I’m a big fan of the GPT3 API and would love something similar for DALL-E which also (ideally) exposes some more internals (e.g. text and image encodings) For GPT3, a notebook style interface that allows you to chain together multiple prompt-based operations would be powerful

codex finetuning

Fine tuning on top of Codex models.

There should be different intensity of each model with different pricing tier. Eg: Dall-E or GPT lite. Also there should be a minimum viable light version that is at a free tier to allow users to build and experiment.

Also, openai can ask users to opt-in to use their input prompt as public information to be shared in openai marketplace. Perhaps those who opt-in to share their prompts have a lower tier pricing. Sorta like with ads and without ads premium.

More control over DALL-E parameters like quality and seed. Options for aspect ratio output. Negative prompt support. And pre-saved selectors like “realistic” that are automatically added to all my prompts if selected. Just look at all the prompt flags of Midjourney and copy them

a dashboard for the fine-tuning API would be really useful

Fine tuning is quite expensive and there is no chance to update fine tuning data. If I do small changes on the data I have to retrain my model. Costs will go up quickly with every change. it is hard to find a business model to cover all cost and make a profitable project.

Tools for fine tuning prompts and measuring the delta between outputs of different prompts across runs. Right now, it's a lot of qualitative eyeballing and saying "I think it's better than before"