The menu: Rosemary and garlic roast lamb, Extra-crispy roast potatoes, Honey-roasted carrots and parsnips, broccoli.
The photo: [1] Roast turkey, mashed potatoes, baby carrots, broccoli, brussels sprouts. No lamb, no parsnips.
If you shoot for what's in the photo, you're going to have a bad time.
You're also going to have a bad time when you try to make an apple crumble with no flour, no sugar, and no butter, because they're not on the shopping list.
[1] https://images.openai.com/static-rsc-4/gyzrX8zp3O2KLEPm0FAqh...
I feel like it can only be successful by luck. Either it's reproducing a recipe verbatim that was tried and validated and tasted by a human, in which case, we didn't need AI for that, just a searchable cookbook. Or, it's making one up. I understand that using RLHF has helped improve the quality of questions about history, programming, or TV show recommendations, but I do not believe that there has been some kind of training regime where people prompt the models for "a recipe that uses X, Y, Z random ingredients," follow the recipe, and then score it. And even if they do, I don't see how a model can learn enough from that besides "this exact recipe is good/bad." '1/2 tsp cumin' may be a great addition to one recipe and not enough for another, so improving the output based on a bunch of scored recipes... I just don't believe cooking is an LLM job.
Maybe some other kind of model that I don't know about.
The user is probably in a grocery store. They need to buy the stuff first. Showing them a picture of a finished product is an irrelevant distraction.
Then, if there is some arguably relevant content you can tack it on afterwards.