An experiment in using AI and automated image tools to create usable images for blog posts, social media and other common use cases.
When you’re sharing content or promoting services, it’s always useful to have a collection of images to use.
We’ve been pondering some alternatives to the sometimes frankly meh stock images available and wondering if the new breed of AI image generation might be the answer.
There’s quite a few machine learning-based image tools out there, but are they really as good as they seem? What kind of sneaky terms of use do they include?
This is what we found.
1. DALL·E 2
The release of a new image generation system DALL·E created a lot of buzz online and is perhaps the best known of the first generation of online image generation tools.
Described as ‘a new AI system that can create realistic images and art from a description in natural language’ there are some pretty neat demos that prove the point.
Developed by OpenAI, their bold aim is to ‘ensure that artificial general intelligence benefits all of humanity’. However, potentially platforms like these may not be good news for everyone, as this article from The Next Web explains.
There’s still a waitlist to access DALL·E 2 at the time of writing. We received an invite after about a week but registration strangely requires verifying a phone number so that’s something that might be a deal-breaker.
You get 50 free credits when you first start. Each ‘DALL·E request’ such as a text prompt for the image generator returns 4 images and uses 1 credit. So that’s something to keep in mind when you start testing the platform.
The interface is by far the simplest to use and I especially appreciated the ‘tips’ that appear on the loading screen.
And the results were quite impressive right from the start.
Our (admittedly ridiculous) text prompt we used for comparison returned some decent results.
But what about actually using these images? Users are now granted full usage rights to images they create, including commercial usage. So this means you can use these on social media and your blog posts, which is our main test case.
All images returned are 1024×1024 without the option to modify the aspect ratio or size.
So these images are not as immediately useful for banners or hero images but the results are potentially useful for other kinds of content. Having said that, they have just announced the new Outpainting feature which lets you add to the image ‘beyond its original border’, a promising addition to the editing options.
2. Midjourney
Midjourney has Just been recently released in Beta and is available via the Discord platform. That’s quite a limitation currently – you do need a Discord account to use it. They reference a web-based version but it wasn’t something we were able to access during this beta period.
There’s a free trial that provides a limited number of prompts, so you can try before you buy. We took the Midjourney bot for a spin via the various newbie channels (type the /imagine command to submit your text prompt).
As with other AI image generators, additional keywords specifying style and other preferences helps you get the results you want.
You can also include the aspect ratio you prefer. For example, for our blog images, we were looking for landscape orientation header images so we used the following prompt:
Amigurumi Librarian –ar 16:9
It’s not as user-friendly as Dalle-e and probably most useful for those who are already comfortable with using Discord (though their documentation is impressive).
Despite the slightly steeper learning curve, we found even our tentative initial results returned high quality images. And the Aspect Ratio (and other parameters you can set) are definitely useful.
You can now add the bot to your own Discord channel too which helps avoid the busy Midjourney discord server when you’re first starting out.
3. This Person Does Not Exist
For something different, we briefly took a break from our ‘Amigurumi Librarian’ mission.
This Person Does Not exist is a site that uses machine learning to create realistic, but completely fake human faces. So basically what it says on the tin. And sure, this is quite a narrow use case, but this can occasionally be useful for times when you don’t want to get into messy territory of using images of real people.
We like them for our user personas, for example. But most other use cases feel a bit creepy and it’s hard to ignore the more troubling potential of these kinds of applications of AI.
It’s also quite difficult to find out if there are any restrictions on usage for this tool or anything much about its origins. It was apparently built by a software engineer by the name of Philip Wang using research released by chip designer Nvidia.
There are no customisation options or anything like that, just refresh the page to get an image of a new ‘person’.
And during our extensive testing (/s), we also stumbled across thiscatdoesnotexist.com. We have zero use case for this version but we are now working hard to come up with one.
4. WOMBO dream
Dream is an app by Wombo which is a bit more art focused than the more general AI image generators above. Rather than a keyword prompt that often includes a particular image style, Dream includes a combination of keyword and an art style selector.
As well as the apps, you can try it out on the website.
We enjoyed the ‘art style’ option with this generator. Something we noticed people using quite a lot as a keyword too.
You could choose from a wide range of styles, from Realistic, Comic, Line Art, Steampunk, Mystical, S.Dali and many more. You also have the option to select ‘No Style’ which, for the sake of scientific enquiry, is what we did.
The images generated were definitely less impressive than Dall-e and Midjourney, retaining some of that AI creepiness factor particularly when it comes to ‘realistic’ people.
But for more creative outputs, this was a fun tool.
It gives you the option to ‘Mint your art as NFTs’ (eyeroll) or to buy a high quality version of your new artwork. If you want to.
5. Stable diffusion
You may have heard about Stable Diffusion as an (open source!) machine learning model and alternative to DALL·E.
Stable diffusion image generation is available via a few different websites. We started Hugging Face – which returned an error of ‘This application is too busy! Try again soon.’
We switched to the DreamStudio open beta, available via : https://beta.dreamstudio.ai/dream
Again, once you have an account, you will be given some ‘complimentary credits’ to try it out. And DreamStudio Lite offers some image settings before you enter your text prompt and hit ‘Dream’. You can set a preferred width and height as well as modifying the Cfg Scale which adjusts how much the image will be like your prompt.
It’s early days with this front end but the results were promising. And the licensing is a bit clearer than some other tools with the FAQs clearly stating that images made are assigned a CC0 1.0 Universal Public Domain Dedication.
So though our Amigurumi Librarians had gone a bit askew, they do get bonus points for clear and fair licensing.
Verdict
While the summer of AI images may be down to the increase in more user-friendly and accessible web interfaces for these tools there is still quite a variation in what these tools can offer.
Admittedly, this was just an initial experiment and a chance to generate some cute (albeit stereotype-laden) images.
It may just be our keyword efforts which we kept simple as part of the testing, but the AI image results can often be just a tiny bit creepy. Particularly when they include ‘people’.
But with new features being introduced at a rapid rate and with the recent announcement of Google’s new text-to-image diffusion model, DreamBooth, there’s definitely a lot of interesting things happening in the world of AI image generation.
It seems that the use of these tools to generate unique and usable images for blog posts and social media sharing is a future that is not too far away at all.