What's the topology like? A lot of the time with this type of generation topology is too dense or high to be practically used in game, especially if you have thousands of assets. It's neat, but it'll take an actual artist to retopologise the model, unwrap it and texture it, at which point, for a lot of game assets, would be easier to just create the manually from scratch.
i go more in depth on the topology in this article: halkonyuncut.com/how-to-use-microsoft-trellis-models-for-game-development/ once you fix the non-manifold edges, you can decimate faces quite effectively
interesting, currently getting an error after 'generate' button.. I'd try the code myself but I'm on cuda 12 and I've found running 11 and 12 causes issues with a variety of things. love the concept though. if blender has issues with the high poly counts, I wonder about meshmixer ad 3dcoat for those steps. but it also looks like it has simplify settings in the export, so not sure about the problem someone else mentioned. would love to try it out though.
I'm currently running trellis locally with CUDA 12.2, the version they said they tested with in their docs. 12.6 would probably work fine though. 20 second inference on my 3090, uses maybe half my vram. I ran the .ps1s from this repo to install trellis and it worked like a charm: github.com/sdbds/TRELLIS-for-windows/ as to the other commenter, I looked into it and the models it spits out have non-manifold geometry, making poly manipulation awkward, but theres an easy fix. linked a blogpost about it in that comment.
@@jamesmurphy7115 was hoping there was a way to do it without using blender to streamline the workflow but that's what I've been doing regardless thanks for the quick response!
Tried using it, and it produces models that are unusable for games. They have errors due to the AI hallucinating, a high poly count and messy topology. When I tried to lower the poly count in Blender, the decimate modifier just garbled the model because it couldn't handle the topology. Sure, you can clean up the model manually, but that saves you perhaps 5-10% of total time spent. What this could be useful, to some degree, is allowing 2D concept artists to quickly visualize how the finished model might look like so they can improve the concept or show it to management. Much more niche, but at least viable.
Agreed. The models I've generated have had 8,000+ faces. However, if you check the "Editing" section of the Trellis page, they give an example prompt that creates a voxel pick up truck. That topology is much more predictable. That part of the model hasn't been released yet though, will post an update when I get my hands on it. Another niche use case I found is when training 2D image models. You can use the different perspectives from a render to give tools like Stable Diffusion much better training data.
I looked into the models some more and noticed there was non-manifold geometry. That's why decimates and subDs looked like crap. If you merge the verts by distance, the modifiers work just fine. I did a writeup with the solution on my blog: halkonyuncut.com/how-to-use-microsoft-trellis-models-for-game-development/
this fork worked perfectly for me: github.com/sdbds/TRELLIS-for-windows/ run the .psi marked with "1," then the "2," one with xformers. worked like a charm on CUDA 12.2
@@halkony Win 11 ? I was trying for so long and I get stuck , if you don't mine can you share your installed python version , pip list versions and stuff o.o where did you install it?
@@smoofwah3552 I'm on win10 with py3.11.6. it should work on win11 tho. clone the repo and run the .ps1 scripts in order and it will create a virtual environment with everything in the requirements.txt. i think i had to call the scripts from an admin console to get it working.
What's the topology like? A lot of the time with this type of generation topology is too dense or high to be practically used in game, especially if you have thousands of assets. It's neat, but it'll take an actual artist to retopologise the model, unwrap it and texture it, at which point, for a lot of game assets, would be easier to just create the manually from scratch.
i go more in depth on the topology in this article: halkonyuncut.com/how-to-use-microsoft-trellis-models-for-game-development/
once you fix the non-manifold edges, you can decimate faces quite effectively
Looks interesting
Thank u for sharing that with us 🙏🏻
Hope u the best
U got a new fan
Love this! Thank you
2024: The eyes of the ai generated chicken fbx look a little to seductive for me. 😂 The problems of the future for sure.
interesting, currently getting an error after 'generate' button.. I'd try the code myself but I'm on cuda 12 and I've found running 11 and 12 causes issues with a variety of things. love the concept though. if blender has issues with the high poly counts, I wonder about meshmixer ad 3dcoat for those steps. but it also looks like it has simplify settings in the export, so not sure about the problem someone else mentioned. would love to try it out though.
I'm currently running trellis locally with CUDA 12.2, the version they said they tested with in their docs. 12.6 would probably work fine though. 20 second inference on my 3090, uses maybe half my vram. I ran the .ps1s from this repo to install trellis and it worked like a charm: github.com/sdbds/TRELLIS-for-windows/
as to the other commenter, I looked into it and the models it spits out have non-manifold geometry, making poly manipulation awkward, but theres an easy fix. linked a blogpost about it in that comment.
Do you know if there’s a way to export the model as an FBX to work with Auto Desk software?
yeah - Import the GLB into Blender and export out of Blender as an FBX
@@jamesmurphy7115 was hoping there was a way to do it without using blender to streamline the workflow but that's what I've been doing regardless thanks for the quick response!
Sounds like Jake Tapper is commenting this video.
Can we run this on mobile with desktop mode?
Yes. The processing is done on external servers, so its like using any other web page.
Looks good for background objects and as 3d references but hero assets and anything animated would need to be modelled by skilled hands.
No, in close future AI does it quicker an cheaper
For the moment……. Not forever
Tried using it, and it produces models that are unusable for games. They have errors due to the AI hallucinating, a high poly count and messy topology. When I tried to lower the poly count in Blender, the decimate modifier just garbled the model because it couldn't handle the topology. Sure, you can clean up the model manually, but that saves you perhaps 5-10% of total time spent.
What this could be useful, to some degree, is allowing 2D concept artists to quickly visualize how the finished model might look like so they can improve the concept or show it to management. Much more niche, but at least viable.
Agreed. The models I've generated have had 8,000+ faces. However, if you check the "Editing" section of the Trellis page, they give an example prompt that creates a voxel pick up truck. That topology is much more predictable. That part of the model hasn't been released yet though, will post an update when I get my hands on it.
Another niche use case I found is when training 2D image models. You can use the different perspectives from a render to give tools like Stable Diffusion much better training data.
I looked into the models some more and noticed there was non-manifold geometry. That's why decimates and subDs looked like crap. If you merge the verts by distance, the modifiers work just fine.
I did a writeup with the solution on my blog:
halkonyuncut.com/how-to-use-microsoft-trellis-models-for-game-development/
You can drop the model into Unreal, enable Nanite, and stop worrying about the poly count.
@Metarig yeah, that's why UE5 Games run oh so smooth. Another case of "sounds great on paper, suck when actually used"
The website is trash, always gives an error no matter what image is set
ngl, i feel bad for the artists. though will come super handy for my own game as a compsci student
What a magnificence, thanks, will help a lot
I'm doing something wrong running this on windows xD so I guess I'll try Linux
this fork worked perfectly for me:
github.com/sdbds/TRELLIS-for-windows/
run the .psi marked with "1," then the "2," one with xformers. worked like a charm on CUDA 12.2
@@halkony Win 11 ? I was trying for so long and I get stuck , if you don't mine can you share your installed python version , pip list versions and stuff o.o where did you install it?
@@smoofwah3552 I'm on win10 with py3.11.6. it should work on win11 tho. clone the repo and run the .ps1 scripts in order and it will create a virtual environment with everything in the requirements.txt. i think i had to call the scripts from an admin console to get it working.
@@halkony I'm doing something totally wrong even the auto scripts don't work I just don't know what it is yet
I create much better 3D models on my channel than any of this AI junk! 💪
Then this tool is not for you simple as.
😂
Sure Jan