Viridian's AI Experiments

[Acidtester]

They’re all great, but that Safari Girl one is incredible!
Really a dream come true to see her looking that good.

[sixgunzloaded]

Yeah, really. Every one of these looks next-level real. The one of Vicky looks like somebody literally dropped a GoPro right smack in front of a seriously stuck girl and startled her. Even zoomed in at magnification.
It's amazing just how fast this is advancing.

You said before that 3D artists had nothing to worry about and gave good reasons for it, and I believe that...but...I dunno, man... Every time we turn around it's better and better. And now we have ones that simulate movement.
Just incredible.

[Viridian]

You said before that 3D artists had nothing to worry about and gave good reasons for it, and I believe that...but...I dunno, man... Every time we turn around it's better and better. Just incredible.

Maybe.

[Jinn]

Love the ancient temple and tomb quicksand. Something more devious about it.

[Viridian]

Bringing back some of the older AI renders of the waterproof suits/waders, now in the best available AI.

[sixgunzloaded]

It really seems to have the hang of backgrounds now. It might have fallen back just a smidge on handling faces, though. It still looks really good but it's like it's trying to overdo it just a little. And what it did to poor Sapph's right arm... There must be a secondary curse in the tomb that mutates the trespassers.
I do love how the walls of the tomb seem to be melting into the mire - sinister looking!
And the girl playing Esme..! I want her in the live-action production!

Edit: I'm curious, let's see what it can do with frames 105 and 106 from the story. How will it handle splashing legs?

[Viridian]

It really seems to have the hang of backgrounds now. It might have fallen back just a smidge on handling faces, though. It still looks really good but it's like it's trying to overdo it just a little. And what it did to poor Sapph's right arm... There must be a secondary curse in the tomb that mutates the trespassers.
I do love how the walls of the tomb seem to be melting into the mire - sinister looking!
And the girl playing Esme..! I want her in the live-action production!

Edit: I'm curious, let's see what it can do with frames 105 and 106 from the story. How will it handle splashing legs?

That was meant to show the pros vs cons of 3D models vs AI. The Sapph frame has her using a climbing axe. That was something I couldn't get the AI to render, so it incorporated it into her arm. That one I left unrefined, because I was most impressed with the sand texture in the hallway. The backgrounds in all of these were breathlessly flawless in detail - and effortless.

The restriction in most of these is actually the source material. It's not that they are low quality, but they reflect the use of 3D models posed in a scene as opposed to a photographer framing the shot. In your sequence, we know the events in each frame prior and after, what the characters are wearing, what they are holding, and so on. So when an arm or prop is held in a blind spot on camera, we still know it's there, but the AI doesn't. If you've arranged the set so that we know there's a staircase, we know it, but the AI doesn't.

TROE - AAHT 093.jpg

In 093, for example, we're supposed to know that Sapph is 1) holding a glowing amulet and 2) catching a red climbing axe. For a frame in a 3D sequence, we can fill in the details. But as a photographic frame, it's difficult - the action is obscured with a cluttered angle, and the lighting is very, very dark. Yes, that's part of the scene, but because everything from the pose to the lighting is _slightly_ off, it doesn't get recognised by the AI. So if we examine 106, the problems with visual clarity are clear.

TROE - AAHT 106.jpg

Consequently, my reproduction of 106 simply cannot get the details.

95ed8ab695e7431fa5b521cc853a771b.png

As a comparison, if I were to recreate the scene from scratch, it might look more like this:

4572059a86f8437d9043a6ceb4040268.png

No climbing axe, as I can't generate that prop no matter what I do.

The clear edge that 3D has is that you can pose your models however you want. However, that does present the challenge that there's no guidance over how you frame your shots (SecondLife stories also have this severe limitation). The problem - being critical - is that many frames are just too dark. This is usually because the scene is meant to show a change in lighting, but in my opinion, there are too many "bad" frames where nothing is visible (I have a similar problem with Aiko's TGS renders). It might literally be in a dark room or an evening swamp, but we kind of have to make it clear to the viewer using "stage" lighting without literally putting them in the dark.

One of the clear pros of the AI renders is that they are freaking amazing at getting the photographic and cinematic lighting. Michael Leach has praised that on my work before - things like outlines, highlights and shadows would be a dream to recreate in a studio.

[Viridian]

It's tough to say that this is "best possible", given how much AI art is advancing. But it's been around 9 months of learning, and given the quality of what I'm producing, I decided to back to the first images in this thread to show the progress. These are the exact same AI images from the beginning put through my current process.

[Viridian]

Next set. I have no doubt that the original renders were woefully inadequate, but seeing how current AI can work with them raw is amazing.

[Viridian]

I've redone the "No Go" set before. What was interesting to me was the ability to add extra details in the render. It was much better at recognising the background as "jungle" instead of replicating the psychedelic mess that the original renders made.