GaussianDreamer: Fast Generation from Text to 3D Gaussian Splatting | Hacker News Reader

GaussianDreamer: Fast Generation from Text to 3D Gaussian Splatting

taoranyi.com

94 pointsbondarchuk16 comments

Here is the discussion for the other recent paper by some of them on 4D Gaussian Splatting

https://news.ycombinator.com/item?id=37905601

How long will it take for everyone else to adapt this technology? More precisely, what is preventing the current companies to use this technology to create a video game? Is it the rendering technique, quality of mesh, easiness to edit, or something else?

So now we have this technology, say in 10 years it is accepted world-wide. What does this imply to the old methods? Are they going to be phased out? Is no one going to use them?

A few years at most. 3d programs will adopt it. Game engines like unreal will adopt it. But right now it requires still too much time. Once computing gets faster by 10-100x it will be there.

I will will eventually become like manufacturing is today. There will still be hand crafted items and even whole games and they will advertise as such. But for the average consumer it won't make a difference if it's auto generated or hand made

Gaussian splats are like 3 months old. They currently can't be relit, and there's currently no way to turn them into a mesh+texture. If/when either of those problems can be solved, I certainly would expect them to be used in the games industry.

> there's currently no way to turn them into a mesh+texture

Turning level sets[1] into meshes is old news, like this[2] for example. Given the Gaussians are smooth, I imagine one could readily subtract a fixed threshold for the sum of the Gaussians at any point and get a zero level set that way.

Similarly, generating texture coordinates for arbitrary geometry is also well known, and used in texture generation algorithms like this[3].

[1]: https://en.wikipedia.org/wiki/Level-set_method

[2]: https://doc.cgal.org/latest/Surface_mesher/index.html

[3]: https://cvmi-lab.github.io/Point-UV-Diffusion/paper/point_uv...

It's theoretically possible to mesh any radiance field. The problem is that the mesh looks like shit.

The gaussian splatting method optimizes the gaussians to lie in the volume of the object, so the approximate the surface well. That's why I figured it might work decently with a level set method.

I'm basing this off looking at the examples from the 4D gaussian splatting[1] posted here the other day[2].

[1]: https://guanjunwu.github.io/4dgs/

[2]: https://news.ycombinator.com/item?id=37905601

For certain fine details or certain not-so-fine details like very flat surfaces, gaussian splatting "looks" right but the underlying geometry at the moment is just not correct.

There are currently implementations for Unreal and Unity, but mainly just for visualization. There are some demos for doing more with the data

https://twitter.com/_kzr/status/1714214841265856932

Can it take an input image instead of a prompt ? The generated model are good but 30 mins is still a lot of time to experiment with prompts.

https://huggingface.co/spaces/jiawei011/dreamgaussian

Interesting! but it looks like a different algorithm

Very impressive stuff. Seems like a huge improvement over Dreamfusion, released only a year ago.

Remember back when you could get a PhD in Computer Graphics by putting like a pixel on a screen?

The rate of innovation is absolutely bonkers.

I put a ton of pixels on the screen every time I turn on my computer monitor, someone get me my PhD ASAP! :)

Reply on news.ycombinator.com