My first (and probably last) attempt at AI qs

[sixgunzloaded]

It's pretty feeble. Just imagine it's below the surface and try not to laugh too hard...

Susan says: "If you're gonna do this to me, at least make it look real!

[Duncan Edwards]

You're just getting started. In all seriousness I've probably gen'd up 30,000 images in the last year and still only like one in a hundred. Nice thing about this is no models to pay and no film to waste.

[Viridian]

Don't be disheartened by the initial attempts. Just look at our AI threads where we started from. It takes a lot of experimentation to figure out what workflow works best for us.

[cerberus]

That's a pretty good effort for a first attempt, you've got the style and look of your other work pretty much nailed. I'd say you weren't far off there. Certainly worth persisting with it a bit longer.

[sixgunzloaded]

Thanks for the encouragement, guys!
I don't know why it wouldn't give me anything other than swirling water around her legs. I tried asking for 'quicksand' and every synonym of it I could think of, and indicating exactly where it should go with their selection tools, but the only other thing it would give me was a washed out mess of random colors.
I also tried putting 'water' in the negative prompt box but it just sort of ignored that.
Basically, this was just a little experiment to see if I could make anything interesting using my other stuff as seed images...or whatever.
I may try again every now and then, but I think I'll mostly just leave it to the pros (you guys).

[quicksand3D]

prompt chocolate pudding ?

[sixgunzloaded]

prompt chocolate pudding ?

lol An interesting idea! Ya never know!

[cerberus]

That chocolate pudding idea could work but may look a bit disgusting, maybe porridge or similar could be worth a try.

I was in the process of writing a few tips, but got "GoDaddy"ed, it thought I was cross-scripting (i.e. including embedded code)! Looks like I'll have to phrase it differently.

My prompts never say "quicksand" except for beaches, I usually use "mud quicksand" and often apply weighting to the whole sinking term. By that I mean I enclose the sinking phrase in parenthesis and add a weighting factor before the closing parenthesis (a colon followed by 1.3 say). This causes the AI to give the term a higher priority, note: it doesn't mean that increasing the weight makes them sink more, but it often increases coverage. Watch out for ambiguous terms, "sinks" can result in your subject sitting in a square pit with taps! Weights can also be applied to the terms in the negative prompt, but results can be a bit variable.

It may be worthwhile just experimenting with prompts for a while to see what gives good results, what you are trying is quite sophisticated really. You wouldn't pick up a guitar for the first time and start trying to play Arnold's Guitar Concerto (before you laugh, yes, this is a real piece of music).

[Viridian]

Thanks for the encouragement, guys!
I don't know why it wouldn't give me anything other than swirling water around her legs. I tried asking for 'quicksand' and every synonym of it I could think of, and indicating exactly where it should go with their selection tools, but the only other thing it would give me was a washed out mess of random colors.
I also tried putting 'water' in the negative prompt box but it just sort of ignored that.
Basically, this was just a little experiment to see if I could make anything interesting using my other stuff as seed images...or whatever.
I may try again every now and then, but I think I'll mostly just leave it to the pros (you guys).

As with cerberus, I never use "quicksand" in my prompt. I use "sand", "sandy mud" or some variant. Note that depending on which AI platform and model you are using, even the prompts "sinking" or "partially submerged" may trigger a water effect. However, when prompted with an image with a quicksand texture, or using inpainting to edit an image, it will quickly pick up the quicksand prompt, whereas a raw generation will often end up with water.

Many platforms now have more advanced tools that allow style and "vibe" checks to capture the tone and features of a premade image. I often recycle my old generated images to create new ones much more consistently than working raw.

Using this image as my style reference:

c2f13048b87e4a16906b601b7e0f677c.jpg

I use that in NovelAI to generate an anime-style reference picture for my next step. Note that the initial output is rather gray and watery. I presume that you were creating a scene from your recent Evil sequence. The problem is that the AI will associate a "fountain" with water, so it will be less likely to generate mud.

evil1.png

I then run that through my Mage app, which uses the more advanced Exo comic model with my blend of prompts and LoRAs, to get this:

evil2.jpg

That's enough to get it looking like grey mud, but let's say we wanted something more earthy. I go back to my NovelAI reference, use inpainting or editing to change the soil colour to a brown:

evil3.png

Then run that again through Mage/Exo and get a much muddier result, as the AI more clearly recognises the brown surface as mud as opposed to water.

evil4.jpg

Should I want to create a 3D style render or a photorealistic image, I have that option to do so with these reference images.

[sixgunzloaded]

Yep, I tried both "sand" and "mud". I tried to use both simple and creative prompts. I tried combinations like "sandy mud" and "wet sand" and "thick clumpy mud". It didn't seem to help.
I wasn't aware of things like the parenthesis or weighting factors, though. I didn't see any sort of options describing those, but that might have been because I was just using their free trial version.
DAZ has an AI program now that uses their Genesis 9 3D figure. I got some fairly decent raw, text-to-image results playing around with it a couple weeks ago. Maybe I'll try feeding it my seed pic and see what it does with it.
Actually, Viridan, I was using my single-frame update pic "Christmas Magick 2024" that I posted in January. No fountains, just a bed and a qs carpet. lol
But I like what you did there! Especially evil4.png. Is there no way to achieve that sort of result using just one AI program? And do it for free? I'm not ready to dump any money into it, at this point.