On AI Image Generation
12th September 2025
My friends and I held a competition to make a photo realistic multiview image using generative AI. Here are some notes about my attempt to craft a winning entry.
Strategy
I decided early on that I would try several AI image generators in order to find the best one. Most of the images I had previously generated were using ChatGPT so I already knew there was one generator that could produce reasonably photo realistic images.
I reasoned that choosing a suitably iconic landmark location should result in a high degree of consistency between images because generators would have been trained from many real images of the landmark. Then maybe, as long as the generator is photo realistic, I "only" need to persuade the generator to select a different pose for the virtual camera and hold any dynamic elements static.
Attempts
Some of these services are using the same/similar generator under the hood. I only considered free to use services. I was careful to limit my requests to a handful at a time and tried to follow an analytical approach because I know these models can use a lot of electricity. For this reason, some of my observations may be proved incorrect as really one should create a large number of results to ensure they are repeatable.
Notes: 1. The links below were correct at time of writing 2. I'm not showing images for all the prompts that I tried, my final image selection for the competition can be found on the competition page.
Deep AI
It's not clear how many images you can generate for free using Deep AI (https://deepai.org/machine-learning-model/text2img). Deep AI Pro allows you to create 500 standard and 60 Genius mode images per month and you can pay more if you need more images. You do not need to login to use this service.
I asked Deep AI to create stereo pair images but it only produced single images. The prompt appears to have no history and asking to change viewpoint to generate a different view didn't work. For those reasons I moved on to other generators.
Prompt: Create a photo realistic stereo pair image of Times Square
Prompt: Create a photo realistic stereo photo of Times Square, with a 1 metre baseline
Midjourney
Midjourney (https://midjourney.online/) does not require a login to use. You may create 15 images per day. Midjourney offers a negative prompt in addition to the main prompt unlike the other services in this article.
When experimenting with different prompts, Midjourney did not seem to pay close attention to the details of the prompt. On several occasions it produced an artistic version of what was asked for, see the middle image below. It's not necessarily a negative it's just that this was not suitable for my purposes.
I believe that the model picked out the words it recognised and did not build an overall context around the prompt. In the last image below I wanted to see if I could deliberately prompt it to make a surreal image.
Prompt: A Tibetan monk standing in front of an elephant with his hand on the elephants trunk. The background is mountainous and a path can be seen with long flags. The image looks like it was made with a stereo camera.
Prompt: The white house, with the sun setting behind it
Prompt: A hot steaming mug of rainbow with a furry bowl of blue star shaped cornflakes on a round table with a supernova tablecloth
Canva
Canva (https://www.canva.com/ai-image-generator/) requires an account but simply uses a one time password sent by email tp login. The site shows a user interface that suggests it is trying to be an app (maybe I should have tried this on mobile).
Canva offers 4 different versions of the generated image, which is nice. It did a decent job of generating a stereo photograph. When I tried more detailed prompts it did well at adding most of the elements I asked for but I hit a limit on the length of a prompt.
Of the four options I am showing the best result in the example images below. Canva results were decent; but in some images shading and depth of field were insufficient to make the image photo realistic.
Prompt: Tibetan monk with his hand on the trunk of an elephant
Prompt: Stereo photograph of the white house with left and right viewpoints
ChatGPT
ChatGPT image generation (https://openai.com/index/introducing-4o-image-generation/) is free but only lets you generate 5 images per 24 hour period. You need to pay to make more images.
Despite my best efforts I could not get this model to understand spatial positioning, Using broad terms like "rear view" seemed to be successful. Like the other models here, if you did get the model to generate a view from another pose then details in the scene changed.
I asked ChatGPT to analyse one of the images it generated so I could challenge it on consistency. Some of the objects were not particularly distinguishable but were identified as objects that had been requested. Is it the same model doing the generation and analysis?
Prompt: Create a photo realistic stereo image of Times Square, refined with: The two images that form the stereo pair should be the same content but from slightly different viewpoints
Prompt: Create an image of a wooden picnic table overlooking the Grand Canyon. A shiny silver robot should be sitting at it, eating lunch with a ginger cat
Prompt: (Following the image to the left) Now rotate the camera around the table 20 degrees to create an image of exactly the same scene at the same instant in time. The robot, cats and food should not move. Do not add any new objects to the table.
Google Gemini
Gemini (https://gemini.google/overview/image-generation/) uses the Imagen 4 model and you need to have a Google login to use it.
Gemini had no problems generating stereo pair images (but are they real stereo pairs?). Photo realism was high, probably better than ChatGPT. It did like to change small details in the scene, e.g. colour or style of a car.
I tried generating images of a scene and not asking for a stereo pair. Again, details changed between views. Using the chat interface to try to correct details felt slightly better than ChatGPT.
Prompt: Create a view of Concorde from above, flying over continuous rolling sand dunes. The sun reflects from the wing edges. The tail has the British Airways livery. refined by Create the same view but zoomed out so that we can see a second Concorde flying in the same direction; the previous image was taken from a window on the second Concorde.
Prompt: (Following on from the image to the left) Now generate an image that looks like the same scene but taken with a stereo camera
Prompt: (Following on from the image to the left) An image from a stereo camera should have left and right views side by side.
Adobe Firefly
Firefly (https://www.adobe.com/uk/products/firefly/features/text-to-image.html) uses its own model called Firefly Image 4. It is a paid service that you need to sign up for however they offer 10 complimentary images as a trial.
There are a lot of options including uploading images for reference or style. This definitely has the look and feel of a professional product compared to the other services that offered a paid tier where paying is only a means to generating more images.
Photo realism was high in general but the model managed to produce cars that did not look like cars (see the last image below). I think in photo mode it should be more careful not to mash up images in this way. If I was paying for this I would be disappointed that images can be obviously AI generated.
I don't think 10 images would be enough for me to evaluate whether I wanted to pay for this service. I also noticed in the T&Cs they don't want you to use Firefly to generate image datasets for training other models which seems overly restrictive to me.
Prompt: A stereo photograph of the Arc de Triomphe in Paris; the baseline for the stereo camera is about 1 metre. The viewpoint shows cars and people in front of the Arc. The weather is sunny but there is a rainbow in the sky. Some of the people are tourists taking photos of each other.
Prompt: (Using the image to the left as the reference image) A photograph of the Arc de Triomphe in Paris. The viewpoint shows cars and people in front of the Arc. The weather is sunny but there is a rainbow in the sky. Some of the people are tourists taking photos of each other. Make sure the image shows the same scene as the reference image but from another viewpoint about 4 metres to the right.
Prompt: (Using a stereo image created by Gemini as the reference image) A stereo photograph of the Arc de Triomphe; using the left-right image style of the reference image. The left and right views should have a wide baseline.
Microsoft Bing
Bing (https://www.bing.com/images/create) uses DALL-E 3.
You must create a Microsoft account to use Bing. I already had a Microsoft account (yes, I purchased a copy of Windows a few years ago). I don't currently have a computer that runs Windows.
I could not sign in. I am using Firefox on Linux but trying to sign in results in a blank page. I checked my ad blocker and privacy extensions but they claimed to not be blocking anything.
I tried again with Chrome and got as far as receiving a code by email to use for the login but the login web page would not accept it.
Final Thoughts
It was a battle to get these generators to produce images according to my specifications, although to be fair I was quite demanding on consistency. The services that use a chat type interface usually gave better results although we only have a few data points here. Sometimes it felt that ChatGPT rushed the job, hoping it was good enough and then it would fix it on my next prompt.
For the stereo pair images I did not confirm that any of them were a true stereo pair. That is certainly a future work and might give insight into how well a model understood the stereo images it has seen.
I preferred the results from Gemini but I can see myself using some of the other services, depending on my needs. I really wanted to try some self hosted algorithms but I did not have time to do so.
Generative AI video generators are sure to be more consistent; at the time of the competition I did not find any free services to test with.