(Early Access) Assisted Culling (LrClassic)

I need to turn this feature off and keep it off. I don’t want it operating without my activating it.

Tried this for the first time yesterday and it did okay. It’s quite quick, and the ability to filter by approved or rejected, and to adjust standards on the fly is very convenient. I appreciate having it in its current form to help with an initial quick pass to reject the obvious “no” photos.

It struggles with some things that I think are viable to improve, though. Here are some thoughts I had when using it.
 

Show Aggregate Counts in Range Pickers

The sliders for subject focus and eye focus would be more helpful if they had some visual indication of where the values of the current selection lie. For example, a histogram range slider like is commonly seen in faceted search scenarios where the values indicated by the histogram are derived from the quantity of photos within each bucket. 

This lets me as a user quickly get a sense for how sharpness is distributed across the set, and therefore what value to choose as my divider between what I'm accepting and rejecting.

Contextual Rejection

Standards for rejection vary based on the options when a human is doing the culling. If there are many properly exposed options for a given composition, improperly exposed options are easy to reject. If all of the options in a composition are improperly exposed, I may try to make one of them work rather than rejecting them all outright. The current Assisted Culling doesn’t seem to have any mechanism for dynamic standards based on similarity despite automated stacking.

“Exposure Issues” 

In its current form, the exposure issue detection is either too sensitive or lacks awareness of the image content. It tends to flag photos of someone in dark clothing with moody shadows (where the skin is properly exposed) as underexposed, likely due to the histogram being heavy on darker values. Being able to tune that sensitivity with a slider like we can for sharpness might be enough to make the current heuristic more usable. Otherwise, a vision-capable model that can determine either the most relevant areas to consider for exposure or that can spit out a target EV for the photo may be worthwhile.

Inconsistent evaluation of near identical images

I have two frames that were taken of the same subject in studio that are exposed identically and use the same focus point, but where the subject is positioned about 5% further away from the center. For one, eyes were detected and the subject focus score is 36. For the other, no eyes are detected and the subject focus score is 64.

Thank you!

Quick question. Is there a way to get the assisted culling to run over the images again and re-score them from scratch?

First attempt for studio style portraits, it rejected closed eyes and flash misfires (underexposed). Good. This was a project that I had already completed so all my manual rejections were picked up by Assisted Culling. 

Algorithm rejected all the full body shots and it is saying subject and eye focus have lower score compared with the Selects. Possibly needs a ratio of body percentage present in the shot to normalise what in focus actually means. Fewer pixels per face area - definitely, but the full body shots are sharp.

Different type of shots with low light and high contrast. Rejected them all. But disable exposure issues and they become Selects. So it seems that blocks of images might need to be taken into consideration.  One third of the shoot was deliberately low light background, overall under exposed if we are thinking histogram. So it’s almost like I need stack-based settings eg. this block of images I don’t care about exposure, but for the others in the folder, I do want under exposure detection. So folder only settings might not take into consideration the variations of shooting style that happen in a session.

Where is the key to turn it off?

where in Preset menu to set or deselect it?

how much resources are kept by the function ?

Well lit headshot, 130 images. The feature missed most of the closed eyes, and all of the partially closed eyes. Doesn’t seem ready.

MacOS Sequoia, LrC 15.2

Have run this on thousands of photos, but it never finds the subject much less the eyes.  This needs alot of work.  It’s basically useless for my purposes.

Just updated to Lightroom Classic 15.2 and tried the "improved" AI-assisted culling. After running an entire studio photoshoot through it, my opinion hasn't changed since the last version - this feature is a total joke.

I shot a session on a Sony A1 at f/5.6, where 95% of the frames are tack sharp. Looking at the results, it feels like the AI is just generating random numbers instead of actually scanning the images for sharpness or detail. For example, in a series of nearly identical shots, the one where I clearly missed focus got a 95% score, while the perfectly sharp ones were rated under 90%.

I don't understand why Adobe is wasting time and resources on a "feature" that performs like a toy. It's completely unreliable for professional work. They should focus on improving actual AI editing models or performance instead of this.

Is anyone actually finding this useful for real-world sessions, or is it just a useless gimmick?

It’s working out fairly well but one thing that is happening is if I go into loupe on a photo and then go back into grid mode, that photo will disappear from the filter of possible rejects and it takes resetting the Assisted Culling filter to get it back on the list.

I would like a culling that removes duplicates.