Awesome and informative video 😀 I love to deep dive into things as well, just never got around to looking at the shorten yet. I think I remember David mentioning at one of the Office hours that the shorten is not exactly one to one with the Model, meaning that it tries to guess some things and is quite accurate there, but it can be off sometimes, because it is not actually tied to the image generation model. I have noticed that some words do actually make a difference in the end result even though the shorten detail view shows them at a really low priority or at 0.
Thanks for doing this video...other "instructors" are only confusing people with claims that specific movies, directors, film types, camera types, etc. have a effect. They might if they were the only prompts, but words like "shrimp" for example, obviously matter more. Midjourney isn't a natural language processor like ChatGPT, but rather just processes a prompt for keywords. I honestly don't think the order of the keywords matters, I've tried to resort them just separated by commas, and there is no difference. I don't think anyone understands how it works, including the creators, because that is the nature of AI, it trains on images it scrapes off the internet, and words that are associated with the images it scrapes. In short, use /shorten on a prompt before you use it. It can help you improve it. Also, long prompts are bad because you can include words that might seem insignificant to you , but are flagged by Midjourney as important and will dilute the significance of more important words in you r composition.
That's a pretty good summary! It's extremely difficult to know what a word will do to a generation - so starting small and then adding as you go - is a good way to progress a prompt
more awesome content. Didn't know about the shorten prompt, just tried it. very cool. What would you say is the advantage of using --style raw? I haven't used it.
There are 2 main things. First, it listens to you more closely, which means it's harder to prompt. A shorter prompt might result in ugly pictures. You need to write a lot, and like specify that you want it to be beautiful lol but I think it gives you more control overall and the other thing is that it has a bias towards real life. So if you prompt 'dog' in --style raw, 4/4 of the pictures will be photorealistic dogs. So in the end, if you have something really specific you want to generate, and you'd like it to look real (rather than illustrated or painted), --style raw is a good thing to use. Thanks for the good question, I'll make a video about this soon.
Most likely, there is a lot of correlation between different words. For instance, with the artists, the fact that Dan Hillier did not seem to matter that much initially could easily be explained by suggesting that when it comes to the latent space of all dimensions from which an image is constructed, he is not significantly different than the other two artists. Their "essence" is already in the image, and if you change the order to nominally alter the weights, the actual representation of the styles does not change. With a slightly more elaborate version of the argument above, I think you can explain the effect of the ordering on weights. You are shuffling the weights but not moving far from where you were in the high-dimensional latent space.
It's a hit and a miss sometimes, but the order does matter. Try and see how it weighs 'depth of field'. It's strange. 0.00 in many shorten prompts, but there's always a difference. Like the person's comment below, the language model and MJ's model are different, so that might be why there's sometimes inconsistency. Thanks for the video. 👍
Does anyone know what website lets you upload multiple images and then just click a button to get a blending animation between each image in one video? I saw a youtuber show it a few months ago and can't find it.
Everybody says cutting out useless words makes for a better prompt, but from I've seen with my own experiments the shorter images lose depth/atmosphere/emotion. Compare the first image of Audrey Hepburn to your shortened images. Yours look flat and lifeless. So what is it with "ignored" words that flesh out an image?
I know exactly what you mean and I should have mentioned it in the video. You could say the extra words don't really matter, or you could say that they add the little bit 'extra' to the images.
I've been checking Midjourney's feed and I've noted some prompts using codes like (#18fc49d7#), and in the docs I haven't seen anything about that prompting technique, have you? Do you know when and how to use it? Great channel!