Best AI Image Generator: ChatGPT Images 2.0 vs Nano Banana vs GPT Image 1.5 on 50+ Prompts
Pat Simmons ran about 50 prompts through ChatGPT Images 2.0 next to GPT Image 1.5 and Google's Nano Banana. Images 2.0 won photos, maps and infographics, and Nano Banana won handwriting and receipts.
No single AI image generator won every round when Pat Simmons ran about 50 prompts through ChatGPT Images 2.0 on the day it launched (tested 22 April 2026) and set most of them beside the same prompt in GPT Image 1.5 and Google's Nano Banana. Images 2.0 won the realistic phone-style photos, the maps, the UI mockups and the infographics, and it drew a clock at 3:33 when Nano Banana's showed four o'clock. Nano Banana was still better at handwriting and receipts, and on posters and embroidered patches all three came out "kind of the same." GPT Image 1.5 lost almost every round.
For a free option, ChatGPT's Free plan makes images, the free version of Google's Gemini app makes them with Nano Banana 2, Meta's Muse Image is free for everyday use, and Canva Craft is open to every Canva user. Pat hasn't put Muse Image or Canva Craft through the same prompts yet.
| What Pat asked for |  GPT Image 1.5 |  Nano Banana | ChatGPT Images 2.0 | Winner |
|---|---|---|---|---|
| World map with capitals | "Did an absolutely terrible job" | "Actually crushed this" | "Way more detailed," with a little gibberish when zoomed in | Images 2.0 |
| Code editor screenshot (UI mockup) | "Absolute nightmare" | "Pretty darn good" | "Maybe better. I think so." | Images 2.0 |
| Handwritten shopping list | "You can see how fake this looks" | "Pretty good job" | "Very realistic, even the paper creases," but it didn't write between the lines | Nano Banana |
| Trader Joe's receipt | Shown, no comment | "More realistic even with the stains and crumbs" | "A little more fake than Nano Banana" | Nano Banana |
| Woman at a Paris cafe (realistic photo) | "That plasticky skin" | "A little too contrasty almost" | "Way more realistic" | Images 2.0 |
| Clock reading 3:33 in a kitchen | Not shown | "Just did four" | "Nailed it" | Images 2.0 |
| Water cycle and Spanish phrase infographics | "Complete gibberish" (Spanish) | "Just feels like AI" (water cycle) | "Something you'd see in a textbook," one misspelled word | Images 2.0 |
| Saul Bass and Soviet-style posters, embroidered patch | "These are all kind of the same" (Saul Bass), "Not a whole lot of difference" (Soviet), "All kind of the same" (patch) | Tie | ||
| 360-degree panorama of Half Dome | Shown, no comment | "Did not properly stitch together" | "Something definitely feels off" | Neither |
| Vintage postcard | Not shown | "I think I like the banana better" | "Not bad" | Nano Banana |
All quotes are Pat's, from his video published 22 April 2026, embedded above. "Not shown" means that model's version didn't appear on screen for that prompt.
Which AI image generator is best right now?
In Pat's side-by-side test, ChatGPT Images 2.0 was the best of the three for most jobs, and Nano Banana was the better pick for handwriting, receipts and a vintage postcard. Pat's biggest standout was the world map: Images 2.0 "crushed all of these cities, countries, where Nano Banana wasn't able to get that detailed."
The winner changes with what you're making. For a realistic photo, a chart, a mockup of an app screen or anything with a lot of small text, Images 2.0 came out ahead. For a fake handwritten note or a receipt that looks like it sat on a kitchen counter, Nano Banana's results looked more real to Pat. OpenAI's older GPT Image 1.5 trailed both almost everywhere.
Is there a free AI image generator?
Yes, several. ChatGPT's Free plan makes images: OpenAI's help page on ads says Free and Go users in the test may see a picture ad while an image is being made, and that switching Free to its ads-free version takes away image generation. AI for Mortals covered how those picture ads work. OpenAI's page doesn't say which image model Free users get, so this page doesn't either.
Google's Gemini app makes images for free with Nano Banana 2, and its newer Nano Banana 2.1 comes "with a Google AI plan," according to Google's Gemini image page (checked 3 November 2026). AI for Mortals has the full breakdown of what Nano Banana 2.1 costs and who gets it free.
Two more are free inside apps you may already use. Meta says using Meta AI with Muse Image "is free for everyday creation," with more available on its paid subscription plans (Meta's announcement); see Meta's free Muse Image in Instagram and WhatsApp. Canva says Canva Craft is available to all Canva users inside Dream Lab, its image tool. AI for Mortals covered Canva Craft, Canva's own image model when it launched.
Which one is best at text in images?
ChatGPT Images 2.0 was the best of the three at text for most jobs in Pat's test, with handwriting as the exception. Text was where Pat saw the biggest jump: "this is probably the biggest improvement that I've noticed."
On a Spanish phrase cheat sheet, GPT Image 1.5's text was "complete gibberish," while Images 2.0 had "the correct phrases all lined up" and fit the text into its boxes, though Pat spotted at least one misspelled word. On a Steam store page mockup for a fake Half-Life 3, "the text rendered perfectly." On a magazine spread, Pat read a line aloud and said, "Wow, it nailed the text." He couldn't judge Nano Banana's spread because its image wasn't high enough resolution to read.
For infographics, Pat said the Images 2.0 water-cycle guide "feels like something you'd see in a textbook," while Nano Banana's "just feels like AI." On a code-editor screenshot, GPT Image 1.5 was an "absolute nightmare," and Pat said Images 2.0 "did better than Nano Banana for actually generating UI mockups."
Handwriting went the other way. Images 2.0's shopping list was "very realistic, even the paper creases," but it "was not able to write in between the lines." On the Trader Joe's receipt, Nano Banana's looked "more realistic even with the stains and crumbs," and Pat's summary was: "I still think for handwriting receipts Nano Banana is better."
Which one makes the most realistic photos?
ChatGPT Images 2.0 made the most realistic photos of people in Pat's test, especially casual phone-style shots. On the prompt for a woman at a Paris cafe, GPT Image 1.5 had "that plasticky skin, that kind of weird color cast," and Nano Banana Pro looked "kind of a low res" and "a little too contrasty almost." Images 2.0 had "that like HDR iPhone look," and Pat said: "I would not be able to tell the difference between this and a real photo."
He called people overall "a measurable improvement from Nano Banana." A police officer's selfie got the same reaction, though he expected some gibberish on the badge if he zoomed in.
Images 2.0 wasn't good at taking direction. When Pat asked it to zoom in on a surfer's face, it was "not great at listening to my feedback," and a fresh chat with the same request still couldn't get in close. Hands were mixed: the cafe photo's "fingers look good," but a museum tour guide's hand still came out wrong ("Still can't get those fingers right").
Pat also tried photos in the style of famous street photographers, a category he calls "near and dear to my heart": "As a former street photographer, this is going to make me cry a little inside." The Henri Cartier-Bresson and Vivian Maier versions came out convincing, down to the motion blur. ChatGPT's guardrails blocked a Saul Leiter prompt until he rephrased it as "in the style of a photographer like Saul Leiter, not actually Saul," and he couldn't get a Daido Moriyama prompt past them at all.
Where Nano Banana still wins
Nano Banana beat ChatGPT Images 2.0 on handwriting, receipts and a vintage postcard in Pat's test, and possibly on the Saul Leiter photo. On the postcard, Pat said, "I think I like the banana better." On the Saul Leiter image, he said "Nano Banana might be a little bit better."
Both models struggled with the 360-degree panorama of Half Dome: Nano Banana "did not properly stitch together" the ends, and the Images 2.0 version had "something definitely feels off." On several poster styles and an embroidered patch, Pat saw little difference between all three, though he kept crediting Images 2.0 with better texture, pointing at "the sheen in the paper" on a Bauhaus poster.
Pat calls clocks something image models are "notoriously bad at." On a clock at 2:30, Pat said "1.5 was way off. Nano Banana was close," and Images 2.0 "pretty much got it." On a kitchen scene with a clock at exactly 3:33, Images 2.0 "nailed it," and Nano Banana "just did four."
Newer image models Pat hasn't tested yet
Pat's results above are from 22 April 2026, and several image models have shipped since. Pat hasn't run any of these through the same prompts, so this page doesn't rank them.
- ChatGPT Images 2.5 is the model OpenAI released in September, after Images 2.0 (our post on ChatGPT's Try On button, which it powers).
- Nano Banana 2.1 costs about half as much per image as Nano Banana 2 on Google's API, is built into Google Search's AI Mode, and comes with a paid Google AI plan in the Gemini app (our Nano Banana 2.1 explainer).
- Canva Craft is the first image model built by Canva's own research team (Canva's announcement).
- Muse Image is Meta's first in-house image model and makes and edits pictures from your own photos (our Muse Image post).
- Runway's image and video generator works inside ChatGPT as a plugin and runs on Runway credits (how Runway in ChatGPT works).
Midjourney, Adobe Firefly, Ideogram and Flux weren't part of this test either.
How Pat tested
Pat generated the ChatGPT Images 2.0 results live in ChatGPT on 22 April 2026, the day it launched, usually starting a new chat for each prompt. The GPT Image 1.5 and Nano Banana results for the same prompts were already laid out in columns on a whiteboard-style canvas, and he dropped each Images 2.0 result into the empty slot beside them. Claude wrote the prompts ("Claude wrote all these prompts, too").
The video covers about 50 prompts in groups: text, people, street photography, video game screens, reasoning tests like clocks, marketing assets, posters in other languages, panoramas, infographics and poster styles. Not every prompt was shown in all three models, and Pat names "Nano Banana Pro" for some prompts and "Nano Banana 2" for others. Each verdict is Pat's own judgment from looking at the images, with no scoring system. ChatGPT's guardrails blocked prompts naming Saul Leiter, Daido Moriyama and the Metropolitan Museum. Rewording got the Leiter and museum images through, and the Moriyama prompt never worked.
Questions people ask
Is ChatGPT the best AI for image generation?
Of the three Pat compared, ChatGPT Images 2.0 won the most rounds: maps, realistic photos, UI mockups, infographics and the 3:33 clock. Nano Banana won handwriting, receipts and the vintage postcard.
What is the most realistic AI image generator?
For photos of people, ChatGPT Images 2.0 in Pat's test. Of its cafe photo he said, "I would not be able to tell the difference between this and a real photo." For handwritten notes and receipts, Nano Banana's looked more real to him.
Is Nano Banana free?
Nano Banana 2 is free in the Gemini app, and Nano Banana 2.1 needs a Google AI plan there, per Google's Gemini image page.
Can AI image generators draw hands correctly now?
Mostly, in Pat's test. He said the fingers looked good on the cafe photo and on hands holding a brochure, but a museum tour guide's hand still came out wrong.
Who wrote the prompts Pat used?
Claude did. Pat mentions it partway through the video: "Claude wrote all these prompts, too."
Why did ChatGPT refuse some of Pat's prompts?
Its guardrails blocked prompts that named a real photographer (Saul Leiter, Daido Moriyama) or a real museum. Rewording as "in the style of a photographer like Saul Leiter, not actually Saul" got the Leiter image through.
For more on the company behind ChatGPT, see what OpenAI makes, in plain English, or browse more AI tools for creators.
Update 3 November 2026
This is the first version of this page. It covers Pat's 22 April 2026 test of ChatGPT Images 2.0, GPT Image 1.5 and Nano Banana, plus the free options and the image models released since then.
Comments
Sign in to join the conversation.