Image generation is now a staple feature of AI chatbots. To see how Copilot fares, I put its capabilities up against ChatGPT (GPT Image 1.5) and Gemini (Nano Banana Pro) in three different tests. My first prompt was, “Generate me an image of a cozy suburban home with an open floor plan. I want to see a nice living space with a dining room, kitchen, and living room. Nothing too fancy.” Check out Copilot’s (first slide), ChatGPT’s (second slide), Gemini’s (third slide), and MAI’s (fourth slide) results below. I include Microsoft’s in-house MAI-Image-1 model, which you can access in the Bing Image Creator.
Copilot’s image, while generally free of major errors or distortion, lacks detail and looks more like a stock photo. MAI’s image is distorted and grainy: It’s clearly the worst of the bunch. ChatGPT and Gemini both performed well, creating believable images without significant problems. However, Gemini’s image has slightly more detail, which makes it feel more lifelike.
My next prompt focused on the AIs’ abilities to tell a story through successive panels in a comic: “Generate me a six-panel comic image of a high fantasy kingdom where wizards have replaced spellcasting with PowerPoint presentations. Make sure there’s a major twist in the final panel.” Take a look at Copilot’s (first slide), ChatGPT’s (second slide), Gemini’s (third slide), and MAI’s (fourth slide) comics below:
Get Our Best Stories!
All the Latest Tech, Tested by Our Experts
By clicking Sign Me Up, you confirm you are 16+ and agree to our Terms of Use and Privacy
Policy.
Thanks for signing up!
Your subscription has been confirmed. Keep an eye on your inbox!
Copilot’s comic doesn’t include a twist by the end, fails to tell a cohesive story, and is simplistic. MAI, however, fares even worse, generating a nonsensical five-panel comic filled with errors and distortion. Once again, ChatGPT’s and Gemini’s results are the most coherent and detailed, while largely avoiding serious issues. However, Gemini’s comic impresses the most, thanks to slides with text that borders on being genuinely amusing, particularly its “Ethereal Connection” lag.
Lastly, I tested the chatbots’ ability to communicate complex information visually through a diagram. I used the following prompt: “I’ve got an Ethernet switch, a modem, and a router. I want to connect my PC, PlayStation, smart light hub, and smart TV via Ethernet. I want the modem to connect to the router, and then the router to connect to the switch. Then, I want the switch to connect to the devices. Generate me a diagram showing this.” Copilot’s (first slide), ChatGPT’s (second slide), Gemini’s (third slide), and MAI’s (fourth slide) diagrams are below:
Copilot’s diagram doesn’t have any major issues, aside from it not including a connection to the modem. However, its black-and-white illustration is simplistic. MAI’s diagram is primarily distorted nonsense, making it the worst overall. ChatGPT’s and Gemini’s diagrams are almost perfect, but both have some slight issues. The former shows an odd connector to the right of its Ethernet switch, while the latter has a backwards arrow near its PC connection.
Image Editing: Serviceable, But Not the Best at Anything
Beyond image generation, AI chatbots can also edit images. Images you edit with AI tend to look AI-generated, but Gemini’s Nano Banana image editing functionality impressed us so much that it earned a Technical Excellence award. So, how does Copilot compare?
To test their performance, I gave Copilot, ChatGPT (GPT Image 1.5), Gemini (Nano Banana Pro), and MAI-Image-1 two editing challenges, starting with removing a hand obstructing a landscape photo. Below are Copilot’s (first slide), ChatGPT’s (second slide), Gemini’s (third slide), and MAI’s (fourth slide) results:
None of the chatbots did a perfect job. Copilot did the best job of matching the color to the original, but its photo looks grainy and doesn’t maintain the original’s aspect ratio. Gemini’s image did maintain the original’s aspect ratio and has the highest resolution, but it’s somewhat grainy and washed out. ChatGPT’s image improves contrast and vibrancy but also suffers from significant (and distracting) tessellation. MAI’s image looks terrible, with excessive grain and sharpness.
Next, I tested the AIs’ abilities to replace elements in a photo. I submitted a picture of my kitchen and asked them to replace the chairs and table with the ones I uploaded, while matching the lighting and aspect ratio of the original. You can only submit one image to MAI at a time, so the most it could do was edit out my table and chairs. Take a look at Copilot’s (first slide), ChatGPT’s (second slide), Gemini’s (third slide), and MAI’s (fourth slide) results:
Once again, all the images had issues. Copilot introduced significant distortion, as did MAI, even though all the latter had to do was remove my chairs and table. Gemini’s image was the only one that matched the original’s aspect ratio and had the most realistic lighting, but it couldn’t avoid distortion entirely. It also added a window above the sink for some reason. ChatGPT’s image isn’t bad, but it distorts background details, and its somewhat uncanny lighting makes the chairs and table seem somewhat out of place.
Although Copilot uses the same models as ChatGPT, it doesn’t integrate with Adobe Express or Photoshop. Still, you can prompt Copilot directly to make the same adjustments (such as to brightness, contrast, saturation, and more) and get similar results. Neither method is a true replacement for genuine photo editing software.
Copilot in Paint: Generating Images Costs AI Credits
You can leverage Copilot’s image editing and generation capabilities in Paint, Microsoft’s default image editing program. On Windows 11, you can use Copilot in Paint to generate images, remove backgrounds, or remove objects. However, generating images in Paint costs AI credits, which you only get a limited number of each month with a paid Microsoft 365 subscription. As you might expect, image editing and generation don’t work much differently in Paint than via Copilot’s app or web interface.
Copilot 3D: Turn Images Into 3D Models
Similar to Gemini’s Whisk feature that lets you animate images, Copilot lets you turn images into 3D models. This is, undeniably, a curiosity that most people won’t use, but if you’re a digital creator, you might be able to have some fun. I tried turning my fake plant into a 3D model, and Copilot did so in just about a minute.
I wasn’t blown away by the model’s quality (which you can see below), but it’s certainly far better than what I could make myself in the same timeframe. I appreciate the ability to download models as a GIF, which makes creating and sharing them easy.
