Image embedding using own backbone model are of lesser quality than demo

Hi and thanks for sharing this fabulous work!
I've been trying to repro the results from the demo.
It seems that the resulted image embedding that returns from the backbone model at modelzoo is different than the one returned when hosting my own backbone model. I'm hosting vit h model.
When using these two different embedding with onnex and gui, the differences becomes very apparent. It seems like the modelzoo returns a much superior result for the embedding. When trying to select some areas using gui I'm getting a lot of missing data/holes for my embedding where things just work perfectly with the model hosted in modelzoo
Would appreciate your help with understanding:
1. Which version of model is hosted in model zoo
2. Is the inference code for getting an embedding any different or contains some confit edits to obtain those results
3. Any other factors you can think of that could contribute to obtaining the superior result
Much thanks about this work!

Compare the modelzoo and self-hosted ViT embedding paths, including the hosted model version, inference code, and any configuration differences. Reproduce the discrepancy through ONNX and the GUI; done means identifying the factor that causes the missing regions or documenting equivalent settings.

Written by the indexing model from the issue text.