When generative AI tools first reached the public in early 2023, the excitement was tied to two new capabilities: long-form text generation and photorealistic image creation. Three years later, models have evolved from party tricks into practical tools for marketers, designers, and everyday users. Some image generators can now repair old photographs with surprising accuracy; others still struggle with hands, text, and cultural references. This hands-on comparison was built to separate the genuinely capable services from those that only look good in demos.

Six paid AI image generators were tested with the same nine prompts. Each generator faced 30 scoring checkpoints, producing more than 50 images in total. The tests covered four broad areas: editing or recontextualizing existing images, creating original images from text, adding readable text to images, and responding to pop-culture prompts. All results were generated in a 1:1 square ratio to make side-by-side evaluation easier.

Key takeaways

• Google’s Nano Banana Pro earned a near-perfect score of 93%.
• OpenAI’s ChatGPT image generator finished second with 74%.
• Midjourney, Adobe Firefly, Leonardo AI, and Canva scored 57%, 54%, 52%, and 43%, respectively.
• Logo generation was the easiest task for every tool.
• Text rendering and facial preservation were often the weakest points.

How the tests worked

The evaluation began with an existing photograph of the reviewer walking outside. The first prompt asked for a recontextualized version with the subject dressed as a US Navy admiral on the bridge of an aircraft carrier, without changing the face. The second prompt asked each AI to restore an old black-and-white photo of a child, preserving facial features, clothing and period detail. The third prompt asked for restoration and colorization of an old truck photograph, with attention to original markings. These three tasks measured editing, restoration, and text accuracy.

The next set of tests focused on creative generation. One prompt required a retro-futuristic logo for a video studio called Space Coast Studios. Another asked each model to create a fantasy scene of a medieval librarian in a candlelit stone library. Two social-media-style prompts were included to test text overlays. One used a senior adult smartphone user; the other used a student near a window with a laptop and coffee. Finally, two pop-culture prompts tested each model’s understanding of cultural references: a fictional fourth Back to the Future movie set in 1920s New York, and a Nightmare Before Christmas-inspired image of an IT professional in a data center.

Google Gemini 3 Nano Banana Pro: Best AI image generator overall

Google’s Nano Banana Pro, available through Gemini 3 on the Google AI Pro plan, won every major category in this comparison. It scored 93%, a result that seemed almost implausible until the side-by-side images were examined. The image editor handled recontextualization with ease: when an original photo of the reviewer was fed into the system, Nano Banana Pro kept the face, glasses, and general posture while putting the subject in an admiral’s uniform on a ship’s bridge. It even changed the angle of the face to suit the new scene. The result looked like a plausible studio portrait, not a crude composite.

Old-photo restoration was equally impressive. A dark childhood image was cleaned up and sharpened while retaining the characteristic hair and clothing of the original. Colorization of a decades-old truck photo was more complicated. Nano Banana Pro first produced a colorized version of the earlier child photo instead of the requested truck, and then placed the reviewer in an admiral’s uniform in the truck image. Those misfires cost points, but the eventual restoration of the truck was strong. The truck lettering was not perfect; what should have read RADIOLOGICAL DEFENSE appeared as BADICLOGICAL DEFEKSE. That remains a reminder that text is still difficult for most AI systems.

Where Nano Banana Pro really pulled ahead was text generation. The logo for Space Coast Studios looked like a genuine brand mark, with a palm tree, rocket, and film imagery that echoed both Florida and video production. No text errors appeared in that sample. Social media prompts produced clean headline text with proper layout. The AI even rendered a smartphone in the senior portrait, though it chose an iPhone rather than the generic flagship phone requested.

The pop-culture category produced the most surprising result. For a fictional fourth Back to the Future movie, Nano Banana Pro created a poster with the familiar logo treatment, a DeLorean, a skateboard, and a 1920s-style New York marquee. The tagline was original and evocative: The Roaring Twenties Got Heavy. For the Tim Burton-inspired data center prompt, the service matched the intended aesthetic while keeping the IT theme visible. The only real annoyance is a watermark in the corner of each generated image, which can be removed but should not require extra effort from paying subscribers.

ChatGPT

OpenAI’s image generator inside ChatGPT ranked second with 74%. It was tested after the arrival of the company’s new ChatGPT Images model, which was tested separately from the earlier version. The new model improved on photo restoration in some ways, but it still exhibited clear weaknesses. The naval recontextualization captured the reviewer’s face, but the uniform was less convincing. Rank insignia were mixed, and the clothing did not match a specific admiral’s uniform with confidence.

The childhood photo restoration was better than earlier attempts; the original face was preserved rather than replaced with an adult’s face. The truck restoration was also improved, but the text changed from RADIOLOGICAL DEFENSE to RADIO CHEMICAL DEFENSE, with an extra space between words. The logo and fantasy librarian samples were solid, and the social media text was correctly placed, although low contrast in some places reduced readability. The biggest failure came in the pop-culture section. The service refused to complete a Back to the Future-style image and a Nightmare Before Christmas-style image in the originally requested form, losing significant points. Work from the previous model showed better cultural responsiveness, but also produced a third hand on the IT professional. ChatGPT remains strong for natural-language iteration in chat, but it is not a dependable choice for every image-editing need.

Midjourney

Midjourney finished with 57%. It is known for highly creative, cinematic imagery, and that strength appeared in the fantasy librarian and the data center prompt. The medieval library was close to the exact image described in the prompt, and the IT professional image was one of the most visually appealing outputs in the entire comparison. For original artistic work, Midjourney remains a favorite.

Its weaknesses were visible in editing tasks. The recontextualized admiral image was not the reviewer at all; the face was wrong, and the background included a ship in a park. The old-photo restoration did not clean up the image and in some ways made it worse. The truck request produced a different truck from the source image and added a random woman who looked like a science-fiction character. Text generation was also missing: Midjourney did not attempt to render the social media headline text in the requested format. For pure imagination, the tool is still excellent. For controlled, practical edits, it is not.

Adobe Firefly

Adobe Firefly scored 54% and was the most restrictive tool in the group. The service refused to upload an original photo that contained a Star Trek-inspired T-shirt graphic, and it blocked a prompt that used the phrase Facebook-style because the AI safety system identified the word style as a possible artist reference. Removing the word allowed the prompt to proceed.

Firefly created an acceptable logo and generated a reasonable social media image of a senior adult, though it misspelled the headline text. A student image had several subtle anatomical problems, including an odd arm and unconvincing hands. The admiral recontextualization did not resemble the source subject. The childhood photo restoration cleaned up well, but the truck colorization ignored existing artifacts. The most restrictive limitation was its refusal to generate pop-culture-inspired images, costing it several points. For commercial users who need safe, brand-friendly output, Firefly is a defensible choice, but its guardrails can get in the way.

Leonardo AI

Leonardo AI finished at 52%. It bombed the photo-editing tasks, declining to participate in recontextualization and restoration in a meaningful way. In creative work, however, it showed moments of real strength. The logo was clean, although it lacked the same regional theme other services produced. The fantasy librarian met the prompt requirements and allowed the inclusion of either gender or age, demonstrating flexibility.

Social media images were less successful. The man’s arm had an odd inner elbow, and the student image placed coffee on a leg in an unnatural way. Leonardo turned the Back to the Future prompt into an actual poster format, but the movie year listing was bizarre: it appeared to reference Back to the Future 4 through 10, and the text degraded into gibberish. The best Leonardo result was the Nightmare Before Christmas-inspired data center worker, which used a photorealistic character that felt genuinely in the spirit of the request.

Canva

Canva’s internal image generator came in last with 43%. It produced a serviceable logo and an acceptable fantasy librarian, but the photo-editing portion went badly off course. The recontextualized admiral image appeared to place the subject on an aircraft carrier with trees growing out of the deck. The truck restoration omitted the truck entirely and placed a floating woman’s head near a strange industrial tower. The social media prompts produced text that was poorly spaced or disconnected. Canva’s image generator is still suitable for users who need to create quick marketing images within the Canva ecosystem, and its Pro and Business plans include other design tools, but it is not a reliable standalone image editor. Its Back to the Future-inspired image contained nonsense text and an unidentified girl, and it denied the Nightmare Before Christmas prompt after allowing a poster for the other franchise.

What the tests reveal

Several patterns emerged from this comparison. Photo restoration remains difficult for most AI tools because it requires preserving identity while fixing damage. Text rendering continues to separate best-in-class tools from also-rans. Pop-culture prompts are unpredictable when the safety filters of a model intervene. Across the board, logo creation is now close to solved, and simple fantasy scenes are generally safe.

The clearest takeaway is that paid AI image generation no longer needs to involve guesswork. Nano Banana Pro is the only model that handled recontextualization, restoration, text, and style without a significant miss. Its 93% score was backed by visible results across all nine prompts. For the subscription price of $19.99 a month, it is now the default recommendation for anyone who needs reliable AI graphics.

Have you tested any of these image generators? Does your own experience match the winner? The choice may depend on whether you value photo cleanup, text accuracy, creative style, or cultural references. If you have a favorite tool from 2026, share it with others who are trying to decide where to spend their subscription budget.

Source:ZDNET News