pfrancke Loc: cold Maine

Using stable diffusion InvokeAI, I used the below prompt and got this image. (this one was the first one out, often I will produce a hundred images before I select one to work on).

-----------Prompt----------------------
gritty soft focus, organic detailed intricate ornate, oil painting graphite drawing profile fantasy fierce stern snarl, sorceress, smokey cloud steam mist cracked dusty windblown insect shell helmet in murky fog, high cheek bones, ivory armor bright scales shimmer at sunset, in style of John Howe and Ciruelo Cabral, Jean-Baptiste Monge [lights, framed, bad perspective, multiple, double, male]

I took the image into Topaz gigapixel and doubled its size. Gigapixel also did some facial corrections for me.

So, anyway, this is my starting place. Now I will play with the image for 10 minutes or so and see if I can make it better or worse.

One thing that I try to do as a goal is to make the foreground arm less prominent and will attempt to draw more attention to the face. In my next update to this thread I will post what I attempted to do and the results.

(Download)

pfrancke Loc: cold Maine

ok, after 15 minutes or so, this is what I end up with. I attempted to sculpt the face and remove some distractions. I sculpted by taking the areas where the curves are (where the form turns) and making them slightly bluer in cast and adding texture and strength to such places. I sharpened there slightly also, and tried to fade areas further away. Anyway, for good or bad, this is what came out.

(Download)

magnetoman Loc: Purbeck, Dorset, UK

I’m surprised at the length of the prompt Piet, and the apparent contradictions it contains (like ‘gritty’ and ‘misty’ or ‘sorceress’ and ‘male’). Is this chosen from experience or simply a matter of ‘throw everything in and see what we get’? I like the result by the way!

pfrancke Loc: cold Maine

the words inside of brackets are [negative prompts] - words are funny to AI, misty mostly (I think) would refer to atmosphere while words like dirty, gritty, etc, would apply to surfaces. Some people write the prompts in sentences, but to my experience (questionable), the AI does not do a good job of understanding context - just plain old word list seems to work as well as a sentence.

Usually I write a prompt and then use it to generate 5 images and see what I get. Then I add and subtract whatever words come to mind to try to push the AI into the direction that I want.

Playing with AI feels a lot like pulling the handle on a slot machine. The same words pretty much got me this image too - vastly different, yet they feel like they are from the same world.

(Download)

NJFrank Loc: New Jersey

I agree with Dave. That is a lot of verbiage. To get the final results. Have you tried it with a few very specific words.

magnetoman Loc: Purbeck, Dorset, UK

Ah, now I get it. I can understand the word list being better than a sentence. I like this option of image, quite dramatic! Thanks for explaining how you decided upon the method you use. When I get to look at AI it will give me a decent start.

pfrancke Loc: cold Maine

yeah, I do not disagree with you guys. I use words to push and pull the AI along. For some reason, the way my clutter head works, I find it easier to add words than to remove words. The AI does not always seem to use all the words in a list, it seems like a probability thing. I can use the word fog and I don't always get fog. But if I use the words fog, smoke, mist - then I seem to have a higher chance to get a thick atmosphere.

pfrancke Loc: cold Maine

I am glad that Frank called you Dave. I knew (once) that you were Dave - now my muddled head can hopefully start remembering that again. Yeah, the AI needs to be pushed and pulled like a donkey - and adding and subtracting words is what I have found to be most effective. And for a given set of words, you need to give it a fair shot to see what it does, a sample of one seldom works. And I prefer the old 1.5 model - since it still recognizes artists names and Those Words are particularly powerful.

pfrancke Loc: cold Maine

also notice that I use words like graphite and oil painting at the same time. The word graphite seems (to me anyway) end up producing stronger edges. And oil painting will give it a better chance to appear painterly, and if I added the words 'photo realistic' while including the others, the image would be more likely to appear more real. Unlike our world, the words used in AI do not seem to be mutually exclusive. They are a pseudo language that triggers the AI in 'muddy' ways.

magnetoman Loc: Purbeck, Dorset, UK

I’m learning from you and the other guys here. At the moment I’m still preferring the thought of compositing my own images rather than have something conjured by AI, but this may well change when I get frustrated by the lack of an image in my collection or the need for some extra drama. It does seem to me that the current , or modern, requirements of the image-maker is drama in super-quick time - and AI seems to fulfil those requirements very well.

SoHillGuy Loc: Washington

I like both images, however, I seem to be drawn more to the second image produced.

pfrancke Loc: cold Maine

your approach requires much greater talent. One of the things I like about integrating AI components and then running image to image is that the result seems to be very unified. An interesting workflow might be to compose as you do and then feed the results into an image-to-image process - much as one might use a filter. And then blend the results back into your original to your taste.

pfrancke Loc: cold Maine

thanks - I was a least a little successful then with my post processing. I find that many AI images are difficult to improve on... (it makes me humble!)

Curmudgeon Loc: SE Arizona

Thanks Piet, beautiful image. A couple of questions if I may.

Does phrase order matter? Does Stable Diffusion put more emphasis on the first phrase than the last.

Do you have any suggestions on how to improve the odds of getting full body shots?

Going to push 100 today so I'm going to spend the day trying to understand Stable Diffusion.

pfrancke Loc: cold Maine

LOL, I believe phrase order Does matter, but not as much as one would expect. For a full body shot, one thing to do would be to have portrait dimensions rather than landscape or square dimensions. Also, you could improve your odds by using words like standing, walking, running, etc. A really good strategy that always works is to find an image that has the composition that you desire, and then do img2img process on it using your descriptions to control what the figure looks like. Best of luck, look forward to seeing successes and some failures too - for sure it is a fun journey.