messy_ai said: Not much variety here yet, and definitely a lot to learn about how best to prompt it, but first results with v5 are looking promising.
Currently rather bad at any kind of "dynamic" scene, but a much higher level of realism and good at the "after" shots..
Did you use text prompts only, or did you also feed some iage prompts to get those results? I also experimented a bit ... but had to use image prompts to really get decent results, except for the rain shot
These look like even the best WAM producers will be out of work soon. These look great I don't have the time to tinker with this stuff but does it do video yet? Or is that in the works?
messy_ai said: Not much variety here yet, and definitely a lot to learn about how best to prompt it, but first results with v5 are looking promising.
Currently rather bad at any kind of "dynamic" scene, but a much higher level of realism and good at the "after" shots..
Did you use text prompts only, or did you also feed some iage prompts to get those results? I also experimented a bit ... but had to use image prompts to really get decent results, except for the rain shot
Honestly I personally have no compunction about using image prompts - I am trying to get the model to produce the image I want it to, usually because of gaps between my annoyingly specific fetish and the content that producers have made.
That said, nor are image prompts the 'hack' that people seem to think they are! It doesn't use the image, it just reverses the process, getting the CLIP model to describe the image, and then using that as an input. I find the best way is to get about 90% there with the main prompt, an then use the image prompts to "tune" it, usually with a fairly low image weight (eg. 0.8 or so, where the min is 0.25 and the max is 2).
I also don't claim any particular expertise yet on v5! It responds very differently to V4. But FWIW, this is the prompt i'm currently playing with:
[celebrity or generic name] in [clothing, full description], sitting in a pool of slime :: [celebrity or generic name] gunged on gameshow, gameshow slime , nickelodeon slime gameshow, [clothing, short description], embarrassed expression, humiliated, gunged, slimed trashed, slimed clothes, wet hair, slimed hair, slime dripping down face --no [things you want to avoid after generating the first pass] --ar 2:3 --chaos 0 --iw 0.8
So for example, that generated the set of initial images I attach to this post. Not terrible, but not exactly what I was after, and a bit 'un-blended'.
Then after the "tuning" phase, I get something that's just a bit more in the direction I was after (though definitely not exactly what I was going for).
WildThang said: These look like even the best WAM producers will be out of work soon. These look great I don't have the time to tinker with this stuff but does it do video yet? Or is that in the works?
Video probably won't be many years, but it's currently still about where still image stuff was in 2020 or so.
messy_ai said: Then after the "tuning" phase, I get something that's just a bit more in the direction I was after (though definitely not exactly what I was going for).
The first of those images - link to full-size below - is probably the best AI WAM image yet produced, by anyone, anywhere. That is an absolute stone cold classic. You have to open it full size to appreciate it but every single thing about it, especially her positioning and the way the gloop has engulfed her suit trousers, is 100% spot on perfect.
Anyone who thinks images of that quality aren't going to be become a commercial success is deluding themselves.
messy_ai said: Not much variety here yet, and definitely a lot to learn about how best to prompt it, but first results with v5 are looking promising.
Currently rather bad at any kind of "dynamic" scene, but a much higher level of realism and good at the "after" shots..
Did you use text prompts only, or did you also feed some iage prompts to get those results? I also experimented a bit ... but had to use image prompts to really get decent results, except for the rain shot
Honestly I personally have no compunction about using image prompts - I am trying to get the model to produce the image I want it to, usually because of gaps between my annoyingly specific fetish and the content that producers have made.
That said, nor are image prompts the 'hack' that people seem to think they are! It doesn't use the image, it just reverses the process, getting the CLIP model to describe the image, and then using that as an input. I find the best way is to get about 90% there with the main prompt, an then use the image prompts to "tune" it, usually with a fairly low image weight (eg. 0.8 or so, where the min is 0.25 and the max is 2).
I also don't claim any particular expertise yet on v5! It responds very differently to V4. But FWIW, this is the prompt i'm currently playing with:
[celebrity or generic name] in [clothing, full description], sitting in a pool of slime :: [celebrity or generic name] gunged on gameshow, gameshow slime , nickelodeon slime gameshow, [clothing, short description], embarrassed expression, humiliated, gunged, slimed trashed, slimed clothes, wet hair, slimed hair, slime dripping down face --no [things you want to avoid after generating the first pass] --ar 2:3 --chaos 0 --iw 0.8
So for example, that generated the set of initial images I attach to this post. Not terrible, but not exactly what I was after, and a bit 'un-blended'.
Are you doing this on a computer? I know midjourney is via Discord so wasn't sure.
That set is indeed incredibly realistic. Even the slightly too smooth textures are actually convincing as they evoke the quality of stuff transferred off analogue TV formats and older jpegs.
But for real though. Can you imagine this fetish, when we can basically pie or slime anyone we've ever been interested in?
It's going to take actual will power not to abuse this tech.
I believe there was a show called "future law" that had an episode about people cooywrighting their likenesses, in order to protect it from being messed with AI or other programs