Venus Laowa 15mm f/4.5 Zero-D Shift W-Dreamer FFS Review
Months after it came out, there is nothing really out there on this lens except for some stock factory pictures and writeups that are predominantly plagiarisms of promotional materials, “why don’t you get to the point” videos, and vapid clickbait reposts of said videos. There are a couple of decent reviews, but I don’t feel like they were really pushing the lens.
So in the Machine Planet tradition of going off half-cocked, I will give you the dirt on this after spending a day shooting the Nikon F version of this in -3º C weather with a Leica Monochrom Typ 246. No need to start simple, or even with the camera body on which this lens (ostensibly) was intended to mount.
A Typ 246 is an all-monochrome, FX, 24mp Leica mirrorless body that can shoot to 50,000 ISO without looking even as grainy as Tri-X. It has a short flange distance, which means that virtually any SLR lens can be adapted to it. It has pattern, off-the-sensor metering, so there is no messing around with exposure compensation or trying to figure out why shift lenses underexpose on Nikon F100s and overexpose somewhat on the F4 (yes, this is true). It also has an inbuilt 2-axis level that you can see in its EVF, a welcome aid when it is cold outside. These features mean that you can use a shift lens handheld. This lens is a ~22mm equivalent on APS-C (DX) nd I believe a ~30mm equivalent on Micro Four-Thirds. This probably is not a lens for MFT, since it is absolutely massive on any MFT body. In fact, it seems really big for a Sony Alpha body…
The physical plant
The first thing you ask yourself about this lens is, “how could a lens out of China possibly cost $1,199?” But this is a shallow (if not also culturally chauvinistic) observation. Your iPhone is made in China, and there is nothing wrong with its lenses. Or apparently, you iPhone’s price. Venus is something of a newcomer in the camera lens market, and it uses the designator “Laowa,” which is a reference to frogs in a well (not kidding… check out the Facebook page). The idea, they say, is to look up at the sky and keep dreaming. That, of course, is possible where the cost of manufacturing a zillion-element, double-aspherical lens is relatively low. The front ring reads “FF S 15mm F4.5 W-Dreamer No. xxxx.” FFS of course stands for “Full-Frame Shift.”
The 15/4.5 lens is available in a variety of mounts. Word to the wise: get the Nikon F or Canon EF version. Nikon has the longest flange-to-focal distance at 46.5mm, meaning that it has the shortest rear barrel, meaning the maximum compatibility with mount adapters (with simple adapters you can go from Nikon to any mirrorless camera, including Fuji GFX). Canon EF is a close second at 44mm. If you have an existing Canon or Nikon system, just take your pick. Your worst choice is buying this lens in a mirrorless version (Canon RF, Nikon Z, or Sony FE), since you will end up locked into one platform exclusively. Remember that this lens has no electronics or couplings, so adapting it is just a matter of tubes.
The lens comes packed in a very workmanlike white box, just like $50 Neewer wide-aperture lenses for Sony E cameras. This is a mild surprise, but nobody maintains an interest in packaging for very long after a lens comes in. Nikon lenses, after all, come in pulpboard packaging that strongly resembles the egg cartons your kids might give to their hamsters as chew toys. The instructions end with the wisdom, “New Idea. New Fun.” And that is very on-point: for most people, photography is about fun.
The lens is a monster, and it’s not lightweight. It feels at home one something at least the size of a Nikon F4 (and balances well on one, btw). On an M camera, you need to employ the Leica Multifunction Grip (or something similar) to effectively hold onto the camera (this combo can still break your wrist…). Weight as ready-to-mount on a Leica is 740g. For comparison, a Summilux 75 (the original gangster heavyweight for M bodies) is 634g. An 18/3.5 Zeiss ZM Distagon is 351g.
The front element is bulbous. And you must remember that you cannot simply set the camera nose down, since (1) the glass sticks out, (2) there is no filter protecting it from damage from the surface the lens rests on, and (3) this is a really expensive lens. This is also a lens whose lens cap you cannot, must not, ever, lose. It is solid, pretty, bayonets on, and probably can’t be replaced. It is not clear why – if you can mount a 100mm filter holder to the front of this lens – that such a holder is not simply built into the lens – if for no other reason than protecting the front element.
I mounted mine with a Novoflex LEM/NIK adapter, which is pretty much the only dimensionally accurate anything-to-M adapter. Proper registration is a big deal because a 15mm lens cell has very little travel from zero to infinity.
The Novoflex’s stepped interior suggests a place to stick a filter — since the lens has no front filter threads — but for reasons discussed below, this is not a big deal. And in the back of the lens, it’s gel filters – or nothing.
First, this lens is easy to handle wearing gloves. Which, given the temperature yesterday, was fortunate.
This is a little bit different from a traditional PC lens, on which turning a knob would make the shift. The Venus has a third lens ring – behind the focus (front) and aperture (middle). This is different from a Nikon PC lens, for example, where the aperture is front and focus is rear.
The shift ring cams the lens back and forth along the direction of shift, 11mm in either direction. You would think this would interfere with focusing or using the aperture ring, but in reality, it’s likely the only ring you would be moving on a shot-to-shot basis. This lens has such staggering depth of field that you will put this roughly on ∞ and forget about the rest, and you will probably turn it to f/8 and leave it there. Shift is locked with a knob that looks like the knob Nikon uses to shift the lens.
There is a small tab that locks the rotation of the shift mechanism, which can be set to 0 for horizontal pictures, 90 or 270 for verticals, and 180 if you are strange. It moves in 15-degree increments. A 28/3.5 PC-Nikkor does not have a lock, which occasionally can make things exciting if you start framing and realize that your shift is now 45 degrees from vertical (or horizontal).
The aperture ring has light clicks and is logrithmic (unfortunately) – each stop at the wide end is the roughly the same amount of movement, but things do bunch up at f/11, f/16, and f/22. It’s puzzling in this price range.
The focus ring has a short throw, infinity to 1m being about 1cm of travel. Set it and forget it. If you’re looking at pictures on the net and wondering why the focusing scale makes it look like the lens focuses “past” infinity, it’s a mystery.
- At the hard stop and no shift, the lens is indeed focused at infinity. But the scale is off.
- At the hard stop and shifted, the focus is still correct at infinity.
I verified optical focus at the stop both on a Nikon F4 with an adapted red-dot R screen (grid/split prism/f<3.5), the Nikon F4’s phase-detection AF sensor, and with the Leica.
To understand the strangeness of the Venus focusing ring, consider that in an old-school, manual focus lens, you typically have three things in synch for “infinity.”
i. The lens is at its physical stop, meaning you can’t turn the focusing ring to make the optical unit get closer to the imaging surface. This is normally an inbuilt limitation. It is not typically a critical tolerance on a lens due to the two adjustments below.
ii. The lens is optically focused at infinity, meaning that an infinitely distant object is in-focus on the imaging surface. This is usually a matter of shimming the optical unit or in some lenses or using a similar adjustment for forward/backward position of the optical unit.
iii. The focusing scale reads ∞. In the old days, this was simply a matter of undoing three setscrews, lining up the ∞ mark with the focus pointer, and then tightening the screws. If you are a super-precise operator like Leica, your lens stop/focusing ring/scale are made as one piece and so precisely that no separately applied focusing scale is required.
When a manufacturer of modern autofocus lenses (or even high-performance manual telephotos) is confronted with design constraints, it generally omits the relationship (i), the physical stop, and (iii) the infinity mark. It will do this on telephotos (like the 300/4.5 ED-IF Nikkor) because heat-related expansion might otherwise prevent a telephoto from actually focusing on a distant object. With AF lenses, hard stops are not the best for the fallback “hunting” mode — and with the user relying heavily on AF anyway, there is no need to inject another thing to check in QC. By the way, on a lot of AF lenses, the focus scale is basically just taped on – eliminating the setscrews.
Cheaper lenses, like the Neewer I-got-drunk-and-bought-it-on-Ebay specials, don’t really couple any of these things precisely. The stop is set so that you can optically focus past infinity and yet when the lens is optically focused at infinity, the focus scale might read somewhere between 10m and the left lobe of ∞.
For reasons that are frankly baffling, Venus uses a different idea entirely, which is to match the collimation and the stop – the hard part – and yet to omit matching the focusing scale. This provides no ascertainable benefit unless the focusing ring is not just a ring but an integral part of the focusing mechanism. I don’t see any setscrews, so maybe this is the explanation. And really, something in this price range should have things line up, even if it means adding one more cosmetic part to make the focusing scale adjustable.
On the surface, this design choice is frustrating to perfectionists and degrades the value of the focusing scale. That said, in 99% of pictures you take with this lens, you’re going to set it to the hard stop and get more than sufficient depth of field for close objects just by virtue of stopping the lens down.
If you are reading this, Venus, the focus scale design needs to be fixed.
There was nothing remarkable about shooting this lens, which is a good thing. As long as you realize it has no electronic connections or mechanical control linkages to the camera it… works like any Leica lens.
They used to advise that PC lenses had to be used on tripods. That was true when (1) cameras did not have inbuilt electronic levels and most did not have grid focusing screens, (2) viewfinders blacked out at small apertures and with shift, and (3) through the lens meters freaked out at the vignetting.
None of those conditions exist with mirrorless cameras, where viewing is off the sensor, focusing is by peaking, and signal amplification makes it possible to frame a picture even closed down to f/16. On the Leica Monochrom, for example, it is very easy to use this lens – no different from using any other with the EVF. The M typ 240 series cameras have inbuilt levels that are visible through the EVF; the later M10s do too. A visible level is absolutely essential if you are going to shoot this (or any shift lens) handled.
Speaking of the sky, the sweep of this lens, its vignetting, and its self-polarization mean that in many pictures, the sky will be darker than you expect. Most people will not mind. I suppose you could mount a 100mm filter to the front or a gel in the back, but this is highly dependent on what you are trying to do, your tolerance for the expense, and the light response of your camera.
One thing you begin to realize is that if you switch from a 28mm PC lens to a monster 15mm PC lens, you go from shifting exclusively up to avoid converging parallels – to also shifting down to cut down on excessive sky. You might think of the shift as the “horizon control” adjustment. The challenge is, at the end of the day, that this is still a 15mm lens with a super-wide field. Unlike a 28 or 35, you need to think about both the top and the bottom of the picture.
One other thing you will see in a couple of the pictures in the article is that a slight forward tilt of the camera can make things look slightly bigger than they should at the top. This is user error and the unintended opposite of converging parallels.
With wide lenses, you need to watch 3 axes of alignment – left/right tilt, front/back tilt, and critically, parallelism to the subject. This last point can be a major irritation with this lens since cameras don’t typically have live indications of whether you are square with the subject.
Note: WordPress scales pictures down and not in a flattering way; if you want pixel-level sharpness comparisons to other lenses, there are other reviews out there that do that.
The jury is still out here – at least until I get a sunny day and hook this up to an A7r ii, which is more representative of cameras most people would use with this lens. But the foreman is asking some of the right questions for the verdict we want. Field curvature is also something that needs more exploration. As it stands, though, the lens seems to be more than sharp enough for its intended purpose.
All wide-angle lenses have degradation toward the edges of the frame. Many cameras don’t have the resolution to make it obvious, but this is a well-known reality. Shift lenses have a bigger image circle, which gives them comparable performance (not stellar, but comparable) performance to normal lenses over a wide area. They are “average,” but average in the sense that they are reasonably sharp over the whole frame, not super-sharp in the center and falling apart at the edges.
Put another way, a shift lens for 35mm is essentially a medium-format lens. Medium format lenses do not have the highest resolution – because they don’t have to. But they do deliver their performance over a wider field. But by shifting the lens, you are bringing lower-performing edges of the field into the 35mm frame.
But… you protest… my AIA book has all of these perfect architectural pictures of xyz buildings.
No, it does not. First, they are tiny, and that with the halftone screens, they give off an impression of being much sharper than they are in reality. Second, if you look at an original print closeup – pixel-peeping on prints was never normal when people made prints – you’ll see that the pointy top of that building is fuzzy because someone used a 4×5 or 8×10 camera and shifted it to accommodate the tall object in the picture. But seeing it in a gallery or an exhibit, you would (i) be standing back from it and (2) paying attention to the center of the frame, which is where most pictorial interest is. That pointy top is in your peripheral, not central, vision. The central part still has adequate performance for the purpose.
For this reason, the sharpness of a shift lens can only really be understood in terms of shift lenses or shifted medium- or large-format lenses: if you leave a little sky above a tall building, you don’t have to confront so much the inevitable performance falloff in those last couple of mm of the frame. All shift lenses have this issue, and it goes both to illumination and sharpness. Go to maximum shift on anything, and you can expect image degradation at a pixel-peeping level in the top third of the image.
So what? This is the same thing that people with shift-capable cameras have faced since… forever.
And why do shift lenses exist? The answer is pretty simple; it’s easier to get to a good result than many types of post-correction. If you plan to do post-correction, you have to use a much wider lens than you normally would, you have to crop (because tilting an image in post makes the field a trapezoid that must be rectified), and you have to have an accurate measurement of the scale of the original object. On this last point, if you don’t know the XY proportions of a building’s windows, perspective-correcting it in post-processing will result in awkward proportions. So if you have a 42mp image that needs serious correction to make a tall building upright and correctly proportioned, you may end up with less than 20mp of image by the time the process is over. And since tilting magnifies the top edge of the image, you are magnifying lens aberrations in the process.
Post-correcting does have one advantage, though, which is that you can use a lens that performs highly across the frame. I do it a bit with the Fujinon SWS 50mm f/5.6 on a 6×9 camera: when you are working with a wide lens, from a 96mp scan, you have plenty of resolution to burn in fixing one degree of inclination. This is not so much the case with a 35mm lens on a 35mm body.
As of this writing, Leica just announced in-camera tilt correction for its 40mp M series cameras. This is an idea long overdue, since the camera knows what lens is mounted (or can be told) and the inclinations at the time of the shot. s.
You don’t escape post-processing with shift lenses, particularly when you have to fix skew between the image plane and the subject (rotation around the vertical axis of your body). PC lenses also have distortion to contend with, and simple spherical distortion sometimes seems less simple when the “sphere” is in the top half of the frame. But the corrective action is far, far milder.
The complication with digital and shift lenses is diffraction. With a shift lens, you need a small aperture to even out the illumination and sharpness, but that small aperture cannot be smaller than the diffraction limit without degrading sharpness overall. That’s f/11 on a Leica M246 and roughly f/8 on a Leica M10 or a Sony A7r II or A7r III series camera.
A further complication with all shift optics is dust. Small apertures, smaller than f/5.6, tend to show dust on the sensor. Shift optics have at least one extra place for dust to get into the camera body (the interface where the shift mechanism slides the two halves of the frame).
Sharpness seems to peak at around f/8 on the Venus, which is not surprising. The sharpness itself is good as well as consistent until the very margins of a shifted frame; I did not need to turn on sharpening on Lightroom. As with all lenses, apparent sharpness is higher on closer objects – because their details are bigger in the image, pixel-level aberrations are not as apparent.
The goal is “Zero-D(istortion).” The lens gets close – and better than most SLR lenses in this range, and certainly better than a lot of SLR PC lenses – but not completely distortion-free. Unshifted, it looks like a relatively mild +2 in Lightroom (the shot above is uncorrected except for slight horizon tilt). Shifted might be a little tougher to correct, but you can either create a preset for Lightroom or use some of the more advanced tools in Photoshop.
Yes. It has flare when light hits it wrong. Check out the picture above. Sometimes it works. Sometimes it is an irritation. Luckily, it does not seem to happen very often,
There is a real tendency to abuse superwides in photography today, usually to disastrous effect due to the inability of photographers to properly compose pictures. Companies like Cosina/Voigtlander have fed into this, as has Venus, with about a dozen high-performing superwide lenses that would have seemed impossible just a few years ago. “Wide” used to mean 35mm; now “wide” tends to mean 24mm, and “superwide” is below 15mm. The Venus has all of the vices of a wide-angle lens, notably posing the question, “what do I do with all this foreground?”
By the same token, shift lenses are very specialized tools. Old-school shift lenses were the least automated lenses in their respective SLR lines; new ones are marginally more automated (mainly having automatic apertures), but they are staggeringly expensive.
The Venus somehow manages to combine the best and worst of all of this. You cannot argue with the optical performance as a shift lens, but the lack of automation (and frankly, ease of use) makes it just as miserable to use on a native SLR body as any old-school shift lens was. You’ll note where people complain about this lens in reviews, that’s what they complain about. I’m not sure that merits much sympathy; you know what you signed up for. What makes the Venus more fun is that it connects to mirrorless bodies that, by virtue of their EVFs, remove a lot of the irritation that would occur using the lens on a traditional SLR body.
Whether you will always be shooting 30-story buildings from 200m away is a matter of your own predilections, and that might be the deciding factor. Unless you are really good with wide-angle shots – or are a real-estate photographer in Hong Kong, you may not have a very solid (or at least somewhat economically viable) use case. But in reality, the market is not driven by professional needs. If it were, the only things that would ever be sold would be full-frame DSLRs, superfast 50mms, and the “most unique wedding I’ve ever seen” presets package for Lightroom.
Pros: solid build quality, clever shift mechanism, wide angle of view,* reasonably low distortion, actually collimated correctly for its native mount.
Cons: non-linear aperture control,** odd (incorrect?) focus scale calibration,** facilitation of compositional errors you never previously imagined possible,* bulbous front element, no inbuilt filter capability, and a lens cap that only mounts one way.
*Qualities that would be inherent to any lens this wide with shift capability.
**Qualities that do not typically belong on lenses in this price range.
Adobe face recognition: beat the system?
The Kobayashi Maru test is not a test of character unless you see the world in terms of “go down in dignity with the [star]ship” or “be a coward.” Or whatever Nick Meyer thought the outcomes would be. Captain Kirk won the test by not accepting a binary decision tree. This is exactly how you should approach any problem that looks like it is unwinnable. Rewrite the simulation. Use a screwdriver as a chisel.
One of the ways you can do this is to ignore the process as presented completely, decide your goal state, and then selectively use whatever is available to get there. Face recognition is exactly such an exercise. Adobe would have you select one of two suboptimal tools (Lightroom Classic or Lightroom) and have you build out the recognition process and leave it in the platform where it started.
Not believing in the no-win scen-ah-ri-o (sorry, Shatner), I started with the first principle:
What is the purpose of face recognition in photos?
This is actually a really good question. The way the process on Lightroom proceeds (either version), you think the purpose is to name every person in every photo and know what precise face goes with every name. This view assumes that you are a photojournalist who needs to capture stuff. You will go bat crazy trying to achieve this goal if your back catalog is hundreds of thousands of pictures and you use Lightroom Classic (“Classic”) as your primary tool.
Let’s face it – you are (at least this year) a work-at-home salary man, not Gene Capa. The real utility of face recognition is to pull up all pictures of someone you actually care about. You need it for a funeral. For a birthday party. For blackmail.
That does not actually require you to identify precise faces, just to know that one face in the picture is the one you want. You already know this person’s name and how the person looks. And even if you didn’t remember, a collection of pictures of that person – no matter who else was in or out of the shot – would have one subject in common. You would know within a few pictures who John Smith was.
Taking this view, a face identification is just another keyword.
It’s not even 100% clear that you would ever need it done in advance, on spec, or before you had a real need to use it.
What do we know about face recognition in LrC vs LR?
Our statement of problem: 250,000 images of various people, some memorable and some not. I want to get to being able to pull up all pictures of John, Joe, Jane, or Bill. And I want this capability to last longer than my patience with Lightroom cloud. I want to be able to ditch Lightroom, even Classic, one day and change platforms without losing my work.
When you are figuring out a work flow, or trying to, it’s helpful to consider what your tools can and cannot do; hence, with Classic and Cloud, start breaking down the capabilities.
- Both recognize faces with rudimentary training.
- Cloud is much faster than Classic and tends to have fewer false hits (due to Sensei)
- Both can do face recognition within a subset of photos.
- Classic can an apply keywords to images that Cloud can see/
- Cloud cannot create keywords that Classic can see.
- LrC has better keyword capabilities, period.
- You can make an album in Cloud and have it (and its contents) show up as a collection in Classic.
- You can put things in one of these items in either program and have it show up in the other.
Do these suggest anything? No? Let’s step through.
Let’s talk about some preliminaries that no one ever seems to address.
Order of operations. If you are starting from zero, you should identify faces in the import every time you import something. Not only are names of near-strangers fresher in your mind, it also prevents the kind of effort we are about to explore.
What’s my name? You must have a naming convention and a normalized list of names. It doesn’t matter whether you pick someone’s nickname, real name, married name, whatever. Whatever you decide for a person must be treated consistently. Is my name Machine Planet? Planet Machine? PlanetMachine? This has implications for Classic, where you can’t simply type a two-word name (Bill Jones) into the text search box without getting everyone named Bill and everyone named Jones. For that you might want to concatenate both names together (unless you want to use keywords in the hierarchical filters). In Cloud, the program can sort by first and last name, so there is value in leaving these separate.
Stay in the moment. Although you might be tempted to run learning against every single picture you have at once, this leads to a congested Faces view (or People view), slow recalculation on Classic and a lot of frustration. Do a day or a week at a time. Or an event. This will give you far fewer faces from which to choose, and fewer faces to identify. Likewise, if there is a large group picture in the set, focus your effort on tagging everyone in it. This will set up any additional Identified People in Classic and will kickstart Cloud.
Who’s your friend? You next need to decide who is worth doing a lot of work to ID. You are not going to do iterative identification (especially on Classic) with people you don’t care about. Leave their faces unidentified. Or better yet, delete the face zones. This is a very small amount of effort in a 200-shot session or a 36-shot roll of scanned pictures.
Start in Cloud. This part is not intuitive at all. Go ahead and sync (do not migrate!) all your pictures to Lightroom mobile. This consumes no storage space on the Adobe plan. If there are a lot that have no humans, use a program like Excire Search to detect pictures with at least one face pointed at the camera. This is a reasonable cut, since there are few pictures you would bother tagging that have one face, solely in profile.
The synch process will take forever. I don’t think there is a lot of point in preserving the Classic folder structure when you do this; I would just make a collection like “Color 2000-2010” in the Classic synched collections and dump your targets into that (n.b. a collection in Classic is just an alias to your pictures; making a collection does not change the folder arrangement on your computer). We are only using Cloud for face recognition; its foldering is too rudimentary and inflexible to be useful – although right-clicking in Classic to make folders (or groups of folders) into synched folders will let you adopt the Classic organization in Cloud, albeit flattened, without re-synching. Again, not very useful. Also, for reasons described further on, you want to have a relatively clean folder panel in Cloud because you will be making some albums, and you don’t need extra clutter.
Ok. Let the synch run its course, or start your identification work on Cloud as it goes. Cloud will start aggregating what it thinks is the same face into face groups, which you then must name. Start naming these according to the convention you chose. I would put the People view to sort by “count,” which naturally puts the most important people at the top (you have the most pictures of them). Let’s say you name one face group “John Smith.”
The process so far is pretty generic. To start crossing things over to Classic, you need to make folders (“albums”) in Cloud. Start with one per important person (“___ John Smith”). Search for that person. Dump the search results into the album. You can always add more later.
Now flip back to Classic. You will see collections under “From Lightroom.” Voilà! One of them is “John Smith.”
Now you can do one of two things.
You can simply make a quick check to make sure there are no pictures included that obviously are not John Smith. But after you do that, or not, you can mass-keyword everything in that collection “John Smith.” If you named John Smith consistently with any pre-existing Classic face identification of John Smith (i.e., not two different variations of the name), your searches will now have the benefit of both tools. Save those keywords down to the JPG/TIFF files (Control-S/Command-S) or XML files (same), and you will forever have them, regardless of whether you leave the Adobe infrastructure. In fact, many computer-level file indexes can find JPGs and TIFFs by embedded keywords (which the index sees as text).
Congratulations. Now you’ve highjacked Sensei into doing the dirty work on Classic.
With a small but not overwhelming amount of creativity, you could use a technique like this to cross-check your past Classic calls.
STOP HERE AND GO TO “CALIBRATING YOUR EFFORTS” UNLESS YOU ARE A MASOCHIST
Second, if you’ve missed your OCD meds, you can also use the results of this to inform your Classic face-recognition process.
a. Select this “From Lightroom–>John Smith” collection and flip to Faces view in Classic.
You are now seeing all “Named People” and all “Unnamed People.” Unnamed people are shown by who Photoshop thinks they are most likely to be. You can sort Unnamed people in various ways, but however you do it, you want to get John Smith?s in a contiguous section where you can then confirm or X out. By going into this in the From Lightroom–>John Smith collection, you are not waiting for recalculations against every photo you have – just the ones that Sensei thought should have John Smith.
So the cool trick is this: if you see 106 pictures in From Lightroom–>John Smith, then you know you are probably going to be done when you have 106 confirmed pictures of him. Or done enough. John can only appear in a picture once. There will be a margin of error due to how closely Classic can approximate Sensei, but you can get to about 90% of the Sensei results without a lot of trouble. This is a bit better than Classic on its own, where more pictures of John Smith at an earlier age might be really buried down in the near matches. Further, Classic is something of a black hole for similar pictures because unlike Cloud with Sensei, there is no minimum required similarity score to be a suspected match.
b. You can, of course, drill down on John Smith as a Named Person. You don’t have much control over how “Similar” pictures are ordered (I believe it is degree of match for the face), but here, you can confirm a much more concentrated set (after you decide how to deal with the “fliers” who are not John Smith).
One other technique I have developed while in the “confirming” stage is that it may be easier to confirm en masse (even if some are wrong) down to the point where the “not John Smiths” are about a third of the results in a row of Similar faces. A small number of “fliers” can be removed by going up to the Confirmed pictures, selecting them, and hitting delete. Trying to select huge swaths of unconfirmed faces in Similar and then unselecting scattered fliers tends to really slow things down. As in Classic really slows down as it tries to read metadata from everything you selected.
Incidentally including and then manually removing a few fliers from Confirmed does not seem to affect accuracy (because every recomputation of similarity is on the then-current set of Confirmed faces – changing that set changes the computation). If you have 99 pictures that are right and one that is wrong, it won’t even change the accuracy appreciably. If in Confirmed, you have 995 pictures that are John Smith and 5 that are not, again, the bigger set of correct ones will predominate future calculations.
Next, at some point, especially with siblings, Classic is going to reach a point where Jane Smith (John’s Sister) is going to show up as a lot of the “Similars” with John Smith. When this happens, go back to Faces (top level, always within From Lightroom–>John Smith), click on her, and confirm a bunch of her pictures. When you go back to Named Person John Smith, a lot of the noise will be gone, and hopefully more John Smiths will be visible in a concentrated set you can bulk-confirm.
Crossing back (optional)
I did write “iterate,” right? You might want to keep your Cloud face IDs as complete as possible, since there is not 100% correspondence between results from the methods used by the two platforms. This is relevant if you have already trained Classic on John Smith.
- In Classic, note the count in your From Lightroom–>John Smith collection. Say it’s 106 pictures.
- Do a search from your Classic Library for all pictures of John Smith. If you used a space in the name, add Keywords to the field chooser menus (via preferences) and select that line.
- Drag all of those results to From Lightroom–>John Smith.
- Flip to Cloud. They are now in that “John Smith” album. Or they will be when it synchs.
- Select all the pictures in the “John Smith” album.
- Hit Control-K (or Command-K) to bring up keywords and detected/recognized faces in the “John Smith” album.
- Now name any faces that are blanks – but should be John Smith.
- Now from the All Pictures view, search for John Smith and drag all his pictures to the John Smith album.
- In Classic, check your count. If it’s say 128 pictures, now, that means that Cloud took your examples and found more John Smiths. And now they are ID’ed in Classic as well.
- Switch to Faces and confirm the 22 additional faces as John Smith. Now both systems have identical results.
Calibrating your efforts
For searches for random people, Cloud is still the best because it requires very little training. That said, for randos, you are using a tool that does not give you any permanent results. That’s probably ok for people who you don’t really care about. Or if you plan to be on Cloud forever.
For close friends and family, you may just run the “Crossing Over” exercise. I would do it in groups: do a bunch of albums on Cloud (say seven people), then do a bunch of naming on Classic (their collections), etc.
If you are really a neat-freak or compulsive, you could use the “Crossing back” step. But Sensei is reasonably good at what it does, so the marginal effect of adding Classic results to Sensei may not be much. If you have Excire, you might use it to find pictures that look like a picture of John Smith, which will give you a third means of concurrence.
The thing to remember about face recognition is that it is miraculous but also imperfect. It has to detect a face and then it has to identify a face. It doesn’t see how you see. Efficiency works at cross-purposes to accuracy.
But it is still vastly better than trying all of this on your own.
Face-off: Apple vs. Adobe face recognition
So here is a question: what’s the best way to catalogue and tag your pictures? Is it Lightroom Classic? Lightroom Cloud? Is it Apple Photos? Is it something else? Maybe it’s a lot of things. If you are a high-volume imaging-type person, you’ve probably wondered how to deal with things like tagging people. The most macabre application, of course, is the funeral collage. But say you have tens of thousands of pictures of family members and want to print a chronological photo album. Then what? Face recognition features in software may be your best bet. From a time standpoint, they may be your only choice. The problem is that different software has different competencies.
Something like Photos is designed to group pictures, more or less automatically, around people, events, dates, or geography. Think of it as your iPhone application on steroids. Photos is not big on user control. It is not even engineered to do anything with folders except display them if that’s how photos were imported.
Face recognition in Photos is incremental and behind the scenes: it only finds faces when you are not actively using the program, and over time, it batches up groups of pictures which you confirm or deny as a named person in your Faces collection. To establish your Faces collection, you have to put names on faces in a frame where faces have been detected. This tends to mean that face recognition proceeds by which faces the user thinks are most important. As it should be.
Unlike Lightroom, Photos does not presume that detected faces are unique. It applies a threshold such that if it detects Faces A, B, C, and D, and they are close enough, they are treated as the same (unnamed) person. As such, naming one person can have the unintended effect of tagging a bunch of false matches. Either way, you can error correct by right-clicking the ones you see that are wrong.
My assessment of Photos is that it is not suitable as a face-recognition tool if you have hundreds of thousands of images, for several reasons:
- Its catalogs are gigantic, even if you use “referenced” images. Photos loves it some big previews, no matter what you do. For scale, my referenced Photos library is 250gb where my entire Lightroom Classic library folder is 40gb (both excluding original image files – so Photos sucks up 6x the space).
- The face recognition process appears to be mostly (if not completely local), it runs in spare processor cycles, and in my experience, can cause kernel panic. Hand-in-hand with this is the fact that you can never actually turn Photos off. It’s part of MacOS.
- There does not appear to be any indication that Photos actually writes metadata to files. So when you move to a new application, you’re starting from zero.
- You can’t really use it in conjunction with a grown-up asset management system like Lightroom.
Photos is, however, good for generating hilariously off-base collections of photos (memories) with weird auto-generated titles (“Celebrate good times” with a crying baby as the cover photo). Or collections based on the date a bunch of pictures taken over decades were scanned (such as my 42,600 pictures apparently taken on December 12, 2008). I actually have no idea where these are generated. But they are funny.
I’m sure Photos is really good for those funeral collages, though.
Lightroom Classic (LrC)
Something like Lightroom Classic (LrC) is designed around manipulating, filtering, and outputting large numbers of pictures at once. This is, indeed, the killer app for handling large volumes of photos, and becomes a single interface for everything. It’s OK, but not great, for face recognition.
To put it mildly, LrC’s face-recognition is processor- and disk-intensive. The best way to use it is to use it on a few hundred photos at a time so that your identifications don’t swamp everything in your collection in a recalculation. LrC is good at showing you different faces all at once, as single images, so you can get cracking on identifying as many new “people” as you have patience for in one sitting.
The top level of the Faces module shows you (i) “Named People” and (ii) “Unnamed People.” You need to name at least one “Unnamed” person to start. After a while, the system will try to start putting names on “Unnamed” people. If you have a Named person named “John Doe” and are presented with an image that is “John Doe?” you can click the check box to confirm it and the X box to remove the suggestion (clicking again removes the detected face zone, such as if the system mistook a 1970s stereo for someone’s face).
Once you have done that, you can drill down on a “Named” person to see what pictures are “Confirmed” and what pictures are “Similar.” Again, to move from Similar to Confirmed requires an affirmative call. Here, you only get a check box. There is no “Not John Doe” option, which means that every possible match is shown, ranked in what LrC thinks is similarity. This is actually problematic because as you confirm more pictures, the number of “Similar” pictures rises exponentially. This puts a huge computational drag on things.
Wherever it happens, confirmation of a face’s identity is an affirmative process that is repeated for each picture (you can select several). This prevents false IDs based on grouping disparate real people into one “face,” but it also makes tagging excruciatingly repetitive. And slow. Highlighting faces to group-confirm or identify can have the “highlight” lagging far after your click. And God help you if you click six pictures and then try to type a name into one to rename all six. It works about half the time. The other half, it auto-completes with a totally unintended name. If you accidentally confirm the wrong face for a given name, you can highlight the errant thumbnail and hit Delete (this is not well documented).
Critically, the top level of the Faces module (where you see all named people as thumbnails) is the only place where the system puts a “most likely name” on unnamed people. Otherwise, looking at any particular “Named Person,” the same person – Bob – might show up as a similar for John Doe. And when you switch to Richard Roe, Bob will show up as a “similar” for him as well. This is part of the reason why people for whom you have 10 actual pictures always show up with 20,000 “similars.”
A big advantage of LrC over other solutions is that you can see and tag faces within specific folders, collections, or filmstrips. This lets you make context-sensitive decisions about who is who. For example, I am pretty sure that my kids did not exist in the 1970s. Or I might know that only 6 people are represented on a single roll of film that constitutes a folder in my library.
When a name is confirmed on a picture, that name is written as a keyword to the metadata in the library. It appears that XMP files (if you chose that option for RAW files) are written with the actual coordinates of faces in the picture, which allows some recovery if you have to rebuild a library from scratch. The important thing is that a picture is keyworded with the right names. Face zones are nice but not quite as critical in the long run because in reality, you only really care whether a picture contains John Doe or Richard Roe, not which one is which in a picture of both.
Always save your metadata to files if working with TIFFs/JPEGs/scans (Command+S) or “always write XMP” with RAW camera files. This helps keep your options open if you want to get divorced from Adobe. Or if your Lightroom library goes wheels-up and you have to rebuild from zero. There is no explanation for why this program just doesn’t write an XMP for every file. It would make things easier.
Lightroom [CC or “cloud”]
What a hot mess. The only thing that really works about Lr CC is face recognition. The rest of it is a flashy, underpowered toy that despite being “cloud” based can still consume massive amounts of hard drive space and processing power. If your photos are in the Adobe cloud, or synched from LrC, the program works with smart previews.
Adobe’s Sensei technology is a frighteningly good face-recognition system. In the People view (mutually exclusive with the Folders view), it takes all of your photos and groups them according to what it thinks is the same face (like Apple Photos). Put a name on that face, and it might ask you if this other stack over here is the same face. It is extremely fast (because it runs in the cloud). Sensei can also identify objects, and to some degree, places in photos. Naturally, the most important people in your life have the highest counts, and you can sort unnamed faces by count and work your way down. Things break down when 400 people have 15 pictures apiece, though…
The system, though, has some amazing limitations that are pretty clearly engineered in by a company that is trying to move everyone to its walled garden. Two of these four bear directly on the issue of why a hard drive – and keeping your own metadata local – is your ladder out of that walled garden.
First, metadata transfers to Lr are one-way. The program can absorb keywords applied in LrC, but not recognized faces/zones, and nothing you input in Lr can ever rain down on LrC. There is no programming-related reason that prevents metadata from flowing the other way, aside from intentionally engineering this out of being possible — so that you are eventually forced to store all your stuff in Adobe’s per-month-subscription storage space. Because paying a monthly to use programs that aren’t really being updated – like LrC – was not bad enough.
Second, you cannot force face recognition on arbitrary subsets of your library, at least very efficiently or intuitively. If you came at this program assuming that it would be like LrC, you would conclude that there is no way to do this. Instead, you have to select a group of pictures and hit Command/Control-K (for “keyword” – how intuitive…) to see the faces present in the picture or group. Lr then shows you the single picture with the face boxes – and the collection of faces in the picture on the right panel. This is great – but why is it so hard to find? You also get the impression that when you do this, the face boxes are generated on the fly. But the critical defect here is that the “named faces” that are shown as thumbnails are even smaller than the other face thumbnails in Lr.
Third, when asked to “consolidate” two faces, there is no way to flip between the two collections. This is an oversight – you are not asked to name a person based on one photo, but for some reason you are asked to make a consolidation decision that could have catastrophic consequences — based on a single fuzzy thumbnail. If in doubt, sit it out.
Finally, you can’t push face recognition data back down to LrC. So if you use LrC, you basically end up with completely separate face-recognition data sets based on the same photos. This is a big-time fail.
Well, in terms of applications you can access for a Mac right now, the options are ok – but not great. Stay tuned for Part 2, in which we look at a way to leverage LrC and LR CC against each other to speed things up.