Mesh Rendering

[csabs3d]

Hello,
Thank you so much for this, truly the best quality mesh results I have come accross.
I'm running it inside docker with a custom gardio app. I have an rtx 3060 12gb gpu and 32gb ram. The training in most cases rather on the fast side. It can take from 30mins to 90mins but I'm okay with that because the mesh is outstanding.
My issue comes during rendering unbounded mesh (I couldn't manage to dail in bounded properly). If I'm using around 150-300 4k images the rendering quietly crashes docker as it's running out of memory. With lower resolution images ie 720p 1080p I can render at the highest resolution but mesh will be much less detailed.
Would it be possible to somehow fix this please? I assume it's storing everything in the gpu memory during extraction and higher resolution fills up the memory causing it to crash.
I'd love to see the results from larger 4k imagesets and also happy to share the gardio app if you wanted to optimise it and add it to your repo

[hbb1]

Hi. Could you share the detailed error information? Was it caused by the Out-of-Memory on the GPU side?

[csabs3d]

I assume it's out-of-memory on the gpu since docker is showing 11.3gb in use and then it jumps to 0 and just stalls there.
This happens if I start training with 400+ 4k images, if I train 400 2k images then I can render unbounded mesh up to around 1240 resolution or with 1080p images I can go as high as 2048 resolution for rendering but the mesh becomes less detailed with images less then 2k.
Because I'm running it in gardio app I don't get for error logs but I will run some tests with commands only and I will get back to you with the error logs.

[csabs3d]

Here's some tests I ran, I didn't stop any of the processes they just stop without any error output.

4k images with just train.py:

Reading camera 339/339 [16/10 16:16:30]
Loading Training Cameras [16/10 16:16:31]
[ INFO ] Encountered quite large input images (>1.6K pixels width), rescaling to 1.6K.
If this is not desired, please explicitly specify '--resolution/-r' as 1 [16/10 16:16:31]
Killed

4k images with resolution 1:

Reading camera 339/339 [16/10 16:24:47]
Loading Training Cameras [16/10 16:24:47]
Killed

2k images

ptimizing
Output folder: ./output/df8b04d6-0 [16/10 16:34:36]
Tensorboard not available: not logging progress [16/10 16:34:36]
Reading camera 569/569 [16/10 16:34:51]
Loading Training Cameras [16/10 16:34:51]
[ INFO ] Encountered quite large input images (>1.6K pixels width), rescaling to 1.6K.
If this is not desired, please explicitly specify '--resolution/-r' as 1 [16/10 16:34:51]
time="2025-10-16T17:40:51+01:00" level=error msg="error waiting for container: unexpected EOF"

169 images 1684x946

Training progress: 100%|█| 20000/20000 [1:15:23<00:00, 4.42it/s, Loss=0.02661, distort=0.00000, normal=0.00428, Points=
[ITER 20000] Saving Gaussians

--mesh_res 2048

Looking for config file in /workspace/output/6d4e620f-f/cfg_args
Config file found: /workspace/output/6d4e620f-f/cfg_args
Rendering /workspace/output/6d4e620f-f
Loading trained model at iteration 20000
Reading camera 169/169
Loading Training Cameras
[ INFO ] Encountered quite large input images (>1.6K pixels width), rescaling to 1.6K.
If this is not desired, please explicitly specify '--resolution/-r' as 1
Loading Test Cameras
export training images ...
reconstruct radiance fields: 169it [00:17, 9.76it/s]
The estimated bounding radius is 4.11
Use at least 8.22 for depth_trunc
export images: 169it [02:41, 1.05it/s]
export mesh ...
reconstruct radiance fields: 169it [00:16, 10.39it/s]
The estimated bounding radius is 4.11
Use at least 8.22 for depth_trunc
Running tsdf volume integration ...
voxel_size: 0.0040142045724376706
sdf_trunc: 0.020071022862188353
depth_truc: 8.22109096435235
TSDF integration progress: 12it [00:30, 1.91s/it]