock images to find the right one, it replaces that with tweaking prompts and waiting for the images to generate. It should be noted that there is some learning curve to using AI as well. For instance, telling it to give you a “film still” or “photograph” if you want a representation of real life which isn’t meant to look illustrated and cartoonish. There’s a million of these tricks and each system has its own small library of commands which helps to be familiar with so you can get an optimal output. Ultimately, AI probably does take a little bit more time, but it also requires more skill. Mindlessly browsing for stock images is still much easier than trying to get a good output from a generative AI (although playing with AI is usually more fun).
Where stock images simply can’t compete at all is uniqueness. When you generate an image with an AI, it is a unique image. Every image generated is one of one. You don’t get the “oh, I’ve seen this before” feeling that you get with stock images, which is especially prevalent for life science / laboratory topics given the relatively limited supply of scientific stock images. We will probably, at some point in the not too distant future, get past the point of being able to identify an AI image meant to look real by the naked eye. Stock images have been around for over a century and the uniqueness problem has only become worse. It is inherent to the medium. The ability to solve that problem is what excites me most about using generative AI imagery for life science marketing.
The Experiment! Ground Rules
If this is going to be an experiment, it needs structure. Here is how it is going to work.
The image generators & stock photo sites used will be:
I was going to include ShutterStock but there’s a huge amount of overlap with iStock, I often find iStock to have slightly higher-quality images, and I don’t want to make more of a project out of this than it is already going to be.
I will be performing 10 searches / generations. To allow for a mix of ideas and concepts, some will be of people, some will be of things, I’ll toss in some microscopy-like images, and some will be of concepts which would normally be presented in an illustrated rather than photographed format. With the disclaimer that these concepts are taken solely from my own thoughts in hope of trying to achieve a good diversity of concepts, I will be looking for the following items:
- A female scientist performing cell culture at a biosafety cabinet.
- An Indian male scientist working with an LC-MS instrument.
- An ethnically diverse group of scientists in a conference room holding a lab meeting. One scientist presents their work.
- A close up of liquid dripping from pipette tips on a high-throughput automated liquid handling system.
- An NGS instrument on a bench in a genomics lab.
- A high-magnification fluorescent micrograph of neural tissues.
- A colored scanning electron micrograph of carcinoma cells.
- A ribbon diagram of a large protein showing quaternary structure.
- A 3D illustration of plasmacytes releasing antibodies.
- An illustration of DNA methylation.
Such that nothing has an edge, none of these are things which I have recently searched for on stock image sites nor which I have previously attempted to generate using AI tools. Note that these are solely the ideas which I am looking for. These are not necessarily the exact queries used when generating AI images or searching the stock photo sites.
Looking for stock images and generating AI graphics are very different processes but they both share one critical dimension: time. I will therefore be limiting myself to 5 minutes on each platform for each image. That’s a reasonable amount of time to try to either find a stock image or get a decent output from an AI. It will also ensure this experiment doesn’t take me two days. Here we go…
Round 1: A female scientist performing cell culture at a biosafety cabinet.
One thing that AI image generators are really bad at in the context of the life sciences is being able to identify and reproduce specific things. I thought that this one wouldn’t be too hard because these models are in large part trained on stock images and there’s a ton of stock images of cell culture, many of which look fairly similar. I quickly realized that this was going to be an exercise in absurdity and hilarity when DALL-E gave me a rack of 50 ml Corning tubes made of Play-Doh. I would be doing you a grave disservice if I did not share this hilarity with you, so I’ll present not only the best images which I get from each round, but also the worst. And oh, there are so many.
I also realized that the only real way to compensate for this within the constraints of a 5-minute time limit is to mash the generate button as fast as I can. When your AI only has a vague idea of what a biosafety cabinet might look like and it’s trying to faithfully reproduce them graphically, you want it to be able to grasp at as many straws as possible. Midjourney gets an edge here because I can run a bunch of generations in parallel.
Now, without further ado, the ridiculous ones…
Round 1 AI Fails
Dall-E produced a large string of images which looked less like cell culture than women baking lemon bars.
Midjourney had some very interesting takes on what cell culture should look like. My favorite is the one that looks like something in a spaceship and involves only machines. The woman staring at her “pipette” in the exact same manner I am staring at this half-pipette half-lightsaber over her neatly arranged, unracked tubes is pretty good as well. Side note: in that one I specifically asked for her to be pipetting a red liquid in a biosafety cabinet. It made the gloves and tube caps red. There is no liquid. There is no biosafety cabinet.
For those who have never used it, Stable Diffusion is hilariously awful at anything meant to look realistic. If you’ve ever seen AI images of melted-looking people with 3 arms and 14 fingers, it was probably Stable Diffusion. The “best” it gave me were things that could potentially be biosafety cabinets, but when it was off, boy was it off…
Rule number one of laboratories: hold things with your mouth. (Yes we are obviously kidding, do not do that.)
That was fun! Onto the “successes.”
Round 1 AI vs. Stock
Midjourney did a wonderful job of creating realistic-looking scientists in labs that you would only see in a movie. Also keeping with the movie theme, Midjourney thinks that everyone looks like a model; no body positivity required. It really doesn’t want people to turn the lights on, either. Still, the best AI results, by a country mile, were from Midjourney.
The best Dall-E could do is give me something that you might confuse as cell culture at a biosafety cabinet if you didn’t look at it and were just looking past it as you turned your head.
Stable Diffusion’s best attempts are two things which could absolutely be biosafety cabinets in Salvador Dali world. Also, that scientist on the right may require medical attention.
Stock image sites, on the other hand, produce some images of cell culture in reasonably realistic looking settings, and it took me way less than 5 minutes to find each. Here are images from iStock, Getty Images, and Science Photo Library, in that order:
First round goes to the stock image sites, all of which produced a better result than anything I could coax from AI. Round goes to stock sites. AI 0 – 1 Stock.
Round 2: An Indian male scientist working with an LC-MS instrument.
I am not confident that AI is going to know what an LC-MS looks like. But let’s find out!
One notable thing that I found is that the less specific you become, the easier it gets for the AI. The below image was a response to me prompting Dall-E for a scientist working with an LC-MS, but it did manage to output a realistic looking person in an environment that could be a laboratory. It’s not perfect and you could pick it apart if you look closely, but it’s pretty close.
A generic prompt like “photograph of a scientist in a laboratory” might work great in Midjourney, or even Dall-E, but the point of this experiment would be tossed out the window if I set that low of a bar.
Round 2 AI Fails
Stable Diffusion is terrible. It’s difficult to tell the worst ones from the best ones. I was going to call one of these the “best” but I’m just going to put them all here because they’re all ridiculous.
Round 2 AI vs. Stock
Midjourney once again output the best results by far, and had some valiant efforts…
Once thing you’ve likely noticed is that none of these are Indian men! While we found good images of scientists performing LC-MS, we couldn’t narrow it down to both race and gender. Sometimes you have to take what you can get! We were generally able to find images which show more diversity, however, and it’s worth noting that Science Photo Library had the most diverse selection (although many of their images which I found are editorial use only, which is very limiting from a marketing perspective).
Round 2 goes to the stock sites. AI 0 – 2 Stock.
Round 3: An ethnically diverse group of scientists in a conference room holding a lab meeting. One scientist presents their work.
This should be easier all around.
Side note: I should’ve predicted this, but as the original query merely asked for science, my initial Midjourney query made it look like the lab was presenting something out of a sci-fi game. Looked cool, but not what we’re aiming for.
Round 3 AI Fails
Dall-E presented some interesting science on the genetic structure of dog kibble.
Dall-E seemed to regress with these queries, as if drawing more than one person correctly was just way too much to ask. It produced a huge stream of almost Picasso-esque people presenting something that vaguely resembled things which could, if sufficiently de-abstracted, be scientific figures. It’s as if it knows what it wants to show you but is drawing it with the hands of a 2 year old.
Stable Diffusion is just bad at this. This was the best it could do.
Round 3 AI vs. Stock
Take the gloves off, this is going to be a battle! While Midjourney continued its penchant for lighting which is more dramatic than realistic, it produced a number of beautiful images with “data” that, while they are extravagant for a lab meeting, could possibly be illustrations of some kind of life science. A few had some noticeable flaws – even Midjourney does some weird stuff with hands sometimes – but they largely seem usable. After all, the intent here is as a replacement for stock images. Such images generally wouldn’t be used in a way which would draw an inordinate amount of attention to them. And if someone does notice a small flaw that gives it away as an AI image, is that somehow worse than it clearly being stock? I’m not certain.
Stock images really fell short here. The problem is that people taking stock photos don’t have data to show, so they either don’t show anyone presenting anything, or they show them presenting something which betrays the image as generic stock. Therefore, to make them look like scientists, they put them in lab coats. Scientists, however, generally don’t wear lab coats outside the lab. It’s poor lab hygiene. Put a group of scientists in a conference room and it’s unusual that they’ll all be wearing lab coats.
That’s exactly what iStock had. Getty Images had an image of a single scientist presenting, but you didn’t see the people he was presenting to. Science Photo Library, which has far less to choose from, also didn’t have people presenting visible data. The three comps are below:
Side Note / ProTip: You can find that image from Getty Images, as well as many images that Getty Images labels as “royalty free” on iStock (or other stock image sites) for way less money. Getty will absolutely fleece you if you let them. Do a reverse image search to find the cheapest option.
Considering the initial idea we wanted to convey, I have to give this round to the AI. The images are unique, and while they lack some realism, so do the stock images.
Round 3 goes to AI. AI 1 – 2 Stock.
Let’s see if Dall-E or Stable Diffusion can do better in the other categories.
Round 4: A close up of liquid dripping from pipette tips on a high-throughput automated liquid handling system.
I’ve seen nice stock imagery of this before. Let’s see if AI can match it, and if I can readily find it again on the stock sites.
Round 4 AI Fails
Dall-E had a long string of images which looked like everything shown was made entirely of polystyrene and put in the autoclave at too high a temperature. You might have to click to expand to see the detail. It looks like everything partially melted, but then resolidified.
Stable Diffusion is more diffuse than stable. Three of these are the best that it did while the fourth is when it gave up and just started barfing visual static.
This is the first round where Midjourney, in my opinion, didn’t do the best job. Liquid handling systems have a fair amount of variability in how they can be presented, but pipette tips do not, and it didn’t seem to know what pipette tips should look like, nor how they would be arranged in a liquid handling system. These are the closest it got:
Very pretty! Not very accurate.
Round 4 AI vs. Stock
We have a new contestant for the AI team! Dall-E produced the most realistic looking image. Here you have it:
Not bad! Could it be an automated pipetting system? We can’t see it, but it’s possible. The spacing between the tips isn’t quite even and it looks like PCR strips rather than a plate, but hey, a microplate wasn’t part of the requirements here.
I didn’t get the drips I was looking for – probably needed to dig more for that – but we did get some images which are obviously liquid handling systems in the process of dispensing liquids.
As valiant of an effort as Dall-E had, the images just aren’t clean enough to have the photorealism of real stock images. Round goes to the stock sites. AI 1 – 3 Stock.
Round 5: An NGS instrument on a bench in a genomics lab.
I have a feeling the higher-end stock sites are going to take this, as there aren’t a ton of NGS instruments so it might be overly specific for AI.
Round 5 AI Fails
Both Midjourney and Dall-E needed guidance that a next-generation sequencer wasn’t some modular device used for producing techno music.
With Dall-E, however, it proved to not be particularly trainable. I imagine it’s AI mind thinking: “Oh, you want a genome sequencer? How about if I write it for you in gibberish?” That was followed by it throwing it’s imaginary hands in the air and generating random imaginary objects for me.
Midjourney also had some pretty but far-out takes, such as this thing which looks much more like an alien version of a pre-industrial loom.
Round 5 AI vs. Stock
This gets a little tricky, because AI is never going to show you a specific genome sequencer, not to mention that if it did you could theoretically run into trademark issues. With that in mind, you have to give them a little bit of latitude. Genome sequencers come in enough shapes and sizes that there is no one-size-fits-all description of what one looks like. Similarly, there are few enough popular ones that unless you see a specific one, or its tell-tale branding, you might not know what it is. Can you really tell the function of one big gray plastic box from another just by looking at it? Given those constraints, I think Midjourney did a heck of a job:
There is no reason that a theoretical NGS instrument couldn’t look like any of these (although some are arguably a bit small). Not half bad! Let’s see what I can get from stock sites, which also will likely not want to show me logos.
iStock had a closeup photo of a Minion, which while it technically fits the description of what we were looking for, it doesn’t fit the intent. Aside from that it had a mediocre rendering of something supposed to be a sequencer and a partial picture of something rather old which might be an old Sanger sequencer?
After not finding anything at all on Getty Images, down to the wire right at the 5:00 mark I found a picture of a NovaSeq 6000. Science Photo Library had an image of an ABS SOLiD 4 on a bench in a lab with the lights off.
Unfortunately, Getty has identified the person in the image, meaning that even though you couldn’t ID the individual just by looking at the image, it isn’t suitable for commercial use. I’m therefore disqualifying that one. Is the oddly lit (and extremely expensive) picture of the SOLiD 4 or the conceptually off-target picture of the Minion better than what the AI came up with? I don’t think I can conclusively say either way, and one thing that I dislike doing as a marketer is injecting my own opinion where it shouldn’t be. The scientists should decide! For now, this will be a tie.
AI 1, Stock 3, Tie 1
Round 6: A high-magnification fluorescent micrograph of neural tissues.
My PhD is in neuroscience so I love this round. If Science Photo Library doesn’t win this round they should pack up and go home. Let’s see what we get!
Round 6 AI Fails
Dall-E got a rough, if not slightly cartoony, shape of neurons but never really coalesced into anything that looked like a genuine fluorescent micrograph (top left and top center in the image below). Stable Diffusion, on the other hand, was either completely off the deep end or just hoping that if it overexposed out-of-focus images enough that it could slide by (top right and bottom row).
Round 6 AI vs. Stock
Midjourney produced a plethora of stunning images. They are objectively beautiful and could absolutely be used in a situation where one only needed the concept of neurons rather than an actual, realistic-looking fluorescent micrograph.
They’re gorgeous, but they’re very obviously not faithful reproductions of what a fluorescent micrograph should look like.
iStock didn’t produce anything within the time limit. I found high-magnification images of neurons which were not fluorescent (probably colored TEM), fluorescent images of neuroblastomas (not quite right), and illustrations of neurons which were not as interesting as those above.
Getty Images did have some, but Science Photo Library had pages and pages of on-target results. SPL employees, you still have jobs.
AI 1, Stock 4, Tie 1
Round 7: A colored scanning electron micrograph of carcinoma cells.
This is another one where Science Photo Library should win handily, but there’s only one way to find out!
Round 7 AI Fails
None of the AI tools failed in such a spectacular way that it was funny. Dall-E produced results which suggested it almost understood the concept, although could never put it together. Here’s a representative selection from Dall-E:
… and from Stable Diffusion, which as expected was further off:
Round 7 AI vs. Stock
Midjourney actually got it, and if these aren’t usable, they’re awfully close. As with the last round, these would certainly be usable if you needed to communicate the concept of a colored SEM image of carcinoma cells more than you needed accurate imagery of them.
Getty Images did have some real SEM images, but the ones of which I found were credited to Science Photo Library and their selection was absolutely dwarfed by SPL’s collection, which again had pages and pages of images of many different cancer cell types:
Here’s where this gets difficult. On one hand, we have images from Midjourney which would take the place of an illustration and which cost me less than ten cents to create. On the other hand, we have actual SEM images from Science Photo Library that are absolutely incredible, not to mention real, but depending on how you want to use them, would cost somewhere in the $200 – $2000 range per photo.
To figure out who wins this round, I need to get back to the original premise: Can AI replace stock in life science marketing? These images are every bit as usable as the items from iStock. Are they as good as the images from SPL? No, absolutely not. But are marketers always going to want to spend hundreds of dollars for a single stock photo? No, absolutely not. There are times when it will be worth it, but many times it won’t be. That said, I think I have to call this round a tie.
AI 1, Stock 4, Tie 2
Round 8: A ribbon diagram of a large protein showing quaternary structure.
This is something that stock photo sites should have in droves, but we’ll find out. To be honest, for things like this I personally search for images with friendly licensing requirements on Wikimedia Commons, which in this case gives ample options. But that’s outside the scope of the experiment so on to round 8!
Round 8 AI Fails
I honestly don’t know why I’m still bothering with Stable Diffusion. The closest it got was something which might look like a ribbon diagram if you took a massive dose of hallucinogens, but it mostly output farts.
Dall-E was entirely convinced that all protein structures should have words on them (a universally disastrous yet hilarious decision from any AI image generator) and I could not convince it otherwise:
This has always baffled me, especially as it pertains to DALL-E, since it’s made by OpenAI, the creators of Chat GPT. You would think it would be able to at least output actual words, even if used nonsensically, but apparently we aren’t that far into the future yet.
Round 8 AI vs. Stock
While Midjourney did listen when I told it not to use words and provided the most predictably beautiful output, they are obviously not genuine protein ribbon diagrams. Protein ribbon diagrams are a thing with a very specific look, and this is not it.
AI 1, Stock 5, Tie 2. At this point AI can no longer catch up to stock images, but we’re not just interested in what “team” is going to “win” so I’ll keep going.
Round 9: A 3D illustration of plasmacytes releasing antibodies.
I have high hopes for Midjourney on this. But first, another episode of “Stable Diffusion Showing Us Things”!
Round 9 AI Fails
Stable Diffusion is somehow getting worse…
DALL-E was closer, but also took some adventures into randomness.
Midjourney wasn’t initially giving me the results that I hoped for, so to test if it understood the concept of plasmacytes I provided it with only “plasmacytes” as a query. No, it doesn’t know what plasmacytes are.
Round 9 AI vs. Stock
I should just call this Midjourney vs. Stock. Regardless, Midjourney didn’t quite hit the mark. Plasmacytes have an inordinately large number of ways to refer to them (plasma cells, B lymphocytes, B cells, etc.) and it did eventually get the idea, but it never looked quite right and never got the antibodies right, either. It did get the concept of a cell releasing something, but those things look nothing like antibodies.
I found some options on iStock and Science Photo Library (shown below, respectively) almost immediately, and the SPL option is reasonably priced if you don’t need it in extremely high resolution, so my call for Midjourney has not panned out.
Stock sites get this round. AI 1, Stock 6, Tie 2.
Round 10: An illustration of DNA methylation.
This is fairly specific, so I don’t have high hopes for AI here. The main question in my mind is whether stock sites will have illustrations of methylation specifically. Let’s find out!
Round 10 AI Fails
I occasionally feel like I have to fight with Midjourney to not be so artistic all the time, but adding things like “realistic looking” or “scientific illustration of” didn’t exactly help.
Midjourney also really wanted DNA to be a triple helix. Or maybe a 2.5-helix?
I set the bar extremely low for Stable Diffusion and just tried to get it to draw me DNA. Doesn’t matter what style, doesn’t need anything fancy, just plain old DNA. It almost did! Once. (Top left below.) But in the process it also created a bunch of abstract mayhem (bottom row below).
With anything involving “methylation” in the query, DALL-E did that thing where it tries to replace accurate representation with what it thinks are words. I therefore tried to just give it visual instructions, but that proved far too complex.
Round 10 AI vs. Stock
I have to admit, I did not think that it was going to be this hard to get reasonably accurate representations of regular DNA out of Midjourney. It did produce some, but not many, and the best looked like it was made by Jacob the Jeweler. If methyl groups look like rhinestones, 10/10. Dall-E did produce some things that look like DNA stock images circa 2010. All of these have the correct helix orientation as well: right handed. That was a must.
The point again goes to stock sites.
Final Score: AI 1, Stock 7, Tie 2.
Conclusion / Closing Thoughts
Much like generative text AI, generative image AI shows a lot of promise, but doesn’t yet have the specificity and accuracy needed to be broadly useful. It has a way to go before it can reliably replace stock photos and illustrations of laboratory and life science concepts for marketing purposes. However, for concepts which are fairly broad or in cases where getting the idea across is sufficient, AI can sometimes act as a replacement for basic stock imagery. As for me, if I get a good feeling that AI could do the job and I’m not enthusiastic about the images I’m finding from lower-cost stock sites, I’ll most likely give Midjourney a go. Sixty dollars a month gets us functionally infinite attempts, so the value here is pretty good. If we get a handful of stock images out of it each month, that’s fine – and there’s some from this experiment we’ll certainly be keeping on hand!
I would not be particularly comfortable about the future if I was a stock image site, but especially for higher-quality or specialized / more specific images, AI has a long ways to go before it can replace them.