Unfortunately, there aren't a lot of relevant data to photo restoration, but imo you can replicate them/ generate your own. I suggest choosing a starting data-set similar to your target data, so if you are trying to restore portrait images maybe the IMDB-WIKI – 500k+ face images with age and gender labels is a good choice. Using OpenCV you can convert these images to black and white (gray scaled) and then you can merge/ blend these images with different old pictures textures. For that see
Adding (blending) two images using OpenCV
.
With some further tweaking (some randomized degradation and augmentation) the resulting images should be quite realistic and close to real data.
Alternatively, you can use restoration tools/ codes like the Algorithmia- API to restore pictures and build a dataset but that will not result in a better restoration neural network than theirs.