Copy link to clipboard
Copied
I've noticed that the AI can't find "People" in my shots when the exposure is low. If I round-trip the shot into PS or something and it comes back brighter, the AI masks the people just fine. This suggests the AI isn't adjusting the luminescense to find people. Please increase the exposure a few times as part of the routine if the AI isn't finding people.
As shot to protect the highlights
I increased exposure in LR Develop module. Still no People found
Baked the exposure by converting the raw to jpg and the AI has no issues finding the People.
Copy link to clipboard
Copied
Why did you not correct/adjust the tonality of the image globally BEFORE making local (Masking) adjustments? This appears to be a workflow issue.
Followup/ from replys, sounds like not a workflow issue, sounds like a bug.
Copy link to clipboard
Copied
Adjusting the exposure by whatever means won't help the AI find people in the orginal raw if it is underexposed as you can see in the screenshots. I have seen the same thing too. It appears the AI works on an unadjusted version of the image always even if you preadjust everything before going into masking. The trick to get it to work is to roundtrip it into Photoshop with the adjusted exposure. If you do nothing in Photoshop and simply save as is, it will now recognize people and objects. This is only an issue for severely underexposed images. If it is just one stop or so underexposed for the people it will still work fairly well.
Copy link to clipboard
Copied
in my post above, the middle photo is the photo after adjustment (raw) showing the AI does not attempt to use this adjustment as an input. Seems like the AI could run the RAW as input first and, when it fails to find People, run the proxy (visible version I see after adjustments in LR) from cache.
Copy link to clipboard
Copied
I suspect, and perhaps you can test, that if you bring the photo into PS, and attempt to use the mask in PS, it will fail as well? From your followup post and the reply by Jao vol, sounds like issue in ACR.