support for >800 faces

[yosun]

seriously.

[buaacyw]

It takes more computation resources for more faces.

[yosun]

it's useless for 800 faces since when u want to convert a bad 3d scanned mesh to "artist made", you start with a high poly mesh

[buaacyw]

This is an academic project; you can't expect it to go from nothing to industry use overnight, right?

[matbee-eth]

This is an academic project; you can't expect it to go from nothing to industry use overnight, right?

Is there any type of continued pretraining or finetuning we could do to achieve > 800?

[yosun]

This is an academic project; you can't expect it to go from nothing to industry use overnight, right?

well, you're certainly great at marketing, even if your research is basically a toy project.

i don't understand why people are excited by the results when existing procedural mesh cleanup techniques are cheaper, faster, and more useful than this.

[buaacyw]

This is an academic project; you can't expect it to go from nothing to industry use overnight, right?

Is there any type of continued pretraining or finetuning we could do to achieve > 800?

Hi! Thanks for your interest. Like LLMs, simply prolong the size of positional encoding vector and train the model on a mesh dataset with meshes more than 800 faces.

[buaacyw]

This is an academic project; you can't expect it to go from nothing to industry use overnight, right?

well, you're certainly great at marketing, even if your research is basically a toy project.

i don't understand why people are excited by the results when existing procedural mesh cleanup techniques are cheaper, faster, and more useful than this.

May I ask which specific methods are referred to by 'existing procedural mesh cleanup techniques'? I will make sure to study them thoroughly.

[yosun]

i'm not sure at what level to start the discussion.

here is a very old demo (no it's not the gen AI InstantMesh from tencent) - it's any mesh to smooth quads from 10 years ago: https://github.com/wjakob/instant-meshes

since then, there have been many advances in procedural non-AI (cheap computationally, consumer GPU takes ms etc) conversion of messy meshes to smooth quads...

in this context, it would be great to see how the results of your solution compare to existing procedural mesh solutions.

hope that by sharing some oldskool techniques, we can more sooner get to the common industry use cases of generated meshes with 800+ faces being usable in your solution.

[matbee-eth]

This is an academic project; you can't expect it to go from nothing to industry use overnight, right?

Is there any type of continued pretraining or finetuning we could do to achieve > 800?

Hi! Thanks for your interest. Like LLMs, simply prolong the size of positional encoding vector and train the model on a mesh dataset with meshes more than 800 faces.

Curious if you've outlined a strategy for doing this?

Perhaps we could do some mass scraping of free 3d models. I have 3 gigabit internet + 100tb of disk space, and experience in scraping.

[buaacyw]

This is an academic project; you can't expect it to go from nothing to industry use overnight, right?

Is there any type of continued pretraining or finetuning we could do to achieve > 800?

Hi! Thanks for your interest. Like LLMs, simply prolong the size of positional encoding vector and train the model on a mesh dataset with meshes more than 800 faces.

Curious if you've outlined a strategy for doing this?

Perhaps we could do some mass scraping of free 3d models. I have 3 gigabit internet + 100tb of disk space, and experience in scraping.

Thanks for reaching out, but I think ObjaverseXL already has enough 3D models for training.