How do I get AI to create an image where everything is where I want it to be?
MidJourney AI Prompt : lonely cowgirl in the foreground looking out at a horse across a diagonal fence which splits the picture.
When at first you don’t succeed….
I came up with a simple scene based around the composition techniques raised in the last post. As AI does not understand the rule of thirds etc. how does one position the frame?
Starting sketch above – Ideally I wanted the fence to disappear to the upper right rule of third point, the first post hold a strong presence on the left third line and the cowboy to occupy a large portion of the right-hand third, thus carrying significant “visual weight” in the composition. The fence was to be the metaphorical barrier of the cowboy’s loneliness and longing to be back in the saddle.
The resulting image ended up pretty close but the path to get there exemplifies the random nature of AI. It is very difficult to control AI and I have no idea if my prompt was just the luck of the “seed” or as a result of careful planning on my part. It took 11 iterations (i.e. 44 images) for the ghost in the machine to produce the best match, most were not even close.
AI you’ve got it wrong! Was it something I said?
I deliberately stated my AI Prompt with an open suggestion of the 4 subject elements – cowboy , fence , horse , prairie – and it was unsurprising that the AI needed a better understanding of what I thought it should be doing. Next I separated the elements into their respective positions – cowboy foreground on the right , fence diagonally , horse background on the left, prairie –
It is fair to say that AI doesn’t understand left or right and it is doubtful that it understands foreground or background either.
AI – make me a picture from my picture
I had read that AI can use picture urls as a prompt and I had also read that it can be a way of getting to closer to positioning things to where you want them to be. I tried incorporating the url of my initial sketch adding the keyworded elements , importantly adding Media Type: photograph but surprisingly it still blended my image to produce a sketch. Neither did it it follow my composition. Back to square one.
Promplets and Weights
In MidJourney a prompt can be divided into sections using a double colon :: called promplets and these promptlets can be weighted. Reading through the FAQs I don’t think I am the only one who has difficulty with the success of this weighting. One of the main issues is that the promptlets need to be relational. i.e. One can’t just type cowboy::3 fence::2 horse::1 to result in the “visual weight” of these nouns, they all need to be connected and repeated with the desired weighting. My understanding is these weights are much more a blending weight than a visual weight e.g. if you wanted a centaur it might do a better job of working out how much horse vs. how much person. With version 5 of MidJourney, I found natural language is much better than comma separated keywords. I am sure this is why my attempts to compartmentalise the nouns did so much worse than a single sentence, no commas, no promptlets – lonely cowgirl in the foreground looking out at a horse across a diagonal fence which splits the picture.
Setting the scene, secret herbs and spices
After saying that a simple single sentence did the trick, I now must tell you the embellishments that were added to create the scene. These were separated from the main sentence (subject), it made a difference. Could I have constructed a succinct paragraph to include these and get the same or better results? – Media Type: photograph, 1970s, kodachrome, Ernst Haas, Environment: prairie, moody sky, Lighting: sunset – one has to note the special wildcard to a perfect composition “Ernst Haas”.