SCORE
Google ImageFX
- Free and good for experimentation
- Can produce engaging images
But when it comes to turning text prompts to images, the company is a bit behind its rivals — at least judging by my testing of Google ImageFX, a free tool that uses its Imagen 2 model. I reviewed ImageFX alongside rivals OpenAI’s Dall-E 3 and Adobe Firefly, and ImageFX fared well in some areas, for example with photorealism and some more conceptual prompts like a lightbulb made out of spaghetti. But I also had lots of problems with distorted anatomy, results that didn’t produce what I wanted and, most annoyingly, innocuous prompts that were rejected because of Google’s overcautious nannying.
Enlarge Image
ImageFX rendered a doctor passably, but the stethoscope and other medical instruments are fraught with errors.
ImageFX tries to prod you along the pathway of creative exploration by processing your text prompt and turning various words or phrases into “expressive chips” — drop-down menus you can tweak. That can be helpful for newcomers trying the technology, but the quality of visual results was inconsistent.
Often, I had trouble generating realistic humans. Fingers and feet and limbs and faces were sometimes peculiar. When prompting for doctors, I got a pretty dour bunch of medical professionals — the opposite of Adobe Firefly skewing toward cheerfulness.
Inanimate objects had problems too. Logos were styled appropriately as 2D illustrations but weren’t graphically convincing. Across more than a dozen attempts to show a doctor, the obligatory stethoscope was never convincingly rendered. It was like a medical instrument from a parallel universe. When I requested a monster truck jumping over a school bus, I got a monster school bus jumping over a truck.
Like all the other AIs, ImageFX failed to count pool balls. Prompted for “There are six pool balls on the green felt of a pool table. A light above illuminates the scene,” ImageFX never gave me six. It sometimes added miniature balls, didn’t include the light above, and duplicated balls. Changing the query to show a single pool ball yielded a table with many.
But I did get good results in some cases, which is why I say you shouldn’t overlook ImageFX. It did the best at all services I tested at rendering the facial emotion required for this prompt: “A product photo with a large collection of cleaning products in a shallow box. The cleaning products are in front of a person who is frustrated at how much work they have to do.”
Google’s ImageFX created this somewhat peculiar image of a man playing pickleball. The man’s features are exaggerated, but the prompt requested a caricature.
For other prompts, though, I couldn’t figure out what triggered the block. Sometimes tweaking the prompt worked, but then I’d try generating again and it would be rejected. That’s the frustrating experience that’ll drive people away.
Rejection of anodyne prompts also was a problem for Google’s Gemini chatbot, which like ImageFX uses the Imagen 2 text-to-image AI model. Google shut down Gemini’s ability to render people after related problems, like the inability to generate images of Black people when requested. Google didn’t shut down ImageFX, which has a different text-processing system. (For example, Gemini can handle very long prompts that ImageFX can’t.)
How engaging are the images?
Generally, ImageFX produces engaging, eye-catching images. Its problems lie with the fundamental image elements, not with the flashiness of the presentation.
ImageFX often would come up with a style it found appropriate, usually with good results in my tests. Logos were punchy. When I asked for a collection of antique scientific instruments, it presented them with the quiet, dusty vibe of a museum. Prompts for Napoleonic-era British Navy scenes produced images in the style of an 18th century etching or hand-painted illustration.
Prompted to create an independent coffee shop logo with red, yellow and green colors and a brown background, ImageFX delivered some so-so results.
ImageFX is reasonably fast most of the time, delivering results in 10 to 20 seconds. Sometimes I’d get impatient and switch away, though.
At times, I had to click “generate” twice, because the first click seemed to succeed only in reconfiguring my prompt. Sometimes ImageFX failed after 20 seconds or so for mysterious reasons and tells you to try your prompt again. Sometimes after that wait, ImageFX just blipped out and erased my prompt as if I’d clicked its “start over” button.
Conclusion
ImageFX delivers on some of the promise of text-to-image AI, though results that were unreal or that didn’t match the prompt were a bit more common than with rivals. If you’ve never tried it, I suggest you give it a whirl. ImageFX has the right price and is a great place to fool around to get a feel for generative AI images.
Google has a major AI effort, though, so expect to see improvements.
Editors’ note: CNET is using an AI engine to help create a handful of stories. Reviews of AI products like this, just like CNET’s other hands-on reviews, are written by our human team of in-house experts. For more, see CNET’s AI policy and How We Test AI.
