I have been meaning to redo the test I did about half a year ago on changing particular image elements for some time now, in order to assess the progress since then.
I headed to Mistral, the french AI company, and asked it more or less the same question as the last time.
Here’s your image of a man in posh clothes walking in the mountains:
Here’s your image of a man in posh clothes walking in the mountains:
Would you like any further refinements?
Perhaps unsurprisingly, the exact same thing happened with Mistral as with Gemini back in June last year; the image theme is the same, but the image is not. I am not sure wheter Mistral uses their own image generation, or some other provider of the image generating models.
Not giving up, I went to ChatGPT 5.2 Thinking and did the same
Here’s your image of a man in posh clothes walking in the mountains:
Here’s the exact same image again (unchanged):
That must qualify as more than acceptable. I notice some of the shadows in the background mountains are very slightly different, as well as the light on his bag is also slightly different, but all in all it is the same.
When I asked it to regenerate the exact same image, I did however not directly get the same image, I rather got a matplolib frame that appeared to show the last image. (I took a snippet of the chat to show the response, as downloading the image did not give the same visual impression). During the thinking I could see it flashing some ideas about difficulty in regenerating exact pixels, and wheter it could just repost the same image. Apparently it was able to run some python code inline and display it.