[bhaney]

This looks... pretty terrible. The images being generated are fine, but the conversion from 2D to 3D is awful. It looks like something poorly lasso-tool'd around the subject, put it on another layer closer to the viewer, and then very poorly interpolated the space that's visible between the two layers when you look at it from an angle.

Am I missing something? I feel like I've seen much better automatic 2D->3D conversions via layering long before this.

[Firmwarrior]

haha, sounds like you're describing every "3d" movie that came out during the attempted 3D TV revolution in the 2010s

The site looks cool to me, I think we're being a little uncharitable to it. It runs at a high framerate and pans around smoothly. If someone or a few people made this in their spare time as a cool demo, it's great IMO

If this is the result of $50,000,000's worth of research and development, maybe it's worth a little scorn

[kingkawn]

Given how 2D looked even a few years ago I’ve got high hopes for this

[joewhatkins]

This is just an off the shelf img2depth model run on top of stable diffusion - I don’t think there’s a novel model or research behind this. People have been doing the same thing in colab for a while.

[kingkawn]

i dont mean this singular method in particular, just the ability to run automatic conversions to 3d or generating 3d assets without needing to hire modelers

[squarefoot]

What I find strange is that it fills the missing details with completely unrelated images. As an example, this "An astronaut meeting the president" uses a layer of grass and trees to fill the missing scenery on Mars.

https://holovolo.tv/v/874a1a

[stuntkite]

It's a tech demo. Six months from now a prompt will be generating interactive environments.

[p1mrx]

Maybe they accidentally trained the AI to fuck up the VR180 camera projection?

The left/right sides of every image contain a different, scaled and rotated image. Some of the discontinuities are visually pretty interesting.

[joewhatkins]

I think this is just created by running a depth prediction model on the output of stable diffusion and then inserting the relevant mesh into a 3D scene. The output of stable diffusion isn’t seamless by default, so those jumps will happen.