INP-Former Performance

[yuanzhao-CVLAB]

Good Job! I noticed that while INP-Former outperforms Dinomaly on MVTec-AD and Real-IAD datasets (as reported in INP-Former paper), its performance on the M2AD dataset—particularly on M2AD-Synergy and M2AD-Invariant—is significantly lower than that of Dinomaly. May I kindly ask if you have any insights into the reasons behind this discrepancy?

[caoyunkang]

Thank you very much for your observation and interest. We believe the current performance gap may stem from unresolved issues in adapting INP-Former to the M2AD dataset—particularly in the synergy and invariant settings. As such, the reported performance of INP-Former may underestimate its true potential on our benchmark. We are actively investigating this and working to resolve the implementation bugs.

[yuanzhao-CVLAB]

Thank you for your reply. I look forward to the good news!