GaussianObject: Just Taking Four Images to Get a High-Quality 3D Object | Hacker News Reader

GaussianObject: Just Taking Four Images to Get a High-Quality 3D Object

gaussianobject.github.io

162 pointssmusamashah23 comments

I wonder how this compares to LGM outputs which also came out recently - https://github.com/3DTopia/LGM - specifically LGM looks like it's generating the additional viewpoints from text or an image and then creating a mesh whereas GaussianObject seems to require 4 viewpoints.

We're getting closer to workflows that can generate good 3d objects and texture them with diffusion models, there's a bit of work in integrating these into existing tooling though - also, animations and rigging for characters still seems to be a big issue (at least I haven't seen any good 3d generation models that can create rigged characters yet).

Animation has always been the achilles heel of new 3D representations, the first thing you should ask when someone claims to have replaced triangles is "how do I animate this?". The answer is generally "you can't" and then the technique fades into obscurity. A decade ago it was voxels coming to replace triangle meshes and that went nowhere.

It is possible to turn a NeRF into a mesh, but the quality of the meshes produced is currently so poor that it doesn't really get us any closer to being able to animate or edit the model.

Depends on how you are going to use the result.

This method is great for creating assets. Photogrammetry is already used a lot.

The rule for a lot if 3D work is: if it looks good, it is good.

But I agree this method won't be suitable for all 3D work.

Photogrammetry requires a lot of manual cleanup and retopology work, but it's worth the trouble because it captures a lot of extremely fine detail which can be baked down into textures and applied to the manually refined model. It's not really clear to me that NeRFs are producing enough detail to warrant that cleanup effort, rather than modelling something from scratch.

I would be interested to see a comparison of conventional SOTA photogrammetry pipelines versus NeRF techniques solving a 3D model from the same set of photos, but these papers only ever seem to compare themselves to other NeRF papers, rather than the software that the industry is already using to solve this exact problem.

Well again: if it looks good it is good.

When you are not creating assets for games or close-up details it can still be very usefull to take a couple of pictures of a tree stump and use it in your environment.

Also photogrammetry from an industry level.

I work in construction, we take photos of the building in situ as a lot of the important elements are encased later (eg. Reinforcement inside concrete)

This level of 3D NeRF is great for that as it's much easier to parse than a folder full of images.

https://github.com/Anttwo/SuGaR looks non-poor to me, but I've not had a chance to play with it properly yet.

They don't show any of the topology, UVs, texture maps, skeletons of the extracted meshes, which makes me skeptical about what they're actually doing to animate them, and whether they've gone through the process of making multiple animations for a moderately complex mesh generated from this tool or just have a few toy examples.

It's still very impressive research, at least.

I can well imagine that reliably getting good topology out is a Hard Problem(tm).

There is work ongoing to animate gaussians:

https://www.youtube.com/watch?v=C4IT1gnkaF0 D3GA: Drivable 3D Gaussian Avatars

https://www.youtube.com/watch?v=q5hHZrRUV-k Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling

It's something, but I find it difficult to get excited about "we trained another NN to perform this specific transformation on this specific type of asset". Does it not generalize? What are you supposed to do if the NN doesn't get it right? It's an inscrutable black box on top of a stack of other inscrutable black boxes, which maybe make things easy if you hit the happy path but make it impossible to reason about the problem if it fails.

I can't say I'm a fan of replacing well defined systems with a big blob of weights that nobody really understands.

This is kind of the attitude that folks at companies had last decade (e.g. early 2010s) which prevented the adoption of deep neural networks until people could no longer overlook their efficacy.

If the NN doesn't get it right, it's absolutely not the end of the world. First of all, triaging the failures usually turns up something interesting / some connecting thread. And as for fixing issues, your options are typically:

1. Retrain with more examples to help it get it right (easier these days since you can take advantage of synthetic data / self- or semi-supervised approaches). Specifically it will help to generate or find examples that approximate the failures.

2. Shape your loss to penalize undesirable behavior / tendencies you've identified in triage

3. Add inputs or outputs that help your net solve the problem better (e.g. side inputs that may help the net not overfit to its available data)

These "inscrutable black boxes," while hard to edit directly, are not impossible to evaluate which means you can just measure to see if they're better than more explicit systems. And most real world systems are patchworks of ML + heuristics to give some desired level of performance.

Furthermore there is a big field of folks dedicated to elucidating what your net is learning (e.g. saliency). The tools are pretty cool -- for example in computer vision you can visualize whether your net is overfitting to background data in your training examples.

A valid critique of all of this is simply that it's expensive. Very expensive to build teams to warehouse and clean huge amounts of data, then build unbiased / fair model evaluations, then build models, then integrate them into real world stuff, usually to be used in conjunction with non-learned algos.

Another valid critique is that many companies have started using models they can't retrain or fix (e.g. off the shelf LLMs). This is dangerous territory since you lose so many of the options I outlined above.

When I say "what are you supposed to do" I'm speaking from the perspective of an end-user, in this case an artist trying to animate a 3D asset. With traditional 3D workflows all of the parameters are in the open, if something doesn't match the desired result it can be tweaked until it does, down to the most minute detail. With this NeRF workflow however, if the NNs aren't cooperating then all the user can do is raise it to someone else and wait to hopefully get a better NN back at some point. I'm just struggling to see what NeRFs bring to the table over existing workflows for it to be worth dealing with fuzzy logic - it seems like a solution looking for a problem.

I guess if you just don't care about the details of what you're making then AI could lower the barrier to producing something mediocre, as it has with 2D image generation, but that's damning with faint praise.

> I guess if you just don't care about the details of what you're making then AI could lower the barrier to producing something mediocre, as it has with 2D image generation, but that's damning with faint praise.

To me this sounds a lot like how people complained in 80s(?) about compilers generating worse code than expert assembly programmers. Yes, it is true, but also quantity can trump quality. And allowing machine to do bulk work, even if mediocre, allows humans to focus on key parts; in code some hot kernels, in 3d modeling maybe some hero assets. Also like compilers have evolved lot in the past decades, I don't see reason why these ML models/NNs/etc could not improve on a similar trajectory. Note that I'm not saying that they inevitably will improve to superhuman levels, but just that at this point they don't seem something that can be completely dismissed.

Furthermore I don't think it's faint praise if tool can enable non-experts to create compelling 3D experiences easily even if they are not perfect and photorealistic. Again I draw parallels to programming; lowering the skill level needed has allowed huge amount of valuable programs to be written even if the code behind them is complete crap

This is a matter of ergonomics and how close this thing is to production readiness; there's no reason the neural-net approach couldn't evolve and open up the parameters you want to tweak. Neural nets are just functions after all. You can refactor functions to expose params.

Without being an expert in this area, I imagine one such way would be to train a second neural net that learns to regress traditional workflow parameters to best approximate the NeRF output, then let a user post-edit those parameters to fine-tune.

Yup, just the fact that topology is often rather poor, you're making it impossible to translate motion capture data for a lot of things. Facial expressions come to mind.

I understand the reasons for different representations for reconstruction, but what we really need is to translate that into a mesh that is actually useful.

Its great to have something as a reference, don't get me wrong, but we are still far away from even being able to discreetly assign materials to the object's various parts.

Another big issue is licensing, most of these are not allowed commercially.

they're not free for commercial use. If you want to use them commercially, you get in contact with the research group that came up with it and pay them a pile of money to use it in your commercial product. For someone like Pixar, I can see them ponying up the cash for it if they think it'll allow them to save on animator hour costs and/or iterate faster.

I fear mesh maybe the modelling equivalent of OOP. Not because it is the best paradigm for all situations, but just through inertia of use over decades

I do a lot of Blender 3d modeling as part of my hobby gamedev stuff.

I think one thing that a lot of folks tend to forget is that quad and tri meshes are only one of SEVERAL 3d modeling techniques. It just so happens that this is the technique that the world landed on in the 90s due to how flexible it is, how performant it is, and algorithms like Catmull-Clark subdivision.

Remember, Toy Story 1 actually used NURB surfaces for all the characters! For a Bug's Life and Toy Story 2, they redid all the meshes as catmull-clark subdividable surfaces. And guess who was CEO and founder of Pixar?

There also techniques like NURBs, signed distance fields, and metaballs (all stuff you can actually use in Blender) for generating 3d surfaces. And of course the techniques used in CAD software.

With the right algorithm and sufficient compute, there could be some sort of completely alternate 3d representation in the future for creating virtual spaces.

This fictional documentary title just popped into my head:

"Cloudy with a Chance of Metaballs: A history of 3D modelling in animated film."

The link is broken for me (may be a transient error), alternative: https://github.com/GaussianObject/GaussianObject

Can you reconstruct a real 3d model from it? What is the accuracy?

As with all NeRFs and Gaussian Splatting, there is no classical mesh data. But there are projects and papers working on the extracting process. Specifically for Gaussian Splatting there is SuGaR. [1]

[1] https://github.com/Anttwo/SuGaR

Reply on news.ycombinator.com