Prompt: Red panda in the wild holding a lime t-sign that reads "ZDNET"
It's been a while since a new text-to-image generator shook up the generative AI space. However, the mysterious Red Panda generator has done just that, climbing up Artificial Analysis's Text-to-Image Arena leaderboards and beating out leading models. Now, the identity has been revealed.
Also: The best AI image generators: Tested and reviewed
On Wednesday, Recraft launched its newest model, Recraft V3, the same model that appeared as Red Panda in the Arena. The leaderboard results show that the model can generate high-quality images with impressive details, quality, and prompt fidelity. However, according to Recraft, its standout is its text generation capabilities.
Some prompts -- such as those involving hands, faces, and text -- are particularly challenging for image generators. Most attempts at text image generation fall short, getting close but messing up one letter, spelling, or making up random words. However, Recraft claims its model can generate anatomically correct images and accurate long text strings.
"The main advantages of Recraft V3 [lie] in text generation quality, anatomical accuracy, prompt understanding, and high aesthetic quality," said Recraft in a blog post. "Recraft V3 is the only model in the world that can generate images with long texts, as opposed to just one or a couple of words."
To see if the claims hold up, you can test Recraft V3 for yourself using the instructions below -- or scroll down to see how it fared on my tests.
How to access Recraft V3
The model is available for free and paid users online and in the mobile app. Getting started is easy: All you have to do is visit the website, click on "Generate AI image," and create a Recraft account or sign in with an existing Google, Discord, Apple, or single sign-on.
Also: Can Apple's new AI photo 'Clean up' tool beat Google's Magic Editor? For some users, yes
Different plans are available to better suit users' needs, starting with a free plan that offers 50 free credits daily and makes all generated images public. The more advanced plans offer higher limits and more advanced features and range from $10 per month to $48 per month.
Once you're in, click on "Create new image," type in a prompt, personalize the settings, and click on "Recraft."
A few results
It took 15 seconds to generate two images. I tested for quality for the first generation using the prompt, "A vibrant, realistic hummingbird perched on a tree." The results were very impressive and comparable with some of the best image generators' takes on the prompt, which you can see on this list. I included one image below.
For the next prompt, I went for something more challenging – hands. I entered the prompt, "Two manicured hands typing on a laptop." The images look OK at first glance. However, when I take a closer look, I can spot some inconsistencies.
Lastly, for the most exciting prompt and biggest challenge, I asked it to generate an image of a computer screen that read ZDNET's brand's mission statement, "ZDNET, tomorrow belongs to those who embrace it today," in electric yellow. I included both results for this one because they were equally impressive.
Not only was all of the text accurately spelled and transferred, but it was also uniformly displayed and spaced out as if a human had placed it there. It was also layered very nicely onto the backdrop, resulting in realistic photos that look like they were taken by a camera. If you look closely, there is some variation in the uppercased words, but that is minimal compared to text results from most other generators that can't even get the letters out.