Yesterday, OpenAI removed the waitlist for DALL·E. Anyone can now sign up and create images from a sentence. This was the last of three major tools to open to the public this summer, all within about ten weeks of one another.
Midjourney, the small independent lab led by David Holz, opened its beta to the public in July, notably through a Discord server. Then, on August 22, Stability AI released Stable Diffusion, including the model itself, which means anyone with an adequate graphics card can run it on their own computer, without filters or a monthly fee. The Wikipedia page on Stable Diffusion provides a useful summary of how it works for readers interested in the technical aspects.
Users can now type "a watercolour of a canoe on a misty lake at dawn" and receive four versions in under a minute. For the general public, the result appears remarkable. For professionals who create images for a living, the implications are of a very different nature.
What are AI image generators good at?
Concepts. This is the most accurate answer.
When an idea must be presented before money is invested in it, these tools are exceptionally effective: a mood for a brand, a draft layout for a poster, or five directions for a book cover, allowing a choice to be made before hiring a designer to produce the final version. Work that once required an afternoon of searching stock image sites and editing in Photoshop now takes ten minutes.
New users often find it difficult to stop experimenting, entering one more idea to see the result.
Where do they fall short?
Hands are a widely observed weakness. Text within an image is rendered as unreadable characters. Anything that must be specific, such as a particular building, a real product or a company logo, comes out close but inaccurate. Many images also share a uniform, glossy and over-lit appearance that is already easy to recognize.
Some of these issues are likely to be resolved, as the models are advancing quickly. Whether the uniformity will be resolved is less certain.
What is at stake regarding rights and artistic style?
This is a contentious issue, and a resolution does not appear imminent.
The models were trained on vast collections of images gathered from the web, including works by living artists who never gave their consent. Entering an illustrator's name in a prompt produces an image in that illustrator's style. For a working artist, this is not a technical curiosity; it means seeing one's livelihood reproduced by a machine that learned from that work without permission.
A few weeks ago, a piece created with Midjourney won the digital art category at the Colorado State Fair, prompting a strong public reaction online. Some of the criticism concerned the prize itself; most concerned what the prize represents.
Our position is as follows. We use these tools. We also consider that artists have a legitimate grievance, and the argument that the process is "just like a human learning from other art" is, in our view, too convenient. A human does not learn from millions of images in a week and then sell the result at scale through a monthly subscription. We expect the question to be brought before the courts.
For a small business, the practical consideration is simpler. Ownership of a generated image has not yet been established, nor whether anyone owns it at all. An image used in marketing may not belong to the business in the way a commissioned photo would. At this stage, we recommend against basing a logo or brand identity on a generated image.
Can a generated image replace a real photo of your business?
No. A prompt can produce a manta ray, but not your manta ray. We spent many years in the dive industry, first developing digital and social media strategies with dive operators, then with Ceningan Divers, the dive resort co-founded on Nusa Ceningan. Around the Nusa islands, manta rays are a major reason visitors come.
A generator can produce a striking manta ray: flawless, luminous and perfectly lit. It will also be a manta that does not exist, above a reef that does not exist, in water no one has ever entered. A diver viewing a dive centre's website perceives this, even without being able to explain why.
What leads to a booking is proof: the operator's boat, its guides and the actual reef on an actual morning, with slightly murky water and other divers in the frame. That photo communicates "this is what you will experience." A generated image communicates "this is what we would like you to imagine."
The same applies to a restaurant, a sugar shack in the Outaouais or a gîte along a river. Customers want to see the actual place, not a more attractive one that does not exist.
How can a small business use AI images appropriately?
- For concepts and moodboards before a real photo shoot or design project: recommended.
- For blog illustrations where no one expects a photo of something real: acceptable, provided the image is clearly an illustration.
- For backgrounds, textures and decorative elements: appropriate.
- For anything representing your actual product, location, team or results: not recommended.
This last guideline is the one most likely to be disregarded, and it is the one that causes the greatest loss of trust when a customer notices.
Authentic images of a real business still require a camera and a skilled professional, which is why we continue to recommend investing in photo and video before any generated imagery. The generators are capable tools, but they cannot visit your business on a Tuesday morning.