[qingdao99]

And it's far better in my opinion - ChatGPT actually creates its own prompts from what you give it and feeds those to DALL-E (you can see this by clicking on the images it returns and reading what the actual prompt was), although you can disable this by just firmly telling ChatGPT not to modify your prompt.

Additionally, the image generation itself is quite different. ChatGPT's DALL-E seems to create much more stylized images - much harder to get plain shots that don't heavily embellish your description.

[TeMPOraL]

OpenAI justified that in a paper the other day, saying that DALL-E 3 performs better on longer, more detailed prompts describing all aspects of the image in rich language - so they put GPT-4 in front to expand the typical user's short and vague prompt, so such users get nicer results by default.

My own observation: this kind of hack is possible only since/with GPT-4 - it takes an LLM this powerful to reliably extend and enrich arbitrary user input into much longer prompt, that's coherent, consistent, and a plausible (to human) interpretation of the original input.

Now this may fan the flames on the "is it or is it not" AI discussions, but: you could almost say that GPT-4 is engaging in creative process here.