Using Photo AI 1.3.0 on windows. I am trying to recover really old photos of varying quality. To be honest I am blown away with the face recovery results. Truly amazing. There are a couple of things I noticed that could be improved.
In several cases the program added a beard or stubble where no beard existed. I even saw this on a woman although must admit she wasn’t the most attractive. My work around here was to blur the chin area prior to loading in Topaz. Maybe the face library being used has a lot of men with beards. It would be nice to be able to selectively filter the library to remove faces with beards. (Alternatively improve the stubble detection capability)
The second thing I am seeing is the program’s willingness to add spectacles to a subject where no specials exist. I think this is in an area that could be improved since in most cases only one lens is added and in no cases are the arms of the spectacles added. If the program is unable to recognize a compete set then it should not add half of one.
Hey @john.ellis-2566! This is a behavior we’re aware of, and just something I think we’ll need to improve as we continue training new face recovery models – if you have examples that you’d be willing to share with us (your original image + the processed ones) I can pass it along to our developers for consideration when they look to make updates to the models.
Files uploaded. The wedding took place in 1949. No men at this time would go to wedding with stubble on chin. Man just behind the bride has stubble and half an eyeglass. Several others show stubble. Other picture take at same event. Again stubble on chin under nose
On a similar note - Topaz Photo AI 1.2.9 produced grizzly-looking distorted human faces from the blurred feathers on the side of the head of a little grebe (a bird)
@preston.werner would you like me to upload the files for inspection?
Hi @preston.werner and thank you for your interest in this matter! I uploaded the original NEF and the DNG produced by Topaz 1.2.9. Look at the grebe on the left
Thanks for sharing the files. Topaz Photo AI is using the detail that is in the original image to create faces. If the starting image quality is low, there will be uncertainty about what the original face looked like.
The Face Recovery model is making a best guess from the information available. These older images with texture and other patterns on film are particularly difficult to get a great result from.
Are you able to smooth out the faces somewhat in your pre-Topaz Photo AI to prevent the AI model from interpreting the texture as stubble?
At the moment it would be difficult to add a feature to customize the faces that are created. It would get complicated very quickly so if possible I’d like to find a workaround.
