That was a great talk, Jack! I have been watching videos about gaussian splatting for over a year, and that made it make more sense than anything else I've seen.
Its Crazy how you can already Find tests of people using Relighting techniques and even Animated Splats using grid of cameras. I've seen a lot of new 3D tech but if the tools allow for Relighting for using in and 3D environment it could truly have a leg to stand on versus the typical 3D methods. SO much potential and its not far from that tipping point.
I had a feeling, but after watching this confirmed, that neither this CEO nor any of the people asking questions understand anything presented here, or what they are using. From notions of "there are only 2 representations of 3d models" to "Mesh" being the only thing supported in 3d apps. (that you need their addons.. why?) To finally that they can't light a ply with the same light as other objects etc... oh my oh my. (Did he say they work on it for 5 years? what?) Go back to school or watch Introduction to Computer Graphics by Cem Yuksel. Splatting is from the 90s, and a separate pipeline in your GPU. (It didn't just "came on" lol ) The reason you don't do this is, 1. you have no actual surface, so no collision, no physics, meaning also no lighting (in path tracing) 2. You just now traded a few pixels of a bitmap, for the same amount of vertecies/points with that color at that position. So not only do you have to store the color, but also the position... great job! (You use textures, since they are vastly less memory hungry than actual vertecies with vertex color, which at a minimum has to be two vec4 floats at 24 bit, meanwhile an image with BC compression only holds a vec3 float, but BC/ block compressed so shared between 4 pixels.) Also, since you have a surface, you can use tileables to add detail finer than the texel density of the texture maps even. Or bake normal detail to lower the vertex count even lower... buuut I digres. 3. What rly put me off was the "rigging problem" of 3D Gaussians... I was like what rigging problem? Just assign some weight to each point and voila! Until I remembered, that even a small .ply file, before meshing, has millions if not trillions of points... You remember how we still animate a lowpoly model, no more than 10k faces, instead of the actual million poly sculpt out of zbrush? Ahh, yes... bcs animating at 0.015fps just isn't fun. Now I am more than interested in the "supposed" paper for SIGGRAPH, where they move a million points based on rigged weights...
Great talk, I can see a bit of convergence between 3DGS and voxel based engines as they share the same issues/shortfalls, that (so far) can only be solved by mesh based models .. ie animation and rigging etc.
Totally agreed! 3DGS and voxel share so many similarities. Compared to mesh-based models, 3DGS has undeniable draw backs; But what really makes me excited about the 3DGS is that it can do things that mesh can never do (like the red flower example I showed at the talk) And compared to voxel, it is so much light weighted and has shown the potential of better editability.
3d splatting people put too much focus on capturing reality, i can also capture cgi scene. Given that's basically a point, a 3d matrice and a color vector as spherical harmonics, we can extend color with other data, basically converge on material pbr data and resolve lighting in deferred.
Yes totally agreed! I've seen some research efforts on representing the texture and materials with 3DGS, can't wait to see what the future looks like :)
Actually I've seen some professional LiDAR Surveying scanners being integrated with 3D Gaussian Splatting to visualize large areas. It's gonna be really cool in the near future
That was a great talk, Jack! I have been watching videos about gaussian splatting for over a year, and that made it make more sense than anything else I've seen.
Its Crazy how you can already Find tests of people using Relighting techniques and even Animated Splats using grid of cameras. I've seen a lot of new 3D tech but if the tools allow for Relighting for using in and 3D environment it could truly have a leg to stand on versus the typical 3D methods. SO much potential and its not far from that tipping point.
Totally agree! And I'm so glad to see many researchers are focusing to make this tech better everyday
It's been 3 months already, but most of the info are still correct.
Wow so great , I love KIRI Engine 🎉🎉🎉
Thank you so much!! I love you too
I had a feeling, but after watching this confirmed, that neither this CEO nor any of the people asking questions understand anything presented here, or what they are using.
From notions of "there are only 2 representations of 3d models" to "Mesh" being the only thing supported in 3d apps. (that you need their addons.. why?)
To finally that they can't light a ply with the same light as other objects etc... oh my oh my. (Did he say they work on it for 5 years? what?)
Go back to school or watch Introduction to Computer Graphics by Cem Yuksel.
Splatting is from the 90s, and a separate pipeline in your GPU. (It didn't just "came on" lol )
The reason you don't do this is, 1. you have no actual surface, so no collision, no physics, meaning also no lighting (in path tracing)
2. You just now traded a few pixels of a bitmap, for the same amount of vertecies/points with that color at that position. So not only do you have to store the color, but also the position... great job!
(You use textures, since they are vastly less memory hungry than actual vertecies with vertex color, which at a minimum has to be two vec4 floats at 24 bit, meanwhile an image with BC compression only holds a vec3 float, but BC/ block compressed so shared between 4 pixels.)
Also, since you have a surface, you can use tileables to add detail finer than the texel density of the texture maps even.
Or bake normal detail to lower the vertex count even lower... buuut I digres.
3.
What rly put me off was the "rigging problem" of 3D Gaussians... I was like what rigging problem?
Just assign some weight to each point and voila!
Until I remembered, that even a small .ply file, before meshing, has millions if not trillions of points...
You remember how we still animate a lowpoly model, no more than 10k faces, instead of the actual million poly sculpt out of zbrush?
Ahh, yes... bcs animating at 0.015fps just isn't fun.
Now I am more than interested in the "supposed" paper for SIGGRAPH, where they move a million points based on rigged weights...
Not enough feature points why not take numerical points overlay
Amazing talk, Jack!
Thanks :)
Great talk, I can see a bit of convergence between 3DGS and voxel based engines as they share the same issues/shortfalls, that (so far) can only be solved by mesh based models .. ie animation and rigging etc.
Totally agreed! 3DGS and voxel share so many similarities. Compared to mesh-based models, 3DGS has undeniable draw backs; But what really makes me excited about the 3DGS is that it can do things that mesh can never do (like the red flower example I showed at the talk) And compared to voxel, it is so much light weighted and has shown the potential of better editability.
GREAT TALK 👏👏
Awwww, thankssss!!!!
At a talk about GS, surely 100% of the audience should have raised a hand.
Hahah, I was actually guessing there were some dev who were just curious about new techs, but yeah I was surely surprised
i try 3dgs but i have a problem when export file i need help
Hey! We are glad to help :) Please feel free to contact us at contact@kiri-innov.com
Also, feel free to just let me know your question in this comment, either way will work :)
3d splatting people put too much focus on capturing reality, i can also capture cgi scene. Given that's basically a point, a 3d matrice and a color vector as spherical harmonics, we can extend color with other data, basically converge on material pbr data and resolve lighting in deferred.
Yes totally agreed! I've seen some research efforts on representing the texture and materials with 3DGS, can't wait to see what the future looks like :)
Insufficient for large areas. There are a lot of programs that do small tasks. It's not very efficient for me.
Actually I've seen some professional LiDAR Surveying scanners being integrated with 3D Gaussian Splatting to visualize large areas. It's gonna be really cool in the near future