The computer has a certain amount of memory allocated for a drawing. When Midjourney reads your prompt, words take up more or less memory depending on how close they are to the beginning and how important they are for Midjourney (like cat is important, hedgehog is less important, a, the are unimportant). When the memory limit is reached, it ignores the rest – at least in V 5.2. That’s why you won’t see some stuff requested in overloaded, long prompts. If you want what’s ignored in your long prompt back, you can start removing other words (also called “tokens” in this context).
(For multi-prompts, there is something more complex.)
Clarinet has mentioned that when the allocated memory is exhausted, V 6.0 may draw poor-quality images – which you should check by upscaling because only then you see the final result.
How does Midjourney read a prompt? Its code is proprietary, but it is likely this way.
With each word and word combination, a drawing AI pulls associations, or contexts, from its database, where text is connected to images. It happens on several levels: the lowest level of colors and textures, then, at a higher level, styles, then objects and actions. This is achieved by several software components working in parallel. It is like several cooks preparing dinner together, each with their specific role.
When the cooks are asked to make something they haven’t before (like, even contradictory), they become very creative; otherwise, they may follow a recipe. The creative work is called Zero-Shot Learning in AI.
Compared to cooks, AI drawing compares the results with the request and may redo the image-making several times.
The versions 5.2, 5.2R, and 6.0 have different algorithms/ways to achieve that image creation process. I don’t know about v5.2R, but 6.0 and 5.2 were also built on different image databases. So it’s no surprise that outputs are very different.
Bouncing between 5.2 and 6 via Vary(Strong) may allow for richer content. Not sure what I am saying (yet?), but it’s quite an exciting technique.
Thanks for reading!
Irina