Two years ago I wrote about GPT-3 and the first machine-written text that sounds human. At the time I thought I had seen the most spectacular part. I was wrong. Over the past few weeks I have spent my evenings with two new tools open on my laptop, DALL-E 2 and Midjourney, asking them for images. Not drawing them, not hunting for them on stock sites, just describing them in words. And they appeared.
I am not a designer. I am an entrepreneur who, over the years, has paid for a lot of product photo shoots, a lot of banners and a lot of illustrations for articles. So I tested these tools with the eyes of someone who buys images, not someone who makes them. Here is what I saw, what excites me and what worries me.
The two tools, in a nutshell
DALL-E 2 comes from OpenAI, the company behind GPT-3. It was announced in April and for months access was waitlist only. On 20 July, OpenAI opened a beta to up to a million people from that waitlist and switched to a paid, credit-based model: each request uses one credit and returns a few variations. You get a small number of free credits, then you buy packs.
Midjourney is a small independent lab that entered open beta in July. The odd part is the interface: there is no website where you type, but a Discord server. You type a command in a channel alongside hundreds of other people, and within a few dozen seconds four images appear. You can upscale the ones you like or ask for variations. It feels chaotic at first, but it has an upside: you see in real time what others are trying and you learn from their prompts.
There is also Imagen, from Google, announced in May, with results that look impressive in the launch materials. But it is not publicly available, so I cannot comment from experience.
What I asked them for
I tried to be practical and picked the kinds of images every online business needs:
- Product photos for an online store. Say you sell ceramic mugs or scented candles: a product on a wooden table, natural light, neutral background.
- Illustrations for blog posts. Abstract concepts such as “a team working remotely” or “e-commerce growth”, in an editorial illustration style.
- Logos. Simple symbols for an imaginary brand.
- A hypothetical travel poster. Vintage style, a Mediterranean coastline, warm colours, the kind of print you would hang in a hotel lobby.
Midjourney has style, DALL-E 2 does as it is told
The difference shows within minutes. Midjourney has its own recognisable look: dramatic lighting, rich textures, compositions that resemble an album cover or a piece of digital painting. Ask for something simple and you get something beautiful, sometimes more beautiful than you asked for, but not necessarily what you asked for. For posters, mood and conceptual illustrations it is hard to beat.
DALL-E 2 is more literal. Describe a white mug on a wooden table next to a window and that is roughly what you get. The images look closer to photography, less “artistic”, but more predictable. For an online store that wants a product mockup or a clean image for an article, predictability is a feature, not a bug. DALL-E 2 also has a handy trick: you can erase part of an image and ask it to fill that area with something else.
Where they stumble
Anyone who plays with them for a few hours hits the same limits:
- Hands. Extra fingers, fingers melting into each other, impossible wrists. It is almost an inside joke among users.
- Text inside the image. Ask for a poster with a city name and you get letters that look like an invented alphabet. For logos this is a serious problem: the symbol can come out interesting, the brand name almost never.
- Faces. They look fine from a distance; up close you notice asymmetries and slightly wrong eyes. On top of that, OpenAI restricts realistic faces and public figures, precisely to limit abuse.
- Consistency. If you need the same character or the same product across ten different images, getting it identical is very hard.
My takeaway: for a real product that a customer will receive at home, you cannot use an AI-invented image as a product photo. That would be misleading advertising. But for backgrounds, lifestyle scenes, concepts and first drafts, the picture changes.
What it means for designers, marketers and online stores
For marketers, the biggest gain is speed of exploration. Instead of describing to a designer, over three emails, what you have in mind, you can generate twenty visual directions in ten minutes and walk into the meeting with them. For ad creatives, where you test many variations, this can shorten the road considerably. A few years ago I wrote about what artificial intelligence can do in marketing, and I was mostly talking about recommendations and segmentation. The visual side felt much further away. It was not.
For small online stores that cannot afford a photo shoot for every campaign, this means blog images, seasonal banners and backgrounds at a low cost. For designers, honestly, I do not think it is an immediate threat. I think it is a new tool, the way Photoshop was in its day. A good designer knows what to ask for, knows how to choose and knows how to finish an image. Today’s AI gives you raw material, often spectacular, rarely ready to publish without extra work.
Prompt writing, a new skill
What surprised me most: the difference between a dull image and a good one lies almost entirely in how you write the request. Style, type of light, angle, camera lens, era, the name of an art movement all change the result. Lists of keywords and prompt guides are already circulating. I would not be surprised to see people soon calling themselves prompt specialists on LinkedIn. It sounds funny today, but fifteen years ago “SEO specialist” or “social media manager” sounded just as funny.
Copyright and the artists behind the data
This is where things get uncomfortable. These models learned from huge amounts of images from the internet, many created by illustrators and photographers who were never asked. When you ask Midjourney for something “in the style of” a living artist, you get something that genuinely resembles their work. Artists have already started to push back, and rightly so. Who owns the generated image? OpenAI now says users can use it commercially, but the question about training data remains open and, I suspect, will end up in court. We saw something similar, on a smaller scale, with NFTs and the debate about art: technology moves much faster than the rules.
The deepfake risk
The second reason for caution: if you can generate a believable photo of a scene that never happened, you can also lie convincingly. Today’s restrictions help, but they are set by a handful of companies. Once the technology is in everyone’s hands, without filters, we will have a serious problem of trust in images. We have already been through waves of fake news and mistrust with text; with images it will be harder, because we are used to believing what we see.
My verdict, at the end of August 2022
It is impressive. It is early. It is imperfect. It does not replace a good designer or a product photographer today, but it changes the way we start any visual project. My advice to entrepreneurs: try them now, while they are new, with a small budget of credits. Not to fire anyone, but to understand what is coming and to teach your team how to write good prompts.
Have you tried them? What worked best for you and where did you get stuck? And, above all, would you pay for an AI-generated image if it were good enough?