Out of Memory (OOM) Errors during mesh generation and refinement

[realstevewarner]

First, let me say that I love this implementation of Trellis2. The ability to generate a mesh, then refine it, then texture it is simply brilliant. That said, I'm running into constant OOM errors even when resolution is set to something low like 512. I'm on a 4090 so I've got 24GB of VRAM but during Refinement or Remeshing it will occasionally spike the VRAM and then everything stops.

I've been able to run model generation and the refiner before, so it really does depend on the model, but OOM is happening more often than not, which is really limiting the use of this tool.

I've tried with Low VRAM turned on and off as well as keep models in memory on and off.

I have a custom build of Comfy just for running Trellis.

VRAM hovers around 15% under most circumstances even during initial stages of mesh generation, but it will then spike and kill things. Is it possible to offload to normal system RAM?

[Thelmos]

I have a similar problem during "Dual Contouring" I don't know if this is your error.
The process has several peaks in VRAM consumption, but the one in this step is usually the biggest.

[Thelmos]

Keeping "dual_contouring_resolution" at 1024 solves the problem in my case, 2048 takes up all the video memory.

[realstevewarner]

Thank you! Yes, I found the same thing. Setting resolution to 1024 works. It's disappointing that on a 4090 we can't get Dual Contouring to work. But I found that the Refiner nodes that VisualBruno added are absolutely amazing and do a remarkable job of cleaning up the initial shape generation. I posted a few results on X here:

https://x.com/SteveWarnerFL/status/2006156726735626423

[visualbruno]

About this OOM exception in "Cumesh remeshing", I'm testing something to improve it.

[visualbruno]

Can you try again with the last commit?
Don't forget to update Cumesh and O_voxel packages

I improved the "Remeshing", you should have less OOM error message

[realstevewarner]

Thank you! I'm tied up today on other tasks but will dive into this tomorrow morning and see if it resolves the issue.

[realstevewarner]

@visualbruno Holy crap. It's working! I'm now getting 1536 generation on my 4090 with no OOM errors.

I am, unfortunately, still getting OOM errors on mesh refinement, however. I fed a 1024 resolution model into the Refiner and attempted to refine it at 1024 resolution. It gave me an OOM.

Is there any way to fix the refiner so it doesn't throw OOM errors?

[Thelmos]

I'm trying to test the new version but new Cumesh and O_voxel wheels instalation failed, it seems they forced torch-2.9.1 install:

python_embeded\python.exe -m pip install ComfyUI\custom_nodes\ComfyUI-Trellis2\wheels\Windows\Torch270\cumesh-0.0.1-cp312-cp312-win_amd64.whl --force-reinstall
....

Attempting uninstall: torch
Found existing installation: torch 2.7.0+cu128
Uninstalling torch-2.7.0+cu128:
Successfully uninstalled torch-2.7.0+cu128
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╸━━━━━━ 11/13 [torch] WARNING: The scripts torchfrtrace.exe and torchrun.exe are installed in 'E:\Dev\ComfyUI_windows_portable\python_embeded\Scripts' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
Attempting uninstall: cumesh
Found existing installation: cumesh 0.0.1
Uninstalling cumesh-0.0.1:
Successfully uninstalled cumesh-0.0.1
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchaudio 2.7.0+cu128 requires torch==2.7.0+cu128, but you have torch 2.9.1 which is incompatible.
torchvision 0.22.0+cu128 requires torch==2.7.0+cu128, but you have torch 2.9.1 which is incompatible.
xformers 0.0.30 requires torch==2.7.0, but you have torch 2.9.1 which is incompatible.
Successfully installed MarkupSafe-3.0.3 colorama-0.4.6 cumesh-0.0.1 filelock-3.20.2 fsspec-2025.12.0 jinja2-3.1.6 mpmath-1.3.0 networkx-3.6.1 setuptools-80.9.0 sympy-1.14.0 torch-2.9.1 tqdm-4.67.1 typing-extensions-4.15.0

---

python_embeded\python.exe -m pip install ComfyUI\custom_nodes\ComfyUI-Trellis2\wheels\Windows\Torch270\o_voxel-0.0.1-cp312-cp312-win_amd64.whl --force-reinstall
......

Attempting uninstall: torch
Found existing installation: torch 2.9.1
Uninstalling torch-2.9.1:
Successfully uninstalled torch-2.9.1
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╸━━━━ 16/18 [torch] WARNING: The scripts torchfrtrace.exe and torchrun.exe are installed in 'E:\Dev\ComfyUI_windows_portable\python_embeded\Scripts' which is not on PATH.
Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
Attempting uninstall: o-voxel
Found existing installation: o_voxel 0.0.1
Uninstalling o_voxel-0.0.1:
Successfully uninstalled o_voxel-0.0.1
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
numba 0.63.1 requires numpy<2.4,>=1.22, but you have numpy 2.4.0 which is incompatible.
opencv-python 4.12.0.88 requires numpy<2.3.0,>=2; python_version >= "3.9", but you have numpy 2.4.0 which is incompatible.
opencv-python-headless 4.12.0.88 requires numpy<2.3.0,>=2; python_version >= "3.9", but you have numpy 2.4.0 which is incompatible.
torchaudio 2.7.0+cu128 requires torch==2.7.0+cu128, but you have torch 2.9.1 which is incompatible.
torchvision 0.22.0+cu128 requires torch==2.7.0+cu128, but you have torch 2.9.1 which is incompatible.
xformers 0.0.30 requires torch==2.7.0, but you have torch 2.9.1 which is incompatible.
Successfully installed MarkupSafe-3.0.3 colorama-0.4.6 easydict-1.13 filelock-3.20.2 fsspec-2025.12.0 jinja2-3.1.6 mpmath-1.3.0 networkx-3.6.1 numpy-2.4.0 o-voxel-0.0.1 plyfile-1.1.3 setuptools-80.9.0 sympy-1.14.0 torch-2.9.1 tqdm-4.67.1 trimesh-4.11.0 typing-extensions-4.15.0 zstandard-0.25.0

[Thelmos]

I managed to restore the configuration and reinstall the wheels without the dependencies (--no-deps).

However, the memory problem persists in Remesh with dual_contouring_resolution=2048

[Rasflanit]

I've noticed that most Trellis 2-related nodes are having low memory issues. I solved this problem by running the program at the minimum resolution first, and then setting the highest. After that, the problems disappeared. In my case, the error occurs at the remesh stage, which makes seven passes at high resolution, and on the seventh pass this error appears. But I have another problem with the fact that the grid doesn't turn out to be whole anyway. This problem is floating, and it may appear after a couple of low-resolution generations, my ana disappears.