Thread by @blog_supplement

Thanks. Multiple people have asked for prompts behind our recent posts. I'll just mention something general in that regard. First, imagine what you want -- that is based on your knowledge of the intended composition. Describe it in words to the engine. The key to note is that

https://twitter.com/santalum_aurum/status/1709635374081859612

this implementation of DALL-E 3 in MS image generator is not like MidJourney. It does not perform mutations on an earlier generated image. So there is randomness: 1. Some of the prompts fail but slight tweaks to the statement while maintaining semantics gives better results.

2. just repeating the same prompt a 2nd time sometimes works better. 3. There are lots of "heartbreaks" -- a near-perfect image ruined by a single irredeemable flaw -- if only we could engineer or mutate these we could get even better productions. I keep those that come near to

the intended mental images even if slightly imperfect. Imperfections are unavoidable in the current AI-- the defect with limbs, facial symmetry persists from earlier versions. However, I think DALL-E 3 has greatly improved. Maybe MidJ is better now but I have not been able to use

after the early days -- it inexplicably kicks me out. The true potential is probably superior to the version we can use -- the duShTa-s mention crippling it in the name of "safety" and prudery. Also, I fully don't understand the "boost" system. You are supposed to get 100 boosts

i.e., apid GPU access, a day. I initially got 100 a day but now it is now renewing. You get more by using bing instead of mahAduShTa-guggulu. So try your own experiments.