1
$\begingroup$

There seem to be multiple aproaches to generating 3d objects from text prompt. What's confusing is that some of them are generating NeRFs (https://arxiv.org/pdf/2308.16512), other's are generating meshes (https://arxiv.org/abs/2404.07191). However, those tasks seem to be very different in their nature (meshes are direct 3d representation and NeRFs compress information about 3d object).

However I don't really see strong division between those tasks in literature (all will call it 3d generation). Is there something I'm missing, and are they actually very similar tasks? Or are they not comparable to each other?

$\endgroup$

1 Answer 1

0
$\begingroup$

After a bit of research, I believe the conclusion is following (June 2024):

While, it is confusing that output of some gen-3d papers are meshes and some are NeRFs or some kind of volumetric data, most of them are NOT directly outputting meshes as they are an awkward representation for learning. Rather they keep some kind of intermediate representation of volumetric data (like triplane-NeRF) and then using algorithms (like marching cubes) to generate meshes.

It is possible to optimize meshes directly, via differentiable mesh generation from "Signed Distance Field" or triplanes - that's what "InstantMesh" is doing.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.