Model training in Image reconstruction problem

I am working on an image reconstruction problem where I am generating single images. I have saved the model for each epoch and tested using five image. The images are generated for some modesl(such as models saved at epochs 171) but just for the model on 172 or 173 epoch, the reconstructed images are very obscure or completely dark. Again, the model saved at epoch ~200 can reconstruct almost perfect images. I am wondering if it means that my model is not really learning properly? I want to mention that the training loss and validation losses are decreasing.

Any suggestions would be much appreciated!

Does this mean that the model stored at other epochs close to 172 is creating “good” outputs while only these two epochs are “bad”? Did you see a loss spike for these epochs?