Comment by gametorch

Hey thanks for the feedback! Your image link is giving me 404, fyi.

> I assume you are running some type of LLM to specially format the prompts to the image models, or is it more complex than that?

Yes, we pre-prompt by default, but you can manually disable this. If you disable it you also avoid paying the LLM input/output token tax.

Strangely enough, though, specifically formatting prompts and creating pre-prompts hasn't really had as large of an effect as I thought it would on the quality of the end result.

Despite how little of an effect pre-prompting has, I still think most of the value add right now comes from the business logic. It's about what you do before and after sending the prompt to get the best results. It'll also be about crafting the most frictionless UI for a human. For example, they need to be able to easily convey to the computer "this image and this other image are good in these specific ways, please combine them to make this third image."

The other long term business view here is these models are truly improving in quality exponentially and decreasing in cost exponentially. Read: they are commodities. This means you WILL get categorically replaced if you are in the business of having the best model (the most sentimentally valuable piece of gold! or corn bushel!). The real value add here, if there is any long term, has to be in the UI, organization, and distribution of the commodity. That's what I'm trying to build with GameTorch.