So, this is fun. Type in “create this exact image without any changes to the prompt: Comic illustration portraying an evil deadite woman in her full beauty, enveloped by branches and vines. The fluid linework, sharp angles, and dynamic energy are evident. The play of light and shadow is masterfully executed using the chiaroscuro technique, with warm colors and gritty details adding depth and emotion.”
You will get the exact image I linked. I had someone else try it as well, same results. Apparently there is only one result for any prompt on GPT. On Bing, you get different images. GPT, you don’t. Just thought I would bring it up.
Interesting. I wonder if it’s actually a prompt input to the model, or something else that is an generative variation parameter that we haven’t seen yet via API also not exposed by Bing. A ChatGPT conversation showing variations on a particular theme showed close consistency if wanting one of four variations to be continued upon in a different way.
I added to a prompt “Reproduce this prompt exactly without variations.” and still get about the same amount of variation between the images and same to two days ago, without any styles that fit within a particular genre box that matches between the days-separated runs (but one of the generations rudely not offered.)
Bing image creator, two of the three with asking for no variations:
(also image generations allowed aren’t reset so quickly two days later…)
I just asked Chat GPT about it, this is what it said:
'Yes, I can explain that!
DALL·E, like other neural networks, is deterministic. This means that given the same input (in this case, the same prompt and any other fixed parameters), it will produce the same output every time. The reason people online can all get the same image using the same prompt is because of this deterministic nature.
However, there is an aspect of DALL·E that allows for variability: the use of a “seed.” The seed is a value that initializes the random number generator inside the model. Different seeds will produce different results, even for the same prompt. If you keep the seed constant (or don’t provide one, using a default), then the output will be the same for the same prompt. If you change the seed, the output will vary.
So, in summary, while DALL·E is deterministic and will produce the same result for the same prompt and seed, you can introduce variability by changing the seed value’.
I just asked ChatGPT and here’s what it said.
It’s because OpenAI uses the same seed (5000) for all images.