DALLE3 is definitely amazingly higher quality but I still feel it's kind of… useless. It's too hard to control in conversation form, because it's not a multimodal LLM but rather just works by rewriting text prompts. ChatGPT doesn't really know what dalle can and can't do, the actual dalle model still frequently fights you and just generates whatever it wants, when it generates prompts it sometimes leaves out or writes conflicting details, and it writes every prompt in "friendly harmless AI voice" so there's superfluous hype adjectives.
This is odd because GPT4V actually is a multimodal model, and asking it to describe images as text works really well IME.
Common failure modes I saw playing with it:
* if you set something in eg "France", ChatGPT will rewrite it to be diverse and inclusive by literally saying "diverse people" a lot, but then also adds a lot of super-stereotypical details, and then dalle ignores half of it. So you get sometimes-diverse French people who are all wearing striped shirts and berets in front of the Eiffel Tower.
* You can have a conversation with it and tell it to edit images, which it does by writing new prompts, but it's dumb and thinks the prompts are part of the conversation. So a second round prompt will sometimes leave out all the details and just say "The girl from before" and produce a useless image.
* Composition is usually boring (very center-aligned and symmetrical) and if you try to control it by using words like "camera", it will put cameras in the picture half the time.
* Typical AI failure modes: "anime" generates a kind of commercial/fanart style that doesn't exist rather than frames from a TV show. You can't generate things if they're hard to explain in English, like "a keyboard whose keys are house keys".
But it's really good at putting things in Minecraft.