Substantial performance degrade in image recognition from TM web results

[USeebi]

A Standard Image model is built from Upload image and the results are almost perfect from Under the hood. The model is tried with additional samples outside the training/validation sets and the tests are still satisfactory.

A Tensorflow.js model is exported but neither the Upload shareable link nor Download my model could reproduce the test results with exactly the same test samples in p5.js with substantial degradation. What did I miss and where to start troubleshoot?

Thanks in advance for any hints.