First, I just wanted to say that DALLE is still way too restrictive. It’s almost like you’re using basic keyword matching and not… you know… the world’s most sophisticated LLM to understand whether or not the prompt violates the rules. Also, the fact that it still starts the generation before testing the prompt is kinda silly. So here are my recommendations:
- Test the prompt before starting the generation
- Give us the ability to test prompts before even hitting submit
- Actually use your own technology to check the prompts instead of keyword matching (or whatever you do right now)
The second topic is a set of feature ideas:
Give us the ability to generate numerous smaller thumbnails very quickly so we can quickly experiment with prompts. Also, include CLIP or whatever its called so we can get some more ideas for prompts. Then, once we have a thumbnail we want, allow us to quickly upscale or regenerate to any size (4k, 8k, etc)
In other words:
- Generate low res thumbnails quickly
- Embed prompt assistance, CLIP, etc
- Allow us to spend more tokens on generating ONE larger image from preferred thumbnails
Anyways, I kinda doubt you guys are working on this anymore, given the dominance of SD and MJ. The quality of DALLE has been utterly left in the dust, so who knows.