What is an AI art generator?
An online artificial intelligence app that creates an image for you, either from some text you write (know as a “text prompt”) or from some other input, like uploading an image or changing numbers with a slider.
New AI Art generators are appearing almost daily now, and its really hard to know what the differences are and which one is best for your ai arts needs so I’m here to clear all that up for you. I’ve spent hundreds of hours generating AI imagery and I’ve used almost all of them. I just can’t stop.
I will be making videos demonstrating how to use them as well as providing sample images wherever possible.
How to use AI Generators
Almost all of them, certainly the Text to Art ones, work like this:
1. You enter a few words describing what you want the AI to create (referrred to as a “text prompt”) hit enter and after a few seconds you get back an image or several images.
2. If you don’t like the output you can change your text prompt and try again. The images will be different every single time.
3. Download the ones you like.
That’s basically it, the rest is all in the dozens of things that are particular to each one, like image size, subject specialties, tweaking options.
With that out of the way, here’s the scoop on the top AI Art Generators:
MidJourney
Pros:
- Tons of (optional) parameters to tweak if you want them.
- Beautiful images with an emphasis on aesthetics and style.
- David, the founder has a strong vision for MidJourneys future as a tool for artists.
- Searchable gallery app with great user interface (UI).
- Image prompt input option to (optionally) generate images using an existing initial image.
Cons:
- No dedicated image creation app. All image generation must be done in the Discord app (!)
This is the best AI Image Generator for artistic imagery. I still use most of the others, but I always keep coming back to MidJourney as my #1. If a beautiful image is more important than a more literal, photo-realistic image, then this is definitely the one for you.
It can generate images in many different styles from 3D/CG to painterly illustrations and it has a ton of ‘tweakability’ if you want more control over your images.
MidJourney has asked that people not make detailed tutorials on how it works since it’s still in its early stages and will still change a lot. In the few months that I’ve been using it it certainly has, but I can’t wait anymore so I’ll just show you a quick overview.
To generate an image, you enter a text description of what you want and you get back four thumbnail images. You can add one or more modifiers to your text to attempt to control the output or give it a reference image to attempt to influence it.
One slightly clunky difference from the others is that MidJourney does not have a dedicated image creation app, the only way in order to generate images with MidJourney is to join their Discord Server and post your text prompts into the chat (like you seee above)
All your generated images become public and are mixed in with stream of images and chats but they also appear in your dedicated library in their beautiful gallery app.
This kinda awkward, non-interface almost stopped me from using MJ, but I am so glad I pushed through my initial discomfort. The rewards are mighty.
Here is what MidJourney gave me when I asked for “slabs of marble leaning against a wall”
Once the thumbnails are rendered (usually in less than a minute) you can choose to ‘upscale’ any one of them by tapping on the number corresponding to the thumbnails (U1 to U4)
If you don’t like any of them you can hit the redo button or the up arrow to re-enter your original prompt and modify it before you hit enter to try again.
I chose to hit U4 to upscale the fourth thumbnail image and this is the result:
Once “upscaled” users with a paid account can choose to upscale again to ‘max’ resolution, or use their other upscaling options to get a different ‘look’.
The image above (also here) is what it looks like when it is “upscaled to max” which in this case, a standard square image, is 1664 x 1664 pixels.
Unlike most image generators, MJ is capable of creating images in any aspect ratio you want. Square, landscape, portrait, ultrawide, ultra tall.
To make a 16 by 9 widescreen image you would add “–aspect 16:9” to your text prompt.
MidJourney offers a constantly growing list of modifiers you can use to try to steer the image generator to make what you are looking for. I will cover this fully as soon as I can. It is one of MidJourney’s unique features and it continues to evolve weekly.
If you purchase a paid account and then also add the privacy option at an additional cost, you can create your images by sending your text prompts in a direct message (DM) to their chatbot and they will remain private and hidden from the public feed.
Despite having no actual image creation app (all done in Discord) MidJourney does have a beautiful, fast, searchable image gallery where you can see and sort all your creations.
The fact that all images are public by default is very intentional. The founder, David talks about this as part of the Midjourney philosophy in his weekly office hours live voice chats on the Discord channel which are well worth listening to.
The thinking is that with all images being public, users learn from each others’ results and text prompts. Also, “…as soon as they become private, it becomes all about work” adds David who is a big fan of creativity over commerce. In fact, the corporate account gives you less images than the personal account and that is intentional.
The result is that the Discord channel is an incredibly active community of almost 1 million registered users generously sharing their images, knowledge and tips.
David and his thoughtful art focused approach to AI Generation is another reason Midjourney is my favourite.
Sample text prompt generated in MidJourney: “giant troll, approaching a small town, seen from below, octane render, 8k, extremely detailed, cinematic lighting –aspect 16:9”
There is always some new experimental feature being tested. Currently it’s a photorealistic option. Here’s a sample photo portrait I generated with it
You can see some more AI generated faces and photo portraits here
DALL-E 2
Pros:
- Photorealistic image output.
- Simple UI with a gallery on the side.
- Fast
cons:
- The images can be ‘flat’ and lack the ‘character’ and style of MJ or Stable Diffusion. It’s output ‘feels’ like it was trained on a lot of camera phone photos rather than artistic, professional photos.
- Fixed, square image size and and fairly low resolution at that. No up-res, landscape or portrait options.
If using MidJourney is like asking your arty friend “Hey, could you illustrate something for me?” Dall-E is like asking “Hey, do you have any photos of…?”
Backed by well funded Open AI (an Elon Musk company) Dall-E 2 is probably the best known of all the AI image generators.
It couldn’t be easier to use. Just enter your text prompt and in about 7 seconds, Dall-E presents you with 4 images. Download the ones you want at 1024 X 1024 resolution. Here’s one from the above example.
The outputs are a lot more photorealistic but that also makes them flatter, with less style.
Here is the second image above at full resolution
You can of course ask Dall-E to create something in any style you like by adding any of your preferred art styles to your prompt, like “impressionist”, “Rennaissance” or “pencil sketch” or simply add something generic like “digital art”
Here’s a more digital art style example, of a giant troll with the full size download at 1024 x 1024, it is soft though and appears to me to be more like an enlarged 512 x 512. Image size is an issue with most of these image generators and creating work with resolution good enough to print is difficult if not impossible.
As well as image generation, Dall-E 2 also allows you to do image editing.
DreamStudio
DreamStudio is a very tweakable Stable Diffusion AI art generator and uniquely, it allows you to generate up to 9 images at a time. At a cost of course, you pay for each image.
Tweakable settings:
Both Width and height are selectable, from 512 pixels to 1024 pixels with any combination allowed for square, portrait or landscape images.
CFG scale allows you to choose how much importance it puts on your text prompt. Sliding it to the right text your prompt. Literally, while sliding it to the left may result in it basically ignoring your prompt all together.
Steps (from 10 to 150) is the setting that will have the biggest impact on your wallet and also your creativity. It’s a slider to increase the amount of time stable diffusion spends processing and generating your image. More steps usually equals more detail but it also costs more.
More detail is not always a good thing, and you can get some great images with just 10 steps which will get you trying more prompts, and the more prompts you try, the better you get at creating great prompts and images. I’ve also seen some terrible outputs even at 150 steps.
Using the step slider, you can quickly generate a lot of drafts at the lowest, cheapest setting and then clicking in the middle of the image you can copy the seed image which will allow you to reproduce it at a higher step value as long as you don’t change the size or aspect ratio.
Like Riku, Dreamstudio is one of the few Stable Diffusion art generators that allow you to choose which Stable Diffusion sampler method to use.
Start using DreamStudio here
Riku.AI Image Generator
Riku is known for making it easier for people to connect with AI apps by connecting you to any number of AI engines like Open AI GPT3, AI21, Cohere, (with paid accounts) or unlimited GPT-J (for free) via a simple user interface. They are helping to make a link between you and the AI apis. If that sounds confusing, don’t worry, because their latest product, the AI image generator, just in Beta as of yesterday, is very easy to use and already allows you to choose which Stabe Diffusion rendering method you use. For more details see my Stable Diffusion sampler comparison here.
Right now they just offer Stable Diffusion (that is plenty for most) but the plan is to allow you to select from multiple AI image models like VQGan and others. NightCafe currently offers this kind of choice already.
Pros
- Multiple Image formats available (square, portrait, landscape)
- Riku’s founder Stuart Lansdale is committed to providing unique and smart services to Riku users, with emphasis on accessing to the finer controls so expect Riku to continually innovate and add their unique flavour to the generator(s)
- Choose your technology/API. Currently only Stable Diffusion but Riku will add other APIs as they go.
Cons
- Can’t set number of generations (Yet!)
Only available in closed Beta for now
Photosonic by WriteSonic
This is a new arrival from the AI text generation app Writesonic.
It looks to me to be using the Stable Diffusion engine via API. Just launched today (August 26th) their is still in Beta and allows visitors to generate 10 free watermarked images. No paid accounts are available as of yet.
Full size (512 x 512 pixels) of image #2 can be seen here.
“Full resolution” image generated by Photosonic using the prompt “one large cube of ice floating in the air over a calm pool of water, at night, octane render, 8k, extremely detailed, cinematic lighting”
While I had some free credits I thought I’d try a more ambitous prompt and ask for the Giant Troll. (you can see what prompts I gave the AI in the screenshot images)
This is the full size image of the troll (512 x 512 pixels)
You can now purchase credits to use Photosonic at 100 image credits for $10
Check out Photosonic here
NightCafe
One of the older more established AI art apps, NightCafe allows you to choose which AI algorithm to use Coheret, VQGan+CLIP and now, as of August 2022 also Stable Diffusion.
I asked NighCafe for the Troll, using the same prompt I gave the others and choosing the Stable Diffusion option I got this:
Definitely a beautful image, but hard to identify as a Troll approaching a small town. But that’s how image generating goes. You often have to roll the dice a few times before you get something you want to keep.
Dream by Wombo
This fast (and light) Android and IOS app is great for quickly trying out text prompts. The output is very low resolution, but its fast and its free!
Playing around with Wombo is a really great way to learn how to communicate with AI via text prompts on your phone.
Wombot
In September 2022, Dream launched Wombot, a discord channel in which you can enter prompts to generate images, just as you do in MidJourney.
After a few seconds, your image will show up in the channel, tagged with your @ username
It’s free to try and a premium subscription gets you: “features like custom resolutions, using WOMBOT in DMs, NSFW image generation, and more.”
You can join and try Wombot yourself here
PhotoLeap
This IOS only app was mainly for retouching your photos (selfies!) until this week when they added an AI image generator to the app (also using Stable Diffusion)
I tried the Troll text prompt a few times and got these:
Wonder AI Art Generator (APP for IOS and Android)
Wonder is an app for your smartphone, much like Dream by Wombo, but it’s images are more stylised and also higher resolution. The ones I saved on my phone and later downloaded to my computer were all 1024 X 1024.
The preview shows them in a more dramatic and eye pleasing Portrait crop, but that can be a little bit confusing as just moments before, you see the full image in a preview, then suddenly it’s cropped but never fear, the full image is there when you download it.
Here are a few images I generated in my first hour with Wonder
There are multiple AI models I E styles to choose from and they have a huge impact on the output. This is both good and bad. It means you’re almost guaranteed a good-looking image, but that also means the AI is putting priority on creating an image with a certain look sometimes at the expense of your text prompt. It is biased in favour of the style, often prioritizing it over your text description.
Beautiful images pour out of this app regardless.
Get the wonder app for IOS here or Android here (Play Store)
The next AI art generators are a bit different from the text to image generators above.
ArtBreeder
This is an older favourite among artists and has several different ‘modules’ that work more by moving sliders than any particular input from you. Each module generates a particular type of output.
While most AI Art generators struggled to make distorted and warped faces, Artbreeder excelled at it. What it lacked in text to image generation it made up in beautiful character and landscape images that you create and tweak with sliders. Very addictive sliders.
Add “parents” to breed two or more images together to create infinite variations of people. As you slide the characteristics sliders the face changes almost instantly.
Artbreeder has a number of other modules that let you create landscapes, general art and or create your own ‘art genes’ for countless hours of tweakable fun.
Paid accounts allow you to generate more images and even create fairly lengthy animations morphing from one landscape or character to another.