If the world has to be static and lighting cannot change i can't see how this is useful for an interactive experience

yeah taka a million high def photos of an alley to get janky simulacra

@@danilluzin under interactive I mean for example museums where you can walk around in your pace and look at objects, which is much more interactive than a 360 camera

It is more on the order of 10's of pictures @@mgntstr

​@@danilluzin for generating world map and spaces. Everything dynamics can be generated separately.

Truman Show vibes

Check the complex game

Mesmerising

Reminds me of a movie where they said they used all the cameras from a cctv setup across the entire city to look into the past when in reality it was an actual time machine of sorts. Can't remember the name. Though if you imagine using all the cameras in London to record and build up not just stills, but full timelines of images this way you could experience the city as if you were walking through it in real time. Which has zero practical application

It resonates with some spiritual experiences :)

你好，可以和你交流一下吗

What gaussian splatting technique is used here?

If I had to guess it just means the images shouldn't have too much lens distortion

This rendering technique is just taking its first baby steps. At the moment it seems like people are mainly using it to render environmental scans that are obtained with photos or video. But I suppose there's no reason why a handcrafted high fidely 3D environment couldn't be converted into data that can then be renderes with gaussian splatting. The issue with it for interactive applications is the static nature of it. I haven't seen anyone use dynamic lighting or reflections with this tech. These photo scanned environments have "baked" lighting that just renders everything seen in the photos, and reflective surfaces are rendered as just space like looking through a window. And I've seen no one do moving parts in the environment. At the moment I supposed that any interactive application would need to mix different techniques. Reflective surfaces would need to be done with common methods and moving objects would need to be made from polygons. But all of these issues will probably be solved relatively quickly!

While much less expansive and striking than the remarkable demonstrations here, I found it was reasonable on an 11700 (non-k) and 3060 12GB (from less than an hour to maybe six hours, depending on settings and total number of images), e.g. with some Sony Xperia 1 (II) imagery, kzfaq.info/get/bejne/j66ldcyDna_Md2Q.html

no doubt that google will use their supercomputers to make 3d gaussian splatting google street view

@@melol69 we would still need commputers to be able to run it, and immm also guessing this takes up a lot of space

Give or take 10 years, when ordinary computers could move data few terabyte per seconds we'll get nice game made using this method

​​@@Setupthemabombthe 5090 will be able to push > 1 TB/s, some OC versions of 4090s already reach 1000 GB/s of bandwith. It's complete nuts

it runs very slowly, requires orders of magnitude more images than they already have, and doesn't really work on a large scale.

They will, in a few years.

It's not as advanced as the demo, but Apple's "Look Around" functionality has smooth transitions, much nicer than Street View. I read somewhere that their cameras are equipped with LIDAR, so they have depth info to work with, Google also has some depth info I think, but it's not as advanced.

@@theneonbop Check out their latest paper on SMERF. It is supposed to be much more efficient.

yes it is possible. but you need data first.