Quicksand Fans

Both of you do a great job and your results are marvelous. That Midjourney is more photorealistic and i like that more by myself. Also the "mud" looks more real and just like in a real location. The ripples Viriadian adds, are not realistic, but are well fitting for the anime-styled pictures his AI generates.

What i'm missing so far, is a certain feeling of real stuckness. I did search my archives and these here are some screenshots that transport this feeling and desperation (of Leila and Sarah-Michelle ) better. This is a thing, AI cannot recreate (yet...?)

I've spent couple hours trying to place a model waist deep in the lake. No way! "Waist deep", "partially submerged", "half submerged", "stepping in", "entering", "wallowing", all these prompts resulted in either model is fully/neck-deep in the water or is mostly above the surface. Any suggestions how to specify the depth of submersion in the prompt?

bogbud wrote:What i'm missing so far, is a certain feeling of real stuckness. I did search my archives and these here are some screenshots that transport this feeling and desperation (of Leila and Sarah-Michelle ) better. This is a thing, AI cannot recreate (yet...?)

Well, MJ definitely has some knowledge of emotions shades. They are even somehow manageable, at least in comparison with depth of submersion

John Junkman wrote:I've spent couple hours trying to place a model waist deep in the lake. No way! "Waist deep", "partially submerged", "half submerged", "stepping in", "entering", "wallowing", all these prompts resulted in either model is fully/neck-deep in the water or is mostly above the surface. Any suggestions how to specify the depth of submersion in the prompt?

Again, not familiar with MJ prompts, but commenting on AI in general based on my NovelAI experience. "Submerged" and its synonyms are typically associated with water, not mud/quicksand, so using them is likely to bring up water. I always use "partially submerged in mud" to at least get something, and put "no water" or a negative prompt to try to get rid of the association. Try other earthy-synonyms, such as "buried". In general, however, the AI doesn't recognise the concepts we have of "waist-deep" or "stepping in", so it can be very challenging to control depth.

However, this is where you can "trick" the AI. The AI will make more of an effort to feature prompts that you include, and exclude negative prompts. With NovelAI at least, adding tags like "face" or "detailed hair" will often force headshots, so the subject might be submerged up their torso or necks to get this right, and adding negative prompts like "legs" will make it less likely to spawn a picture that has legs, which means more likelihood of being immersed deeper.

Viridian wrote:

John Junkman wrote:I've spent couple hours trying to place a model waist deep in the lake. No way! "Waist deep", "partially submerged", "half submerged", "stepping in", "entering", "wallowing", all these prompts resulted in either model is fully/neck-deep in the water or is mostly above the surface. Any suggestions how to specify the depth of submersion in the prompt?

Again, not familiar with MJ prompts, but commenting on AI in general based on my NovelAI experience. "Submerged" and its synonyms are typically associated with water, not mud/quicksand, so using them is likely to bring up water. I always use "partially submerged in mud" to at least get something, and put "no water" or a negative prompt to try to get rid of the association. Try other earthy-synonyms, such as "buried". In general, however, the AI doesn't recognise the concepts we have of "waist-deep" or "stepping in", so it can be very challenging to control depth.

However, this is where you can "trick" the AI. The AI will make more of an effort to feature prompts that you include, and exclude negative prompts. With NovelAI at least, adding tags like "face" or "detailed hair" will often force headshots, so the subject might be submerged up their torso or necks to get this right, and adding negative prompts like "legs" will make it less likely to spawn a picture that has legs, which means more likelihood of being immersed deeper.

I've specified intentionally the "lake" not the "swamp" in that session, to find out how to put model at waist deep into any substance. Regarding "no legs" option / portraits - interesting point, will try it. BTW, looks like this is MJ-specific issue. I've tried several other AIs, and they easily processed the prompt with "waist deep": "beautiful woman hopelessly stuck waist deep in the marsh energetically struggling, photorealistic style". Here are couple of samples:



Regarding "no water" tag - yes it works, but not in the way it's expected ) Water is removed literally, leaving empty hole/ditch or field I've tried dozens of times with the same results:



Regarding "buried" - I've tried it too, and no thanks ) MJ immediately kills the model by this word

Also I've tried to use sample photos kindly provided above by Bogbud (I guess this is Leila / DIDVP) together with the prompt "beautiful woman in overalls stuck and struggling in quicksand" and it worked. So MJ has some "knowledge" of the quicksand substance and "concept" of waist-deep depth. The open question is how to build such scenes by pure prompts

Oh wow, those last pics are awesome and exactly my cup of tea. The mud looks also much more real here.

And it's Leila Hazlett in really serious UK-mud. She did need a rescue in the end...

In the second-last picture, while I know that it is a tire in the background, it is fun to imagine that it is a tentacle rising from the mud, soon to wrap itself around the creature's next meal!

Hello John.

I agree with bogbud. You've cracked the problem of creating believable muddy surfaces using MJ. Now, instead of looking watery it's clearly opaque, realistically coloured, able to show reflected sunlight and also textured just like the real thing. These are significant advances made in just a short time. Well done!

Redjak.

A couple of updated attempts using Stable Diffusion