The validation result are difference when train and test

Has anybody encountered that the validation results are wrong completely while the training is correct and while it’s no problem when place the validation process in the training process.
To clearly, the process is below,
1 i training segmentation and validation, the two process was separate that the training will store a model and val load the model, it’s very bad
2 i val on the process of training, which i place the val process in the process of training, it was no problem. And i use the model stored by training running the validation, it’s still be worst.
Thanks a lot for any help.