Is the model in the web interface different than the one released in the repository?

Hi. I was testing the code and I could notice that the model downloaded from the repository performs slightly worse than the one in your website interface.

For example in the attached image. The left one is the output of the website version, and the right one is the results using the code provided in the README. Both have the prompt "person".

Could you please tell if this is due to a difference in the model or a different post-processing.

Thank you!

Start with the inference code and model-download instructions provided in the README, then compare their output for the "person" prompt with the website interface shown in the attached image. Done means determining whether the discrepancy comes from the released checkpoint or post-processing and documenting the evidence and conclusion in the issue.

Written by the indexing model from the issue text.