Part of any good photo restoration is the ability to correctly prompt your model. In fact, with generative models, what happens is that the system actually recreates every aspect of the image (kind of like Star Trek transporter technology... breaks it down and reassembles). When prompted correctly, the results we're able to get back are spectacular.
JOY CAPTION is outstanding in it field as a "descriptor" model (Image-Text)
Also heck out the improved Alpha-One (All Prompts) version! Now with both formal and informal mode - here