opinions on gaussian splats
back in my day it was called photogrammetry
>>1030965It seems quite limited to small spaces, I doubt you could make a whole city for a game with it, and the storage requirements could be huge for that (I have no idea, maybe it's more efficient, but I doubt it)But I could imagine it being useful for rendering distant foliage or clouds
i dont know why but i am fascinated by it. i do think they will be the future of some sort. it is what voxels always wanted to be
>>1030976We had this hullabaloo over unlimited detail. You use instances to save space and memory
>>1030965Not worth the time or effort beyond "wowing" non's in short format video.
>>1030983>gaussian splatsSince they store view-dependent lighting you can't instance them unless you want the lighting to be all wrong
Really good for static scenes with a layer of objects between the splats and the assets in focus. Pretty good for photorealistic scenes on a low power device or in the unreal mobile render. Not good for dynamic lighting or game interaction conditions.
I think it really shines in VR, where you have the sense of depth and scale. Especially with how it captures transparency and reflections without the problems mesh photogrammetry has. It wastes so many splats on pictures and flat geometry, though. So maybe the idea is a hybrid with textured meshes doing those parts.
>>1030965an AI generate gaussian sploops?
so i have been working on this thingamajig and I thought I should try this out
some collision testing
>>1030987You can fake it by storing light data so that they change their color/position based on the point lights
>>1031038Not good enough for nakidashi, but good enough for a waterfall on a rock that is 1 square inch of a screen.
>>1031052not good enough yet!
>>1031042nta, but is that like how people use normal maps on sprites so they have shading in a 3D environment? Because I was thinking that since the splats are directional, they could have a second color attribute for normals. I guess using a copy of the photos ran through filters/AI to guessimate the angles. (Hopefully better than my example I hashed out with online generators.)It might be convincing, or it might be a mess of blurry edges when a light shines on it wrong.
>>1030965>ten times worse performance for dynamic lighting>rigging any any physics beyond staticbody collisions basically nonexistentDOA "technology">>1030983have fun storing instances of millions of points for what could have been 10000 triangles. There's no point in using this in any sort of dynamic or interactive environment.
>>1031079>lighting>rigging Not what the tech is made for, imo. It's good for static scenery and nonlit props without much of the jank that mesh geometry has. There was a practical example of it in a virtual festival on vrchat some months back. They showed detailed scans of their model kits you could pick up and look at. As I said above, it benefits most from VR for the 3D and scale.
>>1031079Yeah. Use it for static features and use your extra polygons for dynamic things
>>1031111>what is weather?
>>1031113Assuming splats could have something like weight painting, maybe you could have parts like leaves sway a bit. There is also '4D' splatting in the works, which would allow animations with minimal filesize.Otherwise, just pick a different method.
gaussian floor, gaussian wall, gaussian ball, gaussian fog, gaussian light, gaussian shadows
>>1031120Source?
>>1031126my custom "engine" whatever it is
>>1031127sounds cool, can you show the shadow/volumetric test in a higher res screenshot?
>>1031209i think that was before i tested real recursive lighting, with a kernel i was around 30fps i think for 180px height on my 3060
was testing new camera method
was super surprised how well the fog turned out especially. it was my hypthesis, but still am kinda stunned it worked this well. I am just a vibecoder after all
gaussian sculpting?
tests are very promising i would say
