Which interactive visualizer is supported?

[pzarzycki]

Hi!

Setup

I managed to run the model with scripts/train_custom.sh and it produced expected output folder with cameras*.json and *.ply files.

  SSIM :    0.5906495
  PSNR :   19.4915962
  LPIPS:    0.2606166

Test renders are really good!

Issue

But I cannot find any working method to visualizise the results :/

I tried:

• SIBR_gaussianViewer_app viewer (https://github.com/graphdeco-inria/gaussian-splatting)

  • I copied outputs folder from Ubuntu to Windows and tried every possible args combination, including some modificaton for cfg_args (e.g. removing input_path). I can't make it working, it either crashes or complaines on input data

• Blender 4.5.1 with "3DGS Render by KIRI Engine" (3.0.0)

  • Only Point cloud is displayed, crash when attempting to expoer 3dgs

• https://superspl.at/editor

  • failes to import .ply

• https://antimatter15.com/splat/

  • cant load the .ply data

Any hints?
Thanks!

Output content

ls -R outputs/data/walk/2025-08-26_13\:16\:56/
'outputs/data/walk/2025-08-26_13:16:56/':
cameras_all_test.json   cfg_args                                           outputs.log    point_cloud   test
cameras_all_train.json  events.out.tfevents.1756239419.H310M-gpu.111805.0  per_view.json  results.json  train

'outputs/data/walk/2025-08-26_13:16:56/point_cloud':
iteration_40000

'outputs/data/walk/2025-08-26_13:16:56/point_cloud/iteration_40000':
color_mlp.pt  cov_mlp.pt  opacity_mlp.pt  point_cloud.ply

'outputs/data/walk/2025-08-26_13:16:56/test':
ours_40000

'outputs/data/walk/2025-08-26_13:16:56/test/ours_40000':
depths  gt  nvs  per_view_count.json  poses  renders  videos

'outputs/data/walk/2025-08-26_13:16:56/test/ours_40000/depths':
*****.png
*****.npy

'outputs/data/walk/2025-08-26_13:16:56/test/ours_40000/gt':
*****.png

'outputs/data/walk/2025-08-26_13:16:56/test/ours_40000/nvs':

'outputs/data/walk/2025-08-26_13:16:56/test/ours_40000/poses':

'outputs/data/walk/2025-08-26_13:16:56/test/ours_40000/renders':
*****.png

'outputs/data/walk/2025-08-26_13:16:56/test/ours_40000/videos':
depth.mp4  render.mp4

'outputs/data/walk/2025-08-26_13:16:56/train':
cameras.txt  images.txt  ours_40000

'outputs/data/walk/2025-08-26_13:16:56/train/ours_40000':
depths  gt  nvs  per_view_count.json  poses  renders  videos

'outputs/data/walk/2025-08-26_13:16:56/train/ours_40000/depths':
*****.png
*****.npy

'outputs/data/walk/2025-08-26_13:16:56/train/ours_40000/gt':
*****.png

'outputs/data/walk/2025-08-26_13:16:56/train/ours_40000/nvs':
*****.png

'outputs/data/walk/2025-08-26_13:16:56/train/ours_40000/poses':

'outputs/data/walk/2025-08-26_13:16:56/train/ours_40000/renders':
*****.png

'outputs/data/walk/2025-08-26_13:16:56/train/ours_40000/videos':
depth.mp4  gt.mp4  nvs_depth.mp4  nvs_rgb.mp4  render.mp4

[orgicus]

TLDR;
For now, the Scaffold-GS viewer could work after using the convert.py script on the dataset (accounting for images vs input folder names.

@linjohnss Thank you for the continued support. Is my assumption around Scafford-GS correct ? Regarding the screenshot below is that quality expected or are we missing a few tricks ?

More context:

I also had a look for the interactive viewer and couldn't see in the repo, although I can see how the videos are rendered.

An interactive viewer would be nice.

The README does reference Scaffold-GS.
There is a Windows build readily available which works with one of the sample datasets, e.g.:
Scaffold-GS-Viewer\bin>SIBR_gaussianViewer_app.exe --m "E:\DL\LongSplat\outputs\Tanks\Museum\baseline" --s "E:\DL\LongSplat\data\Tanks\Museum"

I did run into errors using it on my custom trained model:

SIBR_gaussianViewer_app.exe --m "E:\DL\LongSplat\outputs\MyCustomModel" --s "E:\DL\LongSplat\data\MyCustomDataset"
[SIBR] --  INFOS  --:   Initialization of GLFW
[SIBR] --  INFOS  --:   OpenGL Version: 4.6.0 NVIDIA 546.65[major: 4, minor: 6]
[SIBR] --  INFOS  --:   Dataset type:
[SIBR] ##  ERROR  ##:   FILE H:\SIBR_viewers\src\core\scene\ParseData.cpp
                        LINE 589, FUNC sibr::ParseData::getParsedData
                        Cannot determine type of dataset at E:\DL\LongSplat\data\MyCustomDataset
[SIBR] --  INFOS  --:   Did not find specified input folder, loading from model path
[SIBR] --  INFOS  --:   Dataset type:
[SIBR] ##  ERROR  ##:   FILE H:\SIBR_viewers\src\core\scene\ParseData.cpp
                        LINE 589, FUNC sibr::ParseData::getParsedData
                        Cannot determine type of dataset at E:\DL\LongSplat\outputs\MyCustomModel

My observations/questions are:

1. Should we use the Scaffold-GS viewer or wait for a LongSplat viewer release ?
2. The Tanks/Museum folder does contain the sparse folder which looks like a standard COLMAP project. The viewer doesn't seem to use poses_bounds.npy, dpt, database.db. We simply need to use the convert.py script (temp renaming or symlinking the images folder to input), right ? (I tested this and it works, however output doesn't look great)
3. I see a lot of lines in the renders when the camera pose isn't closely aligned with gt poses. Is this expected ?

(Also there seems to be a coordinate space mismatch with Scaffold-GS viewer: when using the viewer's snap to +/- controls it appears the camera poses don't match (some axes swapped or rotated 90 degrees ?)

[pzarzycki]

The newly added converter to 3DGS works great!
I was able to open converted .ply in WebGL GS viewers.

python convert_3dgs.py -m PATH_TO_TRAINED_MODEL --prune_ratio 0.6

Additionally, with 3DGRUT converter .ply -> .usdz I was also able to convert it further to USDZ and open scene in Isaac Sim:

python -m threedgrut.export.scripts.ply_to_usd ./point_cloud.ply --output ../scene.usdz

Thank you for this awesome contribution!