I recently tested several so-called AI “Culling” tools for photographers, i.e. software that attempts to help you more quickly pick the best images from a larger set of photos. The most common use cases are for professionals who do events, especially weddings or anything else that has a lot of faces in it. Typical selection criterion are fairly basic at the moment, primarily blurry vs. sharp, eyes open, eye gaze (toward or away from the camera), faces visible vs. obscured, etc.
For my purposes this is only of limited value as I mostly do landscape or other non-people photography. But I’m still hopeful these tools will meet my needs more and more. In particular I’d like to see AI models for good landscape photography that can then help recommend potentially good landscape photos in a set of mine. I’m also very hopeful that I’ll soon be able to train some of these AI models on my own prior photo selections (which are all in the metadata of photos from my use of Lightroom for “catalog” management). This would hopefully let the AI recommend photos that are not just “likely to do well on Instagram” but also and more particularly represent my specific taste, being closer to what I want to represent to and show the world.
So with all that in mind, here are my quick thoughts on several of the existing tools in the space, with their respective functionality as-of December, 2022 (things are changing rapidly across the AI toolspace, so some of the below info may quickly be out of date!).
AfterShoot
This one is my current favorite of the options I have tested.
- OK, right away this impresses for 2 reasons. First, it loads metadata quickly but does not immediately start AI analysis. Second it loads and displays my color ratings from Lightroom, so I can tell if it culls what I already picked! That’s helpful, at least for this stage (though not sure if it has future utility, but I like the feature overall).
- AfterShoot is clearly using the CPU, actually doing something, unlike FilterPixel (see below). You can even adjust how much system resources it uses, nice!
- The Highlights feature is supposed to pick images it thinks might do well on social media, but it seemed pretty poor to me. I mean some looked good, sure, but just as many looked bad. Oh well.
- So out of 700 it picked 331 total. Not bad for a first pass. But bad news: somehow it only picked two of the same images I picked previously myself!
How strange. - You are able to zoom in further in review mode and there is some more detail there, but not full resolution. Still it’s better than FilterPixel, and adequate I think.
- Overall more impressed with this as an app vs. FilterPixel, even though it mostly picked totally different pics than me. Maybe I could train it on my own selections? If they don’t offer a way to just check LR existing picks, I guess a semi-manual process that would still be reasonably quick would be to go back through a few recent albums I already processed and then set color/rating for the ones I ended up publishing, then import that to AfterShoot and use it as the cull selection. Maybe that would train it? Wonder how much it can be trained though.
- AfterShoot support later confirmed that A: my workaround could be effective in the short-term and B: they are working on ways of training their AI model on user’s prior picks, and the app is already supposed to “learn” the style of each person to some degree.
- Woah, I also see now that they have an AfterShoot Edits product in alpha, about to be beta, that will basically auto-edit your photos to your preference based on example past edits. Sweeeeet! My photo hobby might be about to get a lot more fun, hah.
- So hmm, after testing the wedding album, it picked some I had picked, but also had several near-duplicates (or even actual duplicates with different processing). So it seems to still have some issues. If I can train it on my taste it has the best UI/UX and I’d use it, could save time. But that’s unfortunately the only way I could feel comfortable using it at present. Fortunately I can use it to just add metadata (i.e. ratings/colors corresponding to picks/culling) and then that can perhaps guide me in Lightroom, so I can use it as a first pass. But without custom training it is unlikely to be that useful, sadly. AfterShoot Edits may prove to be more useful quicker, in fact. But this is still quite promising and I’ll keep an eye on it.
Update: I have a meeting scheduled with someone from their product team. Will try to report any useful new info after that.
FilterPixel
- First pass analysis of 700 photos on SSD was about 7-10 mins. It stuck at 99% for a while though. Initially some 480 pics out of 700 total were “picked”, so not even cut in half.
- Increased the quality requirements for “focus” and “eye quality” and that got me down to 290 or so, but I can’t really see a major rhyme or reason to why it’s picking. It’s supposed to bundle similar photos for example, which it does if they’re quite similar, but ones that to me are compositionally similar aren’t bundled, and aren’t culled or “picked”.
- Also of note, you can’t view images at full-size and zoom in, it appears to be maybe using the built-in preview images (in the raw file) or something. But this could be OK if I could trust the first pass and then just do remaining comparison and fine-tuning in Lightroom…
- Overall it’s hard to feel like I could trust this, even though this is supposed to be the one that can “learn my style” I think. To be fair these tools are currently aimed at portraits, so might be better to try on a wedding or other people pics set.
- Annoyingly there are no apparent Settings or Preferences, just the main UI sliders and buttons. Hrm.
- OK, now testing it on a wedding set. It had an error at first, didn’t like a PSD file (crashed while it was in the folder). So I moved that and tried again. Still had the error (file was just in a subfolder), and hung on processing again. Tried one more time, got through to 99%, and stopped with little CPU being used, so I just let it sit to see if it finishes, and I’ll test something else in the meantime… No, never finished. Not impressed with stability here.
- Also some annoyingly persistent popup ads/coupons and referral program. They should stay dismissed. Bad first impression here with the stability and popups…
Optyx
- Import process is less slick than AfterShoot, but it does give you a time estimate before processing, which is nice. Even if it is slower than AfterShoot in the end.
- Optyx uses what looks like 2 CPU cores, AfterShoot sometimes used more I think. OK, after initial import, during AI analysis, it uses a good bit more, 3-4 cores, 45-60% CPU. Nice! About 10 mins for analysis, slower than AfterShoot, but not too bad.
- It too shows existing meta data, great! E.g. color red (from Lightroom), etc.
- You can zoom, but not as well as AfterShoot
- Same as AfterShoot, it picked mostly photos I did not. Huh. Worrying? Also more confusing UI, I don’t like it as much as AfterShoot and don’t see any major reason to choose this over it…
Narrative Select
I only tested this briefly as it’s Mac-only and most of my photo processing is currently done on one of my PCs. But I’m open to moving to the Mac or just using it for part of my culling process it here is a notable benefit. Unfortunately the UI for Narrative Select is confusing and awkward to me, and I’m not even sure it auto-culls really. Not that interested in learning it more at the moment, AfterShoot felt much more natural.
Initial Research
Before I tested the above tools, I did a bit of initial, exploratory research of the tool space to see which ones I might want to actually test. Here are my notes on that, for what it’s worth, which include mention of a few other tools I didn’t test (and the reasons why).
- Camera Futura looks kind of older and clunky (UI-wise), has decent features, but does not seem to integrate with Lightroom, it copies or moves images to different folders based on its results, rather than just tagging/metadata
- PostPro Want is a Lightroom plugin, but cloud-based and expensive. Nope.
- AfterShoot claims to adjust to your “style” over time. And does not upload stuff to the cloud. $10/mo, odd pricing for non-cloud. Currently on sale. Will test.
- Narrative Select is non-cloud, but Mac-only. Free version but $15/mo for more than 4 “projects” per month.
- Optyx has extensive meta data customization, Mac+Win, works with LR. Free version or $7/mo. Unclear if cloud-based nor whether trial of Pro is offered, but may be worth testing.
- FilterPixel is also not cloud-based. $100/yr at sale price currently. Somehow seems maybe less good, but possibly worth testing.