Quicksand Fans

I've tried Midjourney AI. It's really interesting, both the "drawing" process and results. Here is some outcome. I can't say that results exactly match the requests, but it worth a look. What do you think?

Are you the guy who does the toy videos on YouTube?

OBI-wan wrote:Are you the guy who does the toy videos on YouTube?

100% No. Square and circle shapes is the top of my art "talent" ) Even the powerful AI fails facing with it

Those are fantastic (the ones from your first post... )

I'm very much into well protected female biologists, nature-conservationists, hunters and the like.

I don't have Discord, so i can't see how that AI works. Could you tell something about it?
In case you feed it with source-photos i could post some for your usage

Thanks for the feedback!

It works pretty simple. Just type in the chat what do you want to get and enjoy (or not ) the result that most likely wouldn't match your request. AI returns quad of lo-res pictures per your request, for example "Beautiful woman in waders struggling stuck in peat":


For each of four tiles you can ask additional variation or upscale to the full-size "final" result, which may be variated too. But these additional variations won't contain a lot of changes, just some "plastic surgery", spare hand or missed leg , nothing dramatic. Upscaling introduces some minor adjustments too: number of fingers, eyes, etc And ta-dam! you get what you get, or not ) Typical outcome ratio is 5-10%


There several tips like: don't use the word "girl", use "woman" instead. The word "beautiful" is a must (now I'm typing it by my own without spellchecker ) Regarding "challenges": Sometimes AI provide irrealistic pictures, like Bosch or Salvador Dali.


F**cking policies prohibit most of important keywords like "breasts", "butt" etc I have no clue how to clearly specify the depth of submersion in the request. Also it's possible to exclude something via "No" keyword, for example "--No water", but AI has no clue how does the bog without water looks like. Also the set of swamp backgrounds AI learned doesn't look big. Regarding the "external" samples - I have just several successful pictures built based on my samples. Looks like its tough task for AI to properly mix two realities. I got a ton of messed pictures when tried, like when fly gets into teleport together with human


But in some cases it succeeds: revariate - revariate - revariate again - revariate damn @#$@# stupid AI!!! - Oh, good boy )


So there is a lot space for AI improvement in our area. Regarding "casual" pictures, it's really f**cking cool. Mostly it's because AI's learning set contained thousands and millions of samples of casual things, not just three pictures from the MPV site and one from Tarzan movie

Seems AI has still to learn something, but it will become much better fast.
Your and also Viridians pics and drawings are the future for sure.

bogbud wrote:Seems AI has still to learn something, but it will become much better fast.
Your and also Viridians pics and drawings are the future for sure.

Speaking of which:

The quality of the outputs from Midjourney are beyond doubt. It does seem that you're coming across similar challenges as I did with NovelAI in narrowing down the prompt to something that specifically matches what we are looking for.

That does make for a big reminder that the ultra realism of the Midjourney output doesn't really suit our specific niche well. Our concept of quicksand is actually quite unrealistic. Aside from artificial pits, quicksand generally doesn't work that way - which means we need to trick the AI into illustrating it.

Viridian wrote:The quality of the outputs from Midjourney are beyond doubt. It does seem that you're coming across similar challenges as I did with NovelAI in narrowing down the prompt to something that specifically matches what we are looking for.

Viridian, could you share your typical requests? Mine requests can be found in the picture names I'm posting.

Here are five more pictures. No "external" samples here, just pure AI and a lot of "Vary" commands. Third picture took a dozen of iterations before reaching acceptable result. Any comments / feedback are welcome

John Junkman wrote:I've tried Midjourney AI. It's really interesting, both the "drawing" process and results. Here is some outcome. I can't say that results exactly match the requests, but it worth a look. What do you think?

I don't come in here very often but someone directed me to this and I'm glad they did. I'm a big believer that in 10 years this stuff will be creating videos that will be difficult to tell from the real thing. For now though this is amazing work. While the images don't really fit with my specific tastes the quality turned out by the imagination of a machine is amazing and an improvement over much I've seen thus far. Would you say the creation machinery is better or do you just have better skills at telling it what to do with itself?

When I put together a prompt, I break it up into three steps. The first relates to style and form: photorealistic, perfect anatomy, detailed face, detailed hair, detailed eyes. With NovelAI at least, the more tags you put, the more the image will focus on them (e.g. "detailed face" will tend to force headshots)
I would also be using lighting, angle and framing tags here if I was using them. I also have a lot of negative prompts (lowres, text, error, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, etc.).

The second is the environment. Because AI doesn't have a concept of "quicksand" the way we do, I blend in a lot of similar textures and descriptors: quicksand, squishy mud, soft mud, deep mud, girl sinking in quicksand, partially submerged in mud, girl sinking in mud, girl trapped in mud, struggling in mud, etc. Add in textures like "ripples" or swap things out like "sand" to get a closer approximation, plus negative prompts like "water" and "reflections" to reduce the watery appearance. Then the appropriate background (jungle, forest, swamp, construction site, etc.). I like to do in-painting for these to give the AI more direction in what I'm envisioning - most of the ripple effects are prompted by my squiggles.

Third is the subject itself, which is normally the descriptor for the character: woman/girl, brown hair, worried, scared, open mouth, parted lips, wide eyes, yellow waterproof jacket, etc.

Without any in-painting, that combination gets me results like these.

I'm not familiar with how Midjourney works. My advice in creating prompts is not to rely heavily on prose. Things like "mired in peat" or "stuck in marsh" are, in a sense, both too broad and too narrow. Most of the entry-level AI tutorials I've seen get people to use prose to begin generating abstract concepts, but you learn quickly that it's hard to get exactly what you want from these requests alone.