Editing an Image

The order

The branch manager wanted a vinyl banner three metres wide for the front of the building, with the bank’s crest, its name in English, and its name in Tamil underneath. He emailed the logo. It was a PNG, 900 pixels across, saved from the bank’s own website some years earlier by somebody who no longer worked there.

The shop’s owner did the arithmetic on the back of the order slip. Three metres is about 118 inches. Nine hundred pixels across 118 inches is under eight pixels per inch.

That number by itself does not condemn the job. Billboards are routinely output at ten to twenty pixels per inch and nobody standing on the road notices, because a photograph read from twenty metres has no detail at that scale anyway and degrades gently. The rule of thumb that says 300 is for something held in the hand.

But this was not a photograph. It was a crest with a fine rule around it, a serif wordmark, and Tamil letterforms whose vowel marks are small strokes sitting above and beside the consonants. Hard edges and small strokes are exactly what a low sample count destroys, and a guest standing two metres from the entrance would be reading it at arm’s length, not from a moving car.

Three routes, none of them free

Upscale it. Upscayl, free, twenty minutes, four times or eight times. On skin and brickwork it would have been remarkable. On letterforms it predicts plausible shapes from the low-resolution evidence, and plausible is not the same as correct. The result looks crisp, which is the dangerous part, because nothing in the file announces that the strokes have been invented.

Redraw it. The crest could be traced as vector in Inkscape in about four hours, and vector has no resolution at all, so three metres or thirty would be the same file. The Tamil wordmark was the problem: the original was set in a licensed font nobody in the shop had, so it would have to be traced letter by letter by eye. That is a full day, and a full day is the wedding-card job that was already on the bench.

Ask head office. One phone call and, if somebody answered, the original artwork. The bank’s marketing department is in Coimbatore and had not replied to an email about a rate card in two weeks. The function was on Friday and vinyl printing needed the file by Thursday morning.

The cost sat on every side. A day of a two-person shop’s capacity is a real loss. So is a banner with a slightly wrong crest, at a function the district collector was attending, produced by a shop whose name would be on the bottom edge.

What the owner actually decided

He ran all three at once, which is the decision worth noticing. He put his assistant on the Inkscape trace on Tuesday morning as insurance, sent the upscaled version to the large-format printer as a single A3 proof so he could look at it rather than argue about it, and telephoned Coimbatore three times until somebody found the artwork.

The A3 proof settled the argument on its own. At one metre the English wordmark looked acceptable. The Tamil vowel marks were visibly the wrong shape (one had acquired a curl it does not have) and the fine rule around the crest had gone slightly wavy where the model had smoothed it.

The job and the machine

Ninety-six pages, 340 student portraits, sixty event photographs and four full-bleed spreads. No Photoshop licence, so the teacher worked in Photopea, which runs in a browser tab, is free, and reads and writes layered PSD. She had used it for two previous yearbooks without trouble.

The trouble arrived with the class-of spread: forty cut-out portraits arranged on a painted gradient background, built at 3,000 pixels square in 16-bit because the gradient banded visibly in 8-bit when she stretched it. Each portrait was a smart object with a mask and a clipped Curves layer. At about the thirty-second portrait the tab reloaded, and two hours of arrangement went with it.

The mechanism is not mysterious. Photopea lives inside a browser tab’s memory allocation, which is roughly two to four gigabytes depending on the browser and the machine, and there is no scratch disk to overflow into. A desktop application in the same position writes to disk and crawls. A browser tab in the same position stops.

The choice

Move to GIMP. Free, tile-based memory management, swaps to disk, so on an eight-gigabyte laptop a large file is slow and keeps working, which is usually the more useful behaviour. The cost is real: GIMP has no smart objects at all, so every scaled portrait is a resampled layer rather than a re-renderable one, and until version 3.0’s layer effects it had no true adjustment layers either, so a tone move is baked into a duplicate. Three weeks before the printer’s deadline, that is a new tool and forty placements to rebuild.

Stay and change the work. Split the document: portraits retouched in one file, the spread assembled in another from flattened TIFFs. Stamp a checkpoint and hide everything beneath it. Drop to 8-bit where the job does not need 16. Work at half size for placement and apply the settings at full. All of that is free and all of it costs her the re-editability she had been relying on.

Borrow the computer lab’s newer machine. Sixteen gigabytes, available one hour a day between practicals, and the school’s rules say files stay on the lab machine. One hour a day for a ninety-six-page book is not a plan, and a file that cannot leave the room cannot be worked on at home in the evening, which is when most of the yearbook was actually made.

What she weighed

The honest comparison was not GIMP against Photopea. It was which parts of this job need which property. The 340 portraits were reproduced at about 800 pixels on the page and needed a crop, a small exposure correction and a spot heal. Nothing there stretches a gradient, so nothing there needs 16-bit, and nothing there needs a smart object. The two gradient spreads needed 16-bit badly and needed almost no layers.

That observation split the job cleanly, and it is the part worth carrying away: the constraint was never the machine in general, it was one property of one part of the work.

The job that actually arrived

Four hundred and twelve items, photographed over three years by three different people against three different grey walls. The new marketplace wanted 2,000-pixel squares on white, the product occupying most of the frame, one main image per listing. Nobody had ever cut any of them out.

A freelancer in Delhi quoted per image, by hand, five working days. The quote was defensible and the money was not in the budget for a migration that was itself a cost. The alternative was the free pipeline: rembg for the masks, ImageMagick to flatten onto white and pad every item to the same canvas so a tall lamp and a wide tray sit at the same visual scale in the grid.

There were two complications, and they pulled in opposite directions.

Ninety of the items are polished brass. A polished surface is an image of whatever is in front of it, which in these photographs was a window, a tripod and a man in a blue shirt. A segmentation model outputs coverage, and coverage is not the right answer for a surface whose appearance is mostly reflection.

Eighteen items were an unreleased range under an agreement with a buyer in Hamburg. Running them through a web background remover means uploading photographs of an unannounced product line to somebody else’s server, under terms of service nobody in the business had read, several of which reserve the right to use uploads for improvement.

Three routes

A web service. Fastest by a distance, and the free tier returns a small image while the full-resolution file is paid. It also disposes of the confidentiality question by answering it badly.

rembg locally. Two hours to install on a machine with no graphics card, then two to five seconds per image on the CPU, unlimited, nothing uploaded. It would still fail on the reflective items, on the teapots with handle gaps, and on the pierced lattice lamps, because those are the predictable failures of a salient-object model.

Pay the freelancer for all of it. Five days, consistent quality, a bill, and no learning that stays in the business.

The decision that actually mattered was not which tool. It was whether to sort the catalogue into automatic and by hand folders before running the batch, which costs a morning of looking at 412 thumbnails and deciding, or to run everything and sort out the mess afterwards.

The part that is not about cut-outs at all

For a catalogue, the individual cut-out matters less than the set matching, and that is a set of decisions to be made once and written into the script rather than exercised four hundred times.

The canvas had to be 2,000 pixels square, because the marketplace said so. The product had to be resized to fit inside 1,800 first and then centred on that canvas, which is what gives a tall lamp and a wide tray the same hundred pixels of breathing space and stops the grid looking wrong even when every file is technically the same size. The background had to be exactly 255 white, because the platform checks automatically and rejects what is not. Shadows had to be present on everything or on nothing: a grid where half the items float is the clearest possible sign of a job done in two sittings by two people. And the file names had to be machine-readable, because 412 of them were going into a spreadsheet.

None of that is a skill. All of it is the difference between a set that looks commissioned and one that looks assembled, and it takes twenty minutes to decide and two seconds to encode.

The setup, and what was wrong with it

A borrowed camera on a tripod, a maroon cloth backdrop taped to a noticeboard, six-minute slots, and a student from the photography club operating it. Daylight came through a window on the left. Fluorescent tubes ran along the ceiling. Everything was on automatic: automatic exposure, automatic white balance, and the camera shooting JPEG because nobody had thought about the card filling up.

By the time a lecturer looked at the files at lunch, sixty-three students had been photographed and two problems were visible on the laptop screen.

The exposures were not consistent, and they were not randomly inconsistent. The camera’s meter reads the frame and pulls towards a middle value. A student with dark skin fills much of the frame with a dark subject, so the meter concluded the scene was under-lit and opened up, or, in the sequence where the backdrop dominated, did the opposite and pulled down. Across the sixty-three, the darker-skinned students were on average noticeably under-exposed, and the failure was systematic rather than scattered.

The backdrop wandered. Automatic white balance re-decides per frame. A frame with a lot of maroon in it reads as a colour cast and gets corrected towards neutral, taking some of the maroon with it. Sampled with a 5 by 5 average, the backdrop measured between 243 and 251 across the set and carried a slight green shift on the frames where the tubes had dominated over the window.

No individual photograph looked wrong. The set, laid out as a contact sheet, looked assembled rather than commissioned.

The decision at lunch

Stop, go manual, reshoot the sixty-three. Correct, and forty-one of those students had already left campus, several of them travelling home that evening and not returning before the prints were needed.

Carry on and fix it in the developer. Nine hours of per-frame work, and the arithmetic of the under-exposed frames is unkind. Lifting exposure afterwards multiplies the signal and the noise together, and on an 8-bit JPEG it does so inside 256 levels. On darker skin, where much of the form is carried by the specular highlights, a heavy global lift also flattens the face, which is a worse outcome than the under-exposure it is fixing.

Switch to manual now and match the first sixty-three towards the second set. Cheapest, and it accepts two sets that a contact sheet will always be able to tell apart.

The cost was on both sides and it was not only technical. Asking forty-one students to come back is asking forty-one people for an hour of a week they had already finished.

The decision made three weeks earlier

There was a fourth cost, and it had been incurred before anybody picked up the camera. The club had set it to shoot JPEG so the card would not fill.

In a raw file, white balance is a pair of multipliers that have not been applied yet, so changing it costs nothing at all. In a JPEG it has been baked in, and correcting it means stretching one channel against the others inside 256 levels. Correcting a green fluorescent cast in a JPEG means multiplying part of the green channel down and the others up; a channel stretched that far bands in the smooth areas and goes noisy in the dark ones, and in the deepest shadows there is nothing left to stretch.

The same argument applies twice over to exposure. A raw file at 12 or 14 bits carries levels in the shadows that an 8-bit gamma-encoded JPEG simply does not have, which is exactly the region the under-exposed frames needed lifting from.

So the frames that most needed correcting were the frames least able to take it, and the reason was a setting chosen for storage on a day when storage was not actually the constraint.

What came in

A colour print, five inches by seven, from a wedding in 1971. Salmon-pink overall. A tear running up from the lower left corner, and the corner itself missing: about two centimetres of what had been a plain wall. And a crease, folded and refolded, running diagonally across the grandmother’s left eye and cheek.

The pink is not the paper yellowing. Chromogenic colour prints hold three dye layers, and they do not fade at the same rate; in the common processes the cyan goes first. Remove cyan from cyan, magenta and yellow and what is left reads as red-orange. That is why every family photograph from this period has the same cast, and it is why the correction is a red-channel curve pulled down in the midtones rather than a warmth slider.

The limit is equally specific. Where the cyan dye has gone, there is no cyan information to amplify. You can neutralise the cast and what you get back is a low-saturation, greyish image, because the data in that channel is absent rather than weak. Pushing further produces channel noise and banding, not colour.

The tear and the missing corner were straightforward: a tear is a tonal line over surviving image, and the corner had contained nothing but wall, so it could be cloned from the opposite corner. Real content, correctly derived.

The crease across the eye was the job.

The decision

The family had brought a second print from the same day. It is badly out of focus and it shows the grandmother from the other side, three-quarters rather than near-frontal.

Rebuild the eye from the second print. Six hours: extract the eye region, warp it to the new angle, match the grain and the tone, sharpen it to sit with its surroundings. The result would be derived from surviving evidence of that person on that day. Six hours is not in a two-and-a-half-thousand-rupee restoration.

Generate or hand-paint an eye. Forty minutes. It would be a plausible eye. It would not be her eye, and it would hang on a wall a grandchild grows up looking at.

Reduce the crease and leave it visible. Two hours, entirely honest, and the family will see the crease every time they pass it.

The studio’s owner does this arithmetic often enough to have a rule about it: the decision belongs to the family, and it cannot be made by them unless somebody explains the difference between the three, which takes about four minutes and is the part most studios skip.

Why the order of work was not negotiable

There was a second decision, smaller and entirely technical, about what to do first.

The instinct on a damaged print is to start with the interesting part, which here was the eye. Doing that costs you the work twice. Restoration runs from large damage to small: archive the untouched scan, recover global tone, correct the global colour, then the tears and missing corners, then the creases and stains, then the dust, then local tone, then grain matching, then output.

The reason is that every large repair changes what the small ones have to match. Repair a crease against a faded, pink, low-contrast original and then recover the contrast, and the repair is now visibly a different tone from everything around it. Recover the contrast first and a surprising amount of what looked like catastrophic damage turns out to be faint: on this print, two marks that had been quoted as stains disappeared entirely once the black point was where it belonged.

Working small to large means redoing the small work after every large repair. Working large to small means doing it once.

The photographs they had

A four-storey building with glass balcony fronts, a tiled forecourt still wet from the morning, and two cars parked along the near side. Shot on a flat white overcast day, because that was the day the keys were available. The agency’s house style, like most of its competitors’, puts a blue sky behind every exterior.

The junior who had been given the files ran an automatic sky replacement on the first one. The masking round the neem tree at the edge of the frame was better than he would have managed by hand in an hour, and the result was obviously wrong in a way he could not immediately name.

It was wrong in the way a sky replacement is always wrong when nobody checks the four agreements.

Direction. The new sky had a bright area low on the right. The building’s soft shadows, such as they were, did not agree with anything, because an overcast day produces no directional shadow at all.

Quality. This is the one that does the damage. The original light was soft, so nothing in the frame casts a hard shadow. Put a clear sky with a visible sun behind it and the scene becomes a bright day in which nothing casts a shadow, which every viewer registers instantly and almost nobody can articulate.

Exposure relationship. Against a bright sky, the building stayed at its overcast exposure and read as dull and slightly grey.

Spill. Nothing in the scene had been lit by the new sky. Roofs, the tops of the cars and the wet forecourt all carried the colour of the old one.

And then the thing that actually killed it: the old sky was still in the building. Grey, in every balcony pane, in the puddles on the forecourt, and along the top of both cars.

Three routes

Replace properly. Choose a sky whose light matches the plate rather than the other way round, add horizon haze, grade the foreground, add spill, and then replace the reflection in every pane of glass, every puddle and two car roofs. That is the better part of a day per photograph, and eleven listings were due.

Improve the sky that is there. A graduated Curves layer over the existing flat sky, twenty minutes, no new content, no reflections to chase. The sky stays grey.

Wait for a clear morning. The photographs are correct and the listing goes live a week later, in a market where the first week carries most of the enquiries.

The check that would have taken thirty seconds

The junior had looked at the replaced sky at fit-to-screen for ten minutes and knew something was wrong without being able to name it. Three checks would have named it, and all three are seconds of work.

Flip it horizontally. The eye habituates to an image quickly, and a mirrored version is unfamiliar again, so errors that had stopped registering reappear. Flip back afterwards.

Shrink it to 300 pixels. Errors of tone, contrast and light direction survive at that size and errors of edge detail disappear. If it still reads as wrong small, the problem is structural rather than in the mask, which is what it was here.

Blur it heavily. Twenty pixels of Gaussian on a flattened copy leaves only the broad tonal and colour structure. The building stayed a dull grey lump under a bright sky, which is the exposure relationship failing, and it was obvious in a way it had not been at full size.

None of those checks needs experience. They need somebody to have been told that they exist.

One frame that works

Out of four hundred photographs from a field visit, one is right: light falling through a doorway onto a teacher at a board, nine children at desks, a wall of exercise books behind. It is a 4:3 frame with the teacher a third of the way in from the left.

The report needed it at five shapes. Square for the donor’s grid, 4:5 for the feed, 9:16 for a story, 16:9 for the website hero, and 3:1 for a banner across the top of a landing page. Cropping to all five works for two of them and destroys the composition in three: the 3:1 crop leaves a strip containing the teacher’s shoulder and half a child.

The obvious answer is generative expand: extend the canvas and let a diffusion model fill the new area, conditioned on the pixels along the boundary. It is genuinely the most useful everyday application of generative imaging in commercial work, and it is the least discussed.

What made it not obvious here

The photograph shows nine identifiable children. Their parents had signed a consent form, written in Hindi and English, covering the NGO’s annual report and its website.

The form says nothing about uploading the photograph to a third-party service. A hosted generative expand means sending an identifiable image of minors to a company whose terms of service nobody in the office had read, several of which reserve rights over uploads.

The communications officer laid out three options in a note to the programme head.

Hosted expand. Two minutes per ratio, the best first result, and it puts the photograph somewhere the consent form does not reach.

Local expand. Fooocus on her own laptop, which has no graphics card. An afternoon to install and understand, then about three minutes per generation. Nothing leaves the machine. It still generates content that was never in that room, and in the first attempts at 9:16 it twice produced an extra child-shaped figure at the lower edge, which is a characteristic failure at the boundary of an outpaint around people.

No model at all. Mirror the edge where the extension is pure texture (floor, wall) and place the photograph on a coloured panel with the report’s title for the two extreme ratios. Exact, free, deterministic, invents nothing, and it reads as design rather than as more room.

The cost on the third option is real and worth stating: it looks like a layout rather than like a photograph, and the donor’s own materials are full of full-bleed photographs. Choosing it is choosing to look different from everybody else in the folder.

The workflow that would have applied either way

One thing was settled before the confidentiality question arose, because it is settled in every job of this shape: you do not run five separate extensions for five ratios.

Extending a canvas five times from the same original produces five different inventions of the same wall, and they do not agree with each other, so the set looks like five photographs of five rooms. The professional route is to extend the master generously in both directions once, keep the extension on its own layers, and then crop that single file to each ratio. One generation, five outputs, and the subject sits where it should in all of them.

The extension itself has to be staged. Asking for a hundred per cent more width in one operation gives the model a thin strip of real pixels along one edge to condition a very large region, and it drifts: the wall changes colour, the floor line bends. Twenty to thirty per cent at a time, with a generous band of the existing image inside each generated region, gives each pass enough context to stay coherent.

And whatever comes back is cleaner than the photograph it is attached to, so grain, vignette and black point have to be unified across the whole extended frame afterwards.

The measurement

Somebody in the office opened the portal’s admissions page in a browser with the network throttled to a slow connection, which is a thirty-second test that nobody had ever run. Fourteen seconds before any picture appeared. The page had thirty-one images on it and every one of them began downloading at once.

The causes were not subtle.

Almost every image was a PNG, because the staff who upload them were told years ago that PNG is the high-quality format. A PNG of a photograph is routinely three to six times the size of a visually identical JPEG, because the format is lossless and designed for flat colour and hard edges. The hero (a group photograph of a prize-giving) was 2.4 megabytes.

Almost every image was 3,000 to 4,000 pixels wide, straight from a camera, displayed in a column about 700 CSS pixels across. On a ratio-2 phone that column needs roughly 800 real pixels. The browser downloads all four thousand of them, then scales down in encoded space with no gamma correction and no output sharpening, which is also why the pictures looked slightly soft.

And nothing was lazy-loaded, so a thumbnail in the footer competed for bandwidth with the photograph at the top of the page.

The choice

An image CDN. Ten minutes to switch on. One master per image and derivatives come from URL parameters (width, format, quality) so the widths and formats problem disappears without anyone re-exporting anything. It is a monthly bill against a budget line that does not exist, and it is a dependency: every image on a government portal would then be served through a private company’s infrastructure, on their pricing and their uptime.

Fix it properly, themselves. Re-export 900 images to three widths in three formats with a shell loop, write srcset and sizes into the article template, and change the upload rule in the CMS so staff cannot post a 4,000-pixel file at all. Free, permanent, and two days of a developer who is the only developer and does not have two days.

Do the cheap half. Fix the twelve pages that carry most of the traffic, set the loading and priority attributes site-wide, and leave the rest until there is time.

The temptation was the CDN, because it is a decision that can be made in a meeting and finished the same afternoon. The argument against it was not technical: it was that the underlying problem (staff uploading camera files) would be hidden rather than solved, and the bill would grow with the site.

Arguing in budgets rather than in settings

The meeting where this was decided nearly turned into an argument about the quality slider, which is an argument nobody wins because a quality number is not comparable between two encoders and does not describe an outcome at all. It scales a quantisation table. Feed it a plain studio backdrop and quality 82 gives a small, perfect file; feed it a crowd under autumn foliage and the same setting gives a large file that still shows artefacts.

What ended the argument was a budget. A full-width photograph on a page gets 100 to 300 kilobytes; a grid thumbnail gets 15 to 40; a whole article page including its images aims under a megabyte, and going over it is a decision somebody makes rather than something that happens.

That converts an aesthetic disagreement into arithmetic. If the hero has to be 1,000 pixels wide and fit 200 kilobytes, the quality setting is whatever achieves that, found by exporting two versions and flipping between them at 100 per cent rather than by nominating a number in advance.