The best AI art generators in May 2023

The best AI art generators have unleashed a controversial new technology on the art and design worlds. Some people love them, and some hate them. Others are curious or are keeping an open mind, interested in seeing how AI image generators might fit into their workflows but with reservations about how they were trained or where they might take us.

Allowing users to turn text descriptions into images, the technology continues to evolve extremely quickly. Recent developments include the announcement of the integration of Adobe Firefly into Google Bard with a mechanism to identify work as AI-generated. One thing seems clear though, if you’re involved in art or design, you at least need to know about the best AI image generators and what they can do.

Below, we list the best text-to-image AI art generators that we’ve tried and we evaluate how they compare in terms of results, versatility, ease of use and cost. We gave each a range of different prompts to see how it produces different styles of AI art, from specific artistic styles to photorealism.

As these tools are developing so fast, capabilities can improve month on month, so we’ll aim to keep this guide up to date to reflect new developments. For tips on how to get the most out of the best AI image generators, see our selection of the best AI art tutorials and our basic guide to how to use DALL-E 2.

Best AI art generators

You know the best AI art generators are starting to become mainstream when Adobe gets involved, bringing the technology to its Creative Cloud suite. Although still in beta, Adobe Firefly is the most practically useful AI image generator that we’ve seen so far for most working designers.

Since it’s part of Adobe Creative Cloud, you’ll eventually have access to it if you already have a subscription for products like Illustrator and Photoshop. At the moment, you need to request access to the beta. Some Firefly tools, including Generative Fill are now available in Photoshop beta.

We’ve found that Firefly has certain limitations because it was trained only on assets from Adobe Stock and public domain content on which copyright has expired, but this provides some peace of mind from an ethical and legal point of view. Adobe promises that it’s developing a compensation model for Adobe Stock contributors, and Firefly gives generations a Content Credentials tag stating that it was created using generative AI.

We think the limitations of its text-to-image capabilities are more than made up for by the sheer practicality of the range of editing tools and easy integration with other Adobe products. It also has a much more user-friendly interface than many of the other best AI art generators. It comprises not one but multiple AI models for different uses: as well as the text-to-image generator, there’s a text effects model and a model to recolour vectors.

It can be used to quickly edit images using text commands to generate content brushes and generate variations of components. In Photoshop, you can use Generative Fill to make over images, using text prompts to generate new elements in non-destructive layers, and Firefly automatically adds details like shadows and reflections to fit the existing image. More tools are expected soon, including text-to-vector, text-to-pattern and text-to-template.

Midjourney initially had a more limited artistic range than some of the other AI art generators and was often used to create painterly fantasy scenes (Midjourney is the tool that won a fine art competition and which was used to create those ‘last selfie’ images that went viral last year). However, Midjourney is the AI image generator that has most rapidly evolved. Now on version 5.1 (perhaps higher by the time you read this), today it can produce stunning realistic images in a wide range of styles.

We found that the interface felt a little strange and counterintuitive at first because you have to request image generations from a robot on the Discord social messaging platform using commands (type ‘/imagine’ to introduce a text prompt), but it doesn’t take too long to get used to. In the public channels, everyone can see your prompt, and you can see theirs, which can be a good way to learn and get ideas for how the model interprets descriptions (on a $48/month  ‘Pro’ plan, you can generate images privately using the ‘/stealth’ command).

And there’s the main downside of Midjourney. There’s no longer a free version, so you’ll need to subscribe to use it. Basic membership, good for up to 200 images, costs $8 a month. Unlimited images will cost you $30 a month. The more expensive subscription plans give you a higher quantity of faster generations.

We’ve found DALL-E 2 to be the best AI art generator for those curious to start exploring the technology – and certainly the simplest and most straightforward of the more advanced tools. It’s well-known, so there are plenty of tutorials, it’s easy to use and it can produce stunning results, including photorealistic images with lots of detail. Getting started is super quick, the main text-to-image function is intuitive to use, and we find the actual process of image generation is relatively quick.

We found that DALL-E 2 is capable of producing clean images that tend to match what was described in the prompt including details such as camera lenses or artistic styles. This text-to-image generator may be a little too limited if you want more control, however. You’re currently limited in terms of the resolution and aspect ratio of the 1024×1024 canvas, and there’s no option to add negative prompts or the range of specific instructions that you can in Midjourney. On the other hand, the inpainting and outpainting editing features offer a lot of flexibility for editing, replacing parts of the image or “uncropping” an image beyond its original frame.

Best of all, DALL-E 2 gives users enough free credits to get started with plus a free top-up of 15 credits each month. They don’t go far but they at least allow you to experiment before you need to buy more credits.

The best AI image generators that we’ve mentioned so far require some form of outlay, at least eventually. The best AI art generator that can be used without paying a penny is the open-source model Stable Diffusion. That potential catch is that it requires a little bit more work. You can run Stable Diffusion locally on your own computer, but you’ll need a powerful-enough machine and you’ll need to know how to install it. Alternatively, you can run it online in several ways, including for free via Google collabs or you can use the model more easily but at a cost via Stability AI’s web app, DreamStudio

In either case, we’ve found that Stable Diffusion is a powerful and versatile AI image generator that can compare to DALL-E 2 in terms of accuracy and the range of styles it can produce. It has the edge in some styles and It provides more flexibility than DALL-E 2 in terms of aspect ratio (although this can cause problematic artefacts) the ability to upscale resolution to 2048×2048, the option to add a negative prompt and to ability to use the seed number to restrict the randomness of the image generation.

There are some other nice features, including a depth-to-image tool, which can infer the perception of depth in an uploaded composition and transfer that to a generated image. The UI isn’t quite as user-friendly as DALL-E 2’s and you’ll need to pick up a bit of terminology, but it’s a very flexible AI image generator. The code is freely available on Github and can be used by developers in their own apps. If you opt for the ready-to-go route of using DreamStudio, it costs $10 for 1,000 credits –enough to generate around 5,000 images using the default settings.

One of the biggest weaknesses of even the best AI art generators (aside from Adobe’s specific text effects tool in Firefly) is text. Tools like Midjourney have started to overcome the problem of human figures with six fingers, but the written word is still usually rendered as some form of alien hieroglyphics.

So it’s the ability to generate text that makes DeepFloyd IF so impressive. This AI art generator is actually an offshoot from the Stable Diffusion team, and it’s similarly free and open source. The imagery itself can sometimes have the edge over Stable Diffusion for certain styles (specify ‘Polaroid’ in the prompt, and the results actually look like a Polaroid). But what’s most notable is that you can add instructions in the prompt to include text in the image. For example, you can request a bar in a desert at night with a neon sign that says “open”, or Lisa Simpson holding a sign that says “Creative Bloq”.

Like with all AI image generators, we found it to be a little haphazard, and phrases still come out partly garbled at times. For some reason, the model has a tendency to add an extra ‘q’ to our name: ‘Creative Bloqq’. But still, this is the best handle of text in an AI image generator that we’ve seen, allowing you to put words not only onto signs but also into collages, fabrics, stained glass windows and more. It’s going to be a dream for meme creators. Deep Floyd IF is currently available as a demo with a research-only non-commercial licence via Hugging Space.

We think Craiyon is still the best AI art generator for anyone who’s just curious about the tech and wants to have a play around without the fuss of having to create an account or buy credits. Simply visit the website, type what you want, and Craiyon will start generating. The downside is that the results can still be total nightmare fuel, especially human figures, which tend to get terribly mangled. It’s also a little slow, and you’ll have to put up with ads unless you pay for a subscription.

This doesn’t mean that Craiyon is unusable, however. It can come up with interesting surprises, and it’s surprisingly diverse, which can make it an interesting tool to use for playing around with ideas. Its features are limited, however. With no in-app editing tools and no image-to-image generator, there’s not a lot you can do in the app itself, although you can now include negative prompts. Bizarrely, you can order T-shirts printed with your generations, which could be scary. Note that Craiyon was previously called Dall-E mini until OpenAI had a word with them.

CF Spark is a bit more than just an AI image generator. It’s part of Creative Fabrica, an online marketplace for crafts, fonts, graphics and other digital assets. There’s a completely free version of the AI art generator, but if you pay for a subscription, you can earn money from your generations when they’re sold on the marketplace.

We’ve found the generator to be one of the best AI art generators in terms of the range of different features that it offers. As well as the text-to-image model, you can generate texts, patterns, transparent PNGs and colouring pages. It also has its own prompt builder tool – yes, an AI tool to help you write your prompts for AI tools. There’s a very busy community feed (users had generated over five million images at the time of writing). Here you can see other people’s creations for inspiration, and you can see other people’s prompts and seeds, so you can try ‘remixing’ their work.

The downside is that there’s not a great degree of control over the generations, with only a limited number of pre-defined artistic styles (kawaii can be cute and some of the illustrative styles can be effective, but this generator isn’t great at photorealism, and it’s ‘hyperrealism’ style is not very hyperrealistic). Generating images using the free plan can be slow too (paying for a subscription allows you to generate images faster and run multiple generations in the queue at the same time). CF is adding more features all the time, so some of this may change.

The best AI art generators we’ve mentioned so far all create 2D images from text prompts. But a recent novelty from the makers of DALL-E is SHAP-E, which does the same thing for 3D. They’re not the most impressive 3D assets by any means, but it’s the best 3D AI image generator we’ve seen to date, and assets can be exported as .glb files to open in Microsoft Paint 3D o Blender or convert into STL files for 3D printers.

The idea is that Shap-E will allow architects to specify structures’ dimensions, materials and colours and game artists to create assets for virtual environments. It’s not quite there yet, but at the rate that AI art generators have advanced in the past 12 months, this will surely be something to watch. The model weight, inference code, and samples can be downloaded on the Shap-E GitHub page and you can install the code using the Python pip command (you’ll need an NVIDIA GPU and a high-performance CPU). There’s an online demo on Hugging Face.

Source link