Average Abstraction: Seventeen lakes, one haze
Seventeen stock photos of lakes, stacked into one picture by Average Abstraction until none of them is left to see.
Seventeen photos from the app's Lake set, aligned and averaged. A gull, a heron, a bench in the snow and a mountain range all went in. A5 at 300 dpi, 2478 × 1652 px.
What does a lake look like if you average seventeen photographs of different lakes? Like none of them. With a handful of photos an average still reads as a double exposure, each picture half there. With seventeen, every motif is outvoted, and what is left is what they share: a warm band where the far shores were, cool grey sky above, darker water below, and vertical streaks from reeds and trunks.
Below is how this one got made in three decisions, and what the download holds. Open the Lake set and follow along.
How this one got made
The app is three pages, one per step, with Back and Next in the controls to move between them.
Acquire, Cluster, Average: the rest of this post follows the same three pages.
Three changes for a bigger group
The link loads the Lake set: 123 landscape photos from Pexels and Pixabay. The sets are stock photos on purpose: for me the piece begins with the search query.
Press Next. With the defaults, Find clusters cuts the Lake set into nine groups, the same nine on every load. For a bigger group, change three controls under Settings before you press it: Strict pixel match to 0, so photos no longer have to match pixel by pixel; Distance threshold to 0.9, so looser likenesses still join a group; and Range of images in cluster to 15–20, so only the big groups are kept. Find clusters now shows three: Select cluster 2 with 18 photos, Select cluster 4 with 17 and Select cluster 5 with 20.
Cluster 4 as Find clusters made it, and the stack it turns into, framed. Nothing in the grouping knew what any of them shows.
The machine proposes the groups; which of them becomes a picture is my call. Untick Select cluster on the other two, keep 4, and press Next. The seventeen photos are by Alexey Chudin, Siarhei Nester, Arina Dmitrieva, Veronika Benz, Damir K, Lera Mk, Serafettin Ünye and an account named Pixabay on Pexels, and by ELG21, ChiemSeherin (two), lizzyliz, Pezibear, jggrz, Kodl68, Jayartphoto and aszak on Pixabay.
Auto alignment pulls the shores into one band
On the third page, Alignment starts at Off. Press Average clusters: the far shores sit at different heights and spread into a wide brown band. Set Alignment to Auto and run it again. The app measures how far each photo has to move to line up with the others and slides each one into place; it reports "Aligned (auto) — mean shift 231.5 px" under the result.
Off and Auto: on the left the far shores blur into a wide band, on the right they gather into one across the middle.
Why 10% is enough at seventeen
Trim mean cutoff (%), under Settings, decides how the stack collapses. At each pixel the app sorts the seventeen values and throws away that share from each end before averaging; at 50% only the middle value is left, which is the median. I made a trimmed mean the default because it drops passing things like birds while keeping more tone than the median.
Ten per cent of seventeen photos is 1.7, and the app rounds down: at 10% it drops the darkest and the lightest value at every pixel and averages the fifteen between. Where the gull flies, its photo is the lightest value, so it goes; even in a plain average it would be only one seventeenth of the colour. At 20% three values go from each end and the haze gains a little contrast. At 50% the median takes over and the picture breaks into saturated patches, with trees from single photos showing through again, because neighbouring patches now come from different photos.
10%, 20% and 50%: the haze holds until the median breaks it into patches.
One pixel where the gull's lower wing was, the lightest of the seventeen values there. The swatches are sorted by brightness; the app sorts red, green and blue on their own, which is why at 50% three photos supply the colour. The framed swatch on the right of each row is what lands in the picture.
Leave it at 10%: the haze is the picture.
From screen to object
Every result gets its own card. Public output is fixed at A5, 300 dpi, and the card's Average (PNG) button hands over the composite with its making written into the file: the set, the grouping, the alignment, the cutoff and the photographers. Sources and credits names them on the page too, with each platform's licence. Delete drops a result you don't want. For stacks made from the curated sets, Order a print of this creation sits under the buttons.
Under each composite: its alignment note, the photos that went in, who took them, the file and the print.
You can average your own photos too. Upload your own on the first page takes 10 to 40 at a time, Crop the whole set to one shape, keeps them in your session only and deletes them within two hours. Those stacks are yours to download; the print order is only offered for the curated sets.
Other controls, one line each
Colour match, Texture match, Structure match, Edge match: which other kinds of likeness count most when grouping.
Correct rotation & scale: also levels tilted frames and matches framing distance, at the cost of a tighter crop.
Alignment strength (%): how far each photo is moved towards the others.
Guided: paint the part of the average that should line up.
Select cluster N images: under each result, untick a photo to leave it out of the next run.
Your turn
Open the Lake set, make a big group of your own, and move one thing: the cutoff, or the alignment. Then send me what you made. I'll ask before I show it anywhere.