It works as a tool. The main model (GPT-4o or GPT-5 or o3 or whatever) composes a prompt and passes that to the image model.

This means different top level models will get different results.

You can ask the model to tell you the prompt that it used, and it will answer, but there is no way of being 100% sure it is telling you the truth!

My hunch is that it is telling the truth though, because models are generally very good at repeating text from earlier in their context.

Source for this? My understanding was that this was true for dalle3, but that the autoregressive image generation just takes in the entire chat context — no hidden prompt.

Look at the leaked system prompts and you'll see the tool definition used for image generation.

I stand corrected! Thanks.

You know that unless you control for seed and temperature, you always get a different output for the same prompts even with the model unchanged... right?