AI image generators have evolved to offer artists most control over the results.
2024 was a relatively subdued year for AI image generation, at least when compared to the attention they gained in 2022 and 2023. Video generation, buoyed by high-profile releases like OpenAI's Sora, became the hot topic.
"I think a lot of tools are seeing the direction AI content creation is going, and just trying to stay up to date, trying to diversify [into video]," says Ross Symons, co-founder of creative AI company ZenRobot. "But I find the tools that are sticking to what they're good at, like Midjourney, are dominating."
Most modern AI image generators continue to use a diffusion model. This type of AI model is trained by teaching a neural network to predict random noise as its gradually added to images. Once training is complete, the process can be reversed to create unique images from random noise. The diffusion model is then paired with a text model, which converts the user's text prompt into tokens that guide the diffusion model's results.
Diffusion models have improved since 2022 but haven't fundamentally changed and, in the past year, the spotlight shifted from the models themselves to the software tools they power. The best AI image generators currently are quick, easy to use, and offer ways to more finely control image generation.
While there's literally dozens of image generators available, Midjourney is the clear frontrunner.
"Midjourney is a very important tool to us," says Chrissie Cremers, co-founder of Aigency Amsterdam. "It still excels in art direction." Symons agrees, saying he "predominately" uses Midjourney.
Released in 2022, Midjourney quickly gained a reputation for detailed, appealing images with a strong artist style. It tends to generate images with vivid, bold colors and a dramatic look that grabs a viewer's attention. Midjourney's diffusion model is proprietary and only available with a subscription to its titular online service.
Creative professionals like Midjourney's user interface which, given the tool's origins, is a bit ironic. The model was initially available only through a Discord chatbot connected to a slapped-together user authentication system. That rapidly changed, however, and Midjourney now offers a beautiful web interface that stands apart from its peers.
While Midjourney is still in favor among AI-powered creatives, 2024 saw the release of one major diffusion model that fills a different role: Flux. The model was created by Black Forest Labs, a company founded by three former employees of Stability AI (the company behind Stable Diffusion).
Symons cited Flux as the most notable new image generation model in 2024. "In terms of realism, it's amazing," he said. Cremers agreed, saying "It's very well suited for photorealism, much more so than Midjourney." Flux also offers strong tools for "in-paint and out-paint," which allows the model to edit images by removing unwanted objects or expanding the image beyond its original scope.
In another contrast to Midjourney, Flux is an open model, meaning it's freely available for anyone to obtain, use, and fine-tune. It can be installed locally to generate images on powerful home and office computers.
Founded in 2022, Leonardo.ai moved quickly to establish itself as a leading choice for individuals and companies looking for professional image generation.
It has largely focused on creating tools that use existing generative models in new ways, though the company did release its first in-house foundation model, Phoenix, in 2024. Leonardo.ai also offers Flux, and it ranks among the better options for those unable to run Flux on their own hardware.
"For character consistency, if you are generating characters that are illustrated or fall in the fantasy space, or in 3D, I think it's good," said Symons. This is achieved with support for Low Rank Adaption (LoRA), which lets users tweak a diffusion model to meet their needs.
Despite that, Dall-E remains among the more prominent tools, especially for amateur use. Now a part of ChatGPT, Dall-E generates passable images quickly. An April 2024 update added image editing, as well, to help users remove or change specific areas of an image generated by Dall-E.
Google Labs, the company's home for its "AI experiments," includes an image generator called ImageFX. Though not widely advertised, and a bit barebones in comparison to other tools on this list, ImageFX has earned a reputation among AI creatives in the know.
"It's a bit of a dark horse. It's not talked about, but it's a good tool. It's free, you can generate four images at a time, and the images are of seriously good quality," said Symons.
ImageFX has a simplistic interface that encourages users to input short prompts. It also provides a list of keywords to guide the user's prompt towards text the model can easily understand. This approach makes it easy to generate detailed images with a specific sense of style. ImageFX is entirely free, too, which is a plus for more casual users.
While the image generators mentioned above stole the show in 2024, they aren't the only tools worth paying attention to. Magnific has a reputation for excellent image upscaling. Krea.ai offers attractive interfaces for third-party models, including Flux.
Creative professionals seem less enthused about image generators from established companies. Getty Images, Shutterstock, and Adobe have generative AI services, but they lack the buzz surrounding leaders like Midjourney and Flux. Cremers blamed the images used for training. "They tend to be trained on a lot of stock imagery, and that reflects in the outcome."
2025 seems likely to be a repeat of 2024. Video generation will continue to take the spotlight as several companies -- most notably Google -- have yet to release their highly anticipated models. Even so, Symons thinks creative professionals will continue to find new ways to use image generators as they become more familiar with the tools.
"I wish people understood the capabilities of AI more. There are certain things you can do, and things you can't," said Symons.
