AI image generator shoot-out: I tested ChatGPT vs Gemini vs Meta AI to crown a winner

[admin]

I performed a side-by-side comparison to evaluate the image generation capabilities of three leading AI models.

AI image generator shoot-out: I tested ChatGPT vs Gemini vs Meta AI to crown a winner : Read more

 

[cam.macleod]

The funny thing is this totally personal preference, because I honestly thought Gemini easily pulled first in every scenario, except the interior design. I was honestly very impressed by Meta though. ChatGPT's always feels extremely artificial somehow.

 

[jansyren]

I feel that Gemini followed your prompt better while the others embellished the picture more. While they might given you more what you hoped for, I think it is better that it doesn't embellish and you have to refine the prompt to get what you envisioned.

Like the view over the ocean, that's very vague, change the prompt to be "a view overlooking a rocky beach with the ocean beyond fading into the horizon." This might have given you a better picture. Because I think it is better that the ai takes the prompt very literal.

 

[jansyren]

The funny thing is this totally personal preference, because I honestly thought Gemini easily pulled first in every scenario, except the interior design. I was honestly very impressed by Meta though. ChatGPT's always feels extremely artificial somehow.

I agree, gemeni followed the prompt closest without embellishing. I tried the following prompt on the living room: "Create an image of a minimalistic living room with a large window overlooking a rocky beach with the ocean beyond fading into the distance" and that's exactly what I got.

 

[AI_Nerd]

Imagen is one of my favorites for generating photos with minimal prompting. It's got good prompt adherence and is fairly effective at generating some good images. I thought they were the best overall. I don't know what Meta uses but it works pretty well.

Dall-E was impressive maybe a few years ago, but I hardly ever use it anymore. There are simply better generators for nearly every kind of image I would want to create.

For the most part I use SDXL and various fine tunes/variants for SDXL, like Dreamshaper XL lightning, when those fail I also use Imagen 3.0 and it's fast variant, or Flux and Flux Schnell.

Really I'm kinda disappointed an open source image generator like Stable Diffusion XL wasn't included. Especially when Dall-E was. I'm not sure which version is free, but honestly I don't get great results from 3, or 2, and I haven't used the first in a long time.

 

[AI_Nerd]

I agree, gemeni followed the prompt closest without embellishing. I tried the following prompt on the living room: "Create an image of a minimalistic living room with a large window overlooking a rocky beach with the ocean beyond fading into the distance" and that's exactly what I got.

Imagen (Alphabet/Googles image generator) is really very good for photos. It tends to have stronger prompt adherence than some other generators, so what you type is largely what you get.

 

[AI_Nerd]

The funny thing is this totally personal preference, because I honestly thought Gemini easily pulled first in every scenario, except the interior design. I was honestly very impressed by Meta though. ChatGPT's always feels extremely artificial somehow.

AGREE, I hardly ever use Dall-E anymore. It's very expensive and even for creating cartoon or "artificial images" there are simply better, and cheaper generators. I think OpenAI has easily one of the best LLMs out there, but they are not even worth mentioning with image Gen. I don't know much about Dall-E 3, whether it's a CNN, GAN, a VAE/transformer like Dall-E or a Diffusion Model like Dall-E 2, but I've never cared much for the images any of them have generated, other than when I was new to image generation and they were the first one I tried.

 

[brysoncreighton]

Interesting outcome. 100% personal preference and based on your breakdowns of each image I found it odd that you chose Meta in most of the sections. Your closing statement also seems to contradict itself.

Gemini seemed to stick to your prompt without adding additional components or making assumptions. As other users mentioned the ocean scene is a great example, I'd rather the model follow explicit instruction and refine via prompt editing than to be forced into the world of embellishment.

Having used a plethora of image generators I lean towards realism and this is Gemini's default approach. It not only followed your prompts to a T, it also rendered the most realistic result. This can be difficult to achieve with most platforms.

Overall good write up! I was surprised by Meta's performance.

 

[jeffey29]

I shared what you said about Meta AI with Meta AI . Here is the reply from the AI
Here's a potential response from Meta AI:

"Thank you, Tom's Guide, for the comprehensive review! We're thrilled to be declared the overall winner, demonstrating Meta AI's commitment to delivering high-quality, realistic images.

We appreciate the acknowledgment of our strengths in detail, coherence, and prompt understanding. Our team will continue refining fine details, balancing stylization and realism, and ensuring consistency across prompts.

We're proud to be among innovative leaders like Google Gemini and ChatGPT, driving advancements in AI-generated imagery.

Your feedback fuels our growth. We look forward to future evaluations and continued improvements!"

 

[ryant71]

"Sponsored by Meta"