P: Use AI to automatically rotate images to the correct orientation

That 76% exact-angle number lines up with what we keep finding too — "does this need rotation" and "which exact amount" are genuinely different-difficulty problems. We see almost the same split in our own numbers: general models are decent at flagging that something's off but specifically confuse 90° with 270° (documented independently in RotBench, EACL 2026) — the model's directionally right that a turn is needed, just picks the wrong one half the time. Your 95/76 split is a clean, independent confirmation of that same pattern from a completely different test set. Nice data.

Relevant data point if it helps: I've been building a dedicated rotation-detection model (not a general vision LLM) and get 99% on clean quadrant rotations, ~97% across the full realistic range including fine-tilt corrections, down to ~84% on the hardest cases (severe tilt, no clean quadrant to anchor on). Makes sense Gemini's weaker on the exact angle - general vision models aren't trained specifically for this. Happy to share more if useful - BlueVeta.com/upright if you want to see the real numbers.

[View this post in your web browser. It contains formatting and images that don't appear in email.]

Using my Any Vision plugin, I tested how well Google Gemini might guess the correct amount a photo needs to be rotated to have the correct visual orientation (0, 90, 180, or 270 degrees). I used a test set of 318 diverse photos (portraits, landscapes, cars, animals, plants, etc.).

Gemini did very well at detecting whether some rotation was needed, getting it correct 95% of the time.

But it didn’t do so well at determining the right amount of rotation to apply (0, 90, 180 or 270), getting that right only 76% of the time.

This suggests that a model specifically trained on rotation might do acceptably well.

I tested several prompts. The best one was:

Does the photo need to be rotated left, rotated right, or rotated upside down to appear in the correct visual orientation?  Output one of the following, with no other words:
No rotation
Rotate left
Rotate right
Rotate upside down			

In my lifetime of experience, there can be no perfect AI solution for this.  Sure, images with people, trees and other such easily recognizable objects can be handled correctly, but what about abstract art, blurry smears and solid colors?  My request is for a quality solution, as best as it can be.  But no matter how good it is... I would never trust it, and would find myself checking all of the results regardless.

Your point is made.  This new feature can simply be applied to a selection.  Then, if desired, the user can verify the results using the existing toolset.  So re-working the toolset is not required.

Why do you request and defend an inferior solution? This is a feature request, so request a quality solution! Don't settle for something that helps a bit but still requires extra work. Ask for something that is so good that you don't need all this. You can always manually correct the very few mistakes if and when you see an image that is still not rotated correctly. This would not be an unreasonable request. AI is already good enough that it should make very few mistakes in this case.

This is really quite funny. You say 'please add this feature, but do not make it flawless. Make it so that I still have to waste some time checking the results". 😁

I see the "one as a time" as a Loupe View where you can very quickly right arrow your way through many images quickly.  Just a bigger version of the grid view.  You would have the option to show only the suggested changes. While this would be similar to what we have now, the AI would add the time saving benefit of having pre-rotated the majority (if not all) of the images that needed to be rotated.  The user would still have the ability to rotate them manually.  For any image the AI has changed, LRC would somehow indicate this to the user by, for example, setting the suggested changed images with a color label, or some other indicator that made them stand out from the others.

I didn't comment on "or in large batches", I commented on "either one at a time". Confirmation in large batches would be more useful, but really this ought to be good enough that it does not need any confirmation at all.

"or in large batches"

Meaning that you would get a grid view where you can quickly see many photos at once and select a large range of images to confirm them all at once.  Very much like PhotoSweeper does when presenting duplicate photo matches.

Confirming the suggestion one at a time would mean you now need to confirm thousands of photos. That is hardly more convenient than rotating them. This would only work well if the AI is good enough to trust it and let it do its work, so you will only have to correct one of its few mistakes every now and then.