On the issue of differences in training visualization and testing effectiveness

[GAPLX]

I used my own dataset for retraining, and the visualization results during the training process were excellent. However, when I tested using the saved model, the results were very poor. Can you provide guidance on where the problem lies? I have been troubled for a long time.

The following is a comparison of visual and testing image effects，the first two are the visualization results during the training process, while the last two are the results of model testing: