Detailed configuration for Deterministic MedCLIPSeg

[WangKai-449]

Congratulations on MedCLIPSeg being accepted to CVPR 2026. I found your work very inspiring. I am currently trying to reproduce the Deterministic MedCLIPSeg setting, and I would like to ask whether you could share the detailed experimental configuration used in your paper.
Thank you very much for your help.

[TahaKoleilat]

Hi @1300092160,

Thank you for your interest in our work. The deterministic version is the same as the original implementation with the only difference being that it uses the vanilla Cross-attention instead of the proposed Probabilistic Cross-attention here. You can keep the configuration the same as the original implementation since the training is pretty stable and you will have to set the number of samples to 1 like here since there is no variation.

Best,
Taha

[WangKai-449]

Hi @1300092160,

Thank you for your interest in our work. The deterministic version is the same as the original implementation with the only difference being that it uses the vanilla Cross-attention instead of the proposed Probabilistic Cross-attention here. You can keep the configuration the same as the original implementation since the training is pretty stable and you will have to set the number of samples to 1 like here since there is no variation.

Best, Taha

Thank you for your reply. Could you share the code for this part? I tried using standard cross-attention with the number of samples set to 1, but the gap from the results reported in your paper is still quite large. It is possible that I made some mistakes in the implementation details.