Interactive Demo Results Not Matching Results on Webpage

[joshmansky]

Hello,

I am trying to reproduce the text-based scene query result shown on the SceneSplat webpage for scene 09c1414f1b, specifically the vacation / art example.

I downloaded the 09c1414f1b scene and the released language checkpoint from Hugging Face:

lang-pretrain-concat-scan-ppv2-matt-mcmc-wo-normal-contrastive.pth

Then I ran standalone inference and loaded the output feature tensor in the mini viewer. However, my query result for vacation does not match the qualitative result shown on the webpage. In particular, the surrounding magazines/books also receive high similarity, so changing the threshold does not cleanly isolate only the travel/vacation magazine like the webpage example.

Website/reference result:

My result:

I have a few questions:

1. Was the webpage qualitative result generated using the released Hugging Face checkpoint, or a different/internal checkpoint?
2. What threshold value was used for the text-query visualizations? Was the same threshold used for all examples, or tuned per query?
3. Was vacation / art evaluated as two separate prompts (vacation and art) or as one literal prompt string?
4. Was any extra postprocessing/smoothing applied to the query highlights in the webpage render?
5. Is there a specific inference seed or command you recommend for reproducing the qualitative website results?

For context, I am running on Windows with WSL and an RTX 3090 Ti. I made some local changes to avoid OOM during inference by accumulating the full feature tensor on CPU rather than GPU, and I also made viewer-side changes to display the scene from WSL in my Windows browser. I do not expect those changes to affect the saved language feature tensor, but I wanted to mention them in case there is a detail I am missing.

Any guidance for reproducing the webpage result would be greatly appreciated.