This is neat but Splats are not really mean to be edited in this way.
Splats are sort of like byte code, they are the compiled and optimized representation of reflected light as semi-transparent guassians.
Or you can think of them as the PDF equivalent of a Google or Word Doc. All the logic is gone, and you just have final optimized results.
Generally when you edit PDFs, the results are not great and you cannot make major edits because the layout won't reflow, etc.
So while this is cool, I don't think it will take off unless there is another innovation in terms of either using AI to "reflow" the lighting and surfaces after an edit, or inferring more directly the underlying representations (true surface properties and the light sources.)
Hi Ben! I would argue that it is very useful for splats to be edited in this way. I couldn't have built this application without SuperSplat for isolating, cleaning, transforming and optimizing/compressing the PLY:
- cleaning up noisy GuassianSplats is useful. There are often stragglers floating around in space that need to get deleted.
- compression/optimizing them is useful.
This being a cleanup and compression tool makes sense, but I guess I don't call that an "editor."
I guess I was more arguing against the idea that this is a viable "editor" where one can combine and manipulate in more radical ways Gaussian Splats. The current technological approach doesn't make this a feasible use case.
- Copy & Paste: e.g. delete a tree and fill the hole with a copied patch of grass
- Color Adjustments: tinting, brightness, etc.
If these aren't editing ops, I don't know what is. :)
Sure, you _could_ go back and recapture photogrammetry or rerun training, but that's super costly in terms of time. SuperSplat lets you make simple edits quickly and easily.
In theory if you delete something you have to recompute global illumination and remove cast shadows in the immediate environment of the removed object, but that information is baked in the gaussian splats. I think that's the kind of limitation the parent comment is talking about.
To be as accurate as possible, yes, you need to consider lighting/shadows. But trust me, in many circumstances, you can copy+paste gaussians and it looks 'good enough'. It depends on the scene and the edit you want to make.
I don't like the "byte code" analogy for Gaussian splats. If they were like that, then we could apply compiler optimization and those sorts of math techniques to them. But they are Probability Distribution Functions with transforms, so the math tools we have to work with them are the similar to those in signal processing -- resampling, quantizing, estimating, etc.
In that model, we don't compile them, we train them; we don't run them, we sample/rasterize them.
This link came up on HN before and was a great refresher/expander on the math of Guassians which allow all this. [1].
Since Gaussians can be estimated, neural networks can model/generate them. Researchers are using this for 4D work and mesh extraction. The NNs run at lower frame rate informing the 3DGS running at interactive rates.
You are right that it is ephemeral and really a weird trick of the eye and we need new ways to edit/create it. Vectors/pixels have had a lot more time to grow tooling. People are working on it, just the toolbox is different. Very cool stuff will be coming up, I bet!
I don't really think this is true. Gaussian splats certainly came from a context where an opaque representation is expected and normal, but they ended up being an entirely comprehensible format. They're not as simple to operate on as an SDF or voxel representation, but I think they're on par with triangle mesh geometry. A transformed fuzzy sphere is about as complex as a triangle, and spherical harmonic colors, while more conceptually difficult than textures, have fewer moving parts.
I guess in theory what you say could be correct, however in practice this tool has been very helpful for client work of editing, cleaning, cropping and even slight modification of Gaussian splats. I could see a similar argument for raster images in general -- they are hard to edit as you're modifying individual pixels and it's not efficient, but we've seen tools grow from MS Paint to modern Photoshop to become very useful. I think the same could be said here -- it's just early and we're at the "bytecode" level as you say.
I found it interesting, I'd heard of guassian splats but not really appreciated how they worked, but this let me play with a model; so I'm not saying necessarily useful but instructive.
Did anyone build a text-to-splat 3D generation model? Seems like it would be pretty straightforward? Should make it really easy to generate assets for video games.
Aren't gaussian splats incompatible with most common game development styles? No shadows, no rigging/animation as main issues - or maybe I'm misunderstanding / behind on the research - please correct me.
Is there a reading guide to the maths behind Gaussian splats? All the resources I could find either assumed lots of knowledge (including what a "3D Gaussian" even is), or were written for complete lay-person (and probably includes some AI grift).
Even this article suffers from the same issues? Just in the first few paragraphs: what are spherical harmonics? I read the Wikipedia article but it all went way above my head. There really needs to be a maths-for-smart-but-not-math-phd-people resource.
I wish all these guides used better notation. They all use the scariest possible Greek symbols (pi epsilon feta etc) that are hard to make sense of without a degree in anthropology, instead of nicely named variables like programmers use.
Any tips for an app to use on iOS to capture the necessary .ply data?
Scaniverse is a great app by Niantic that can do this on-device, but it isn't very customizable and can't export its raw scanning data (exported .plys do not have the data this editor requires).
Not sure if I'm missing something, but the submission title says "open-source" and in the tool's help menu there's a link to the repo (https://github.com/playcanvas/supersplat), the tool runs in your browser, there's no server involved besides a Web-Server hosting the files.
Those were the times when you could actually rely on Google to give you the right results for such a query near the top, and more importantly, give you the same results it would give to the person you're telling to Google it.
Splats are sort of like byte code, they are the compiled and optimized representation of reflected light as semi-transparent guassians.
Or you can think of them as the PDF equivalent of a Google or Word Doc. All the logic is gone, and you just have final optimized results.
Generally when you edit PDFs, the results are not great and you cannot make major edits because the layout won't reflow, etc.
So while this is cool, I don't think it will take off unless there is another innovation in terms of either using AI to "reflow" the lighting and surfaces after an edit, or inferring more directly the underlying representations (true surface properties and the light sources.)
https://playcanv.as/e/p/cLkf99ZV/
Integrating AI is an interesting topic and something that certainly has potential.
- cleaning up noisy GuassianSplats is useful. There are often stragglers floating around in space that need to get deleted.
- compression/optimizing them is useful.
This being a cleanup and compression tool makes sense, but I guess I don't call that an "editor."
I guess I was more arguing against the idea that this is a viable "editor" where one can combine and manipulate in more radical ways Gaussian Splats. The current technological approach doesn't make this a feasible use case.
- Copy & Paste: e.g. delete a tree and fill the hole with a copied patch of grass
- Color Adjustments: tinting, brightness, etc.
If these aren't editing ops, I don't know what is. :) Sure, you _could_ go back and recapture photogrammetry or rerun training, but that's super costly in terms of time. SuperSplat lets you make simple edits quickly and easily.
In that model, we don't compile them, we train them; we don't run them, we sample/rasterize them.
This link came up on HN before and was a great refresher/expander on the math of Guassians which allow all this. [1].
Since Gaussians can be estimated, neural networks can model/generate them. Researchers are using this for 4D work and mesh extraction. The NNs run at lower frame rate informing the 3DGS running at interactive rates.
You are right that it is ephemeral and really a weird trick of the eye and we need new ways to edit/create it. Vectors/pixels have had a lot more time to grow tooling. People are working on it, just the toolbox is different. Very cool stuff will be coming up, I bet!
[1] https://news.ycombinator.com/item?id=41912160 I've also re-learned Fourier transforms to appreciate similar concepts.
- SuperSplat dev :)
- https://www.meta.com/en-gb/experiences/gracia/25784099001234...
- https://www.gracia.ai
EDIT: yep - https://gsgen3d.github.io/
Individualkex also has a couple videos on the high level ideas: https://youtu.be/GQXDjzNWuPc?si=zlAN7dO9STGATKad
Any tips for an app to use on iOS to capture the necessary .ply data?
Scaniverse is a great app by Niantic that can do this on-device, but it isn't very customizable and can't export its raw scanning data (exported .plys do not have the data this editor requires).
[1] https://github.com/playcanvas/supersplat
I remember a time when it was considered unpolite to ask a question without googling first. Is it still the case?
Yes