How to make an image higher resolution when the file is a museum label screenshot

Higher resolution means more pixels on the edge you will print. Upscale the image 4× when 2× still leaves the paper short.

The catalog editor asked how to make an image higher resolution, and the attachment was not a photograph of the chair. It was ScreenShot-label.png, 1170 by 2532, a capture of a phone thread. Inside a message bubble sat a crooked photo of a wall label from a small design museum: side chair, 1924, oak and leather, gift of the H. M. Ellis estate. The editor needed that label in a printed catalog, and the production checklist said 7 inches wide at 300 pixels per inch. I restore pictures for families and for a few clients who publish, and a label is the kind of file that punishes optimism. Letters are either the letters, or they are a different date.

TL;DR. The usable crop, label-bubble-crop.png, was 780 by 540, and the cap height of the word Side was 14 pixels. A 4x pass, label-sr4x.png at 3120 by 2160, rewrote 1924 into something that read 192A, turned Ellis into Eilis, filled the holes in the lowercase letters, and grew wood grain inside the word Oak. The museum press PDF already had the letterpress as vectors, so the catalog retyped those lines. One graphite addition, seat restored 1978, N. Pell, was only in the photographs. I took that line from IMG_8841.HEIC, upscaled it 2x, redrew the broken 9 and the spurred P by hand, and placed it at 4.5 inches.

The thread was part of the picture

People say screenshot when they mean three different objects. This one was literal. Somebody stood in the gallery, photographed the label, texted the photo, and then a second person captured the thread. The PNG of the capture is sharp, because a screenshot of a screen is just the screen. The photograph inside the bubble is not. Messaging had recompressed it. You can see 8 by 8 blocks in the warm gray card of the label, and the black letterpress has gray fringes where the phone sharpened the strokes before the compress.

The first crop the editor made, label-bad-crop.png, was 800 by 560 and included a margin of the blue bubble, about six pixels of interface blue along the left edge. I know the width because I sampled it. The blue sat near 0, 122, 255, which is not a color that exists on a museum label. Any model that sees a strip of pure interface blue next to warm gray will explain it. On their test, which they named label-bad-upscale.png and which I refused to open twice, the blue had become a soft shadow down the left of the card, as if the label were lit from a window on the right. The gallery light was overhead. The shadow was the bubble.

I recropped inside the photograph, not on the bubble edge. label-bubble-crop.png is 780 by 540. No blue. A little of the plaster wall remains at the corner, which I wanted, because the plaster tells you this is a picture of a mounted card and not a flatbed scan. The card is slightly keystoned. The top edge is shorter than the bottom by about twelve pixels across the 780. I left the keystone alone until I knew whether we were keeping the picture of the card at all.

Fourteen pixels, counted on the glass

I zoomed until the word Side filled the display and counted the cap height. Fourteen pixels from the top of the S to the baseline. The stroke of the S is two pixels of near-black plus a one-pixel gray halo. The counter of the lowercase e in leather, the hole that makes it an e and not a c, is one pixel of gray card showing through, and on two of the e's that hole is already closed by the halo. This is the file. A checklist that asks for 2100 pixels across is asking me to invent the holes.

Seven inches times 300 is 2100. 2100 divided by 780 is about 2.7. Not a clean scale factor, which never stops anyone from pressing 4x and cropping. I wrote 14 and 2.7 on the job sheet. If a later export looks crisp and the cap height has become 56 pixels of confident wrongness, I will know the scale and the crime.

How to make an image higher resolution is a reasonable question when the picture is a face and you are honest about guessing skin. It is a bad question when the picture is a date. A face can be a little soft and still be the person. A 4 that has become a 9 is a different chair.

What 2x did, which was merely thick

I ran 2x first on the tight crop and saved label-sr2x.png, 1560 by 1080. I used the AI image upscaler because that is the tool path, and I watched the letters rather than the plaster.

The strokes thickened. The gray halo was pulled into the letter, so every word got bolder by about a pixel at the new size, which is half a pixel of source. The date still read 1924 if I knew it was 1924. If I covered the filename and asked myself cold, the 4 was soft but open. Ellis was Ellis. The counters of the e's were still precarious, some open, some shut. The word Oak stayed a word. No grain inside the strokes yet. The plaster gained a little stipple it did not earn.

At 1560 pixels wide, a 7 inch placement would be about 223 pixels per inch. The checklist wanted 300. A production manager can sometimes waive that for a caption-sized documentary photo. This editor's printer would not waive it. The checklist is in the RIP ticket. So 2x was not going to be the hero at 7 inches, and I did not want it to be. Letters at 223 pixels per inch that are already haloed look like a bad scan of a bad photocopy. You do not frame that across a column and call it the object record.

I kept label-sr2x.png because it was the least dishonest enlargement, and because I suspected we would not use it for the letterpress at all. Suspicion is not a plan. The 4x file made the plan obvious.

The 4x file rewrote the wall

label-sr4x.png is 3120 by 2160. I let it run so I would have the failure in the folder, not as a story I told without evidence. The plaster looked like plaster from a furniture catalog. The card looked clean and expensive. The type looked like type until I read it.

The date. In the source, 1924, the 4 is open at the top right. In the 4x file the opening collapsed and a spur closed across it. At arm's length I read 1929. At 100 percent I read 192A, the 4 having grown a crossbar that did not connect to the right place. I circled it in red on a proof JPEG named label-sr4x-DATE.jpg and sent that one circle to the editor before I sent anything pretty. She wrote back that the chair is 1924 in every other source. Good. We agreed the pretty file was a corrupted label.

Ellis. The second l picked up a dot of the i's halo and the model separated the stroke into Eilis. It is the kind of error a tired proofreader misses because the word is a name and names are allowed to look odd. H. M. Ellis is on the credit line of the exhibit. Eilis would have been a new donor. I am not in the business of minting donors.

leather. Three of the counters filled. The word did not become another word. It became a dark worm with ascenders, which at column width might have passed as "letterpress texture." It is not texture. It is the hole in the e being treated as noise in a black stroke. Models that have learned photographs think holes in dark shapes are often dirt. Counters in type are not dirt.

The word Oak grew a material

This one I would not have predicted, and I should have. The label says Oak, and the wall beside the card is a warm brown, and the chair in the same gallery photo is oak. Inside the strokes of Oak, only in the 4x file, a faint grain appeared. Short dark lines following the stroke direction, like a photographer had shot a wooden letter. The source strokes are flat black ink. I sampled them at 20, 18, 16 or so. The 4x strokes vary from the teens into the 40s along those grain lines. The model illustrated the meaning of the word.

That is a different failure from a soft 4. A soft 4 is a loss. A wooden Oak is an addition. The catalog is about a chair that is oak, not about a label printed on oak. If someone later trained a search on the picture, or a conservator glanced at the materials line, the grain inside the glyph is a false material. I duplicated the crop, desaturated it, and the grain was still there. It was not a color fringe. It was drawn.

I did not run a third pass with a different creativity setting. Two scales had already sorted themselves. 2x thickened. 4x rewrote. The in-between was not going to become a typeface. For a longer argument about when a file from a camera can survive enlargement, the vacation-frame note is not this job, and the portrait note is not this job either. What I owed the editor was a way out of the screenshot.

The press PDF made the pixel count boring

I asked a question I should ask on the first call. Does the museum have the label text in a document. The editor sighed, which I have learned is often a yes. An hour later I had press-release-chair.pdf. Page two is the object tombstone, set in a real font, vectors, the date 1924 as 1924, Ellis spelled with two l's and no phantom i, leather with counters you can fall into. The gift line matches the wall. The chair title matches the wall.

Setting that text in the catalog face took the designer forty minutes, including a short fight about whether to keep the museum's punctuation. Forty minutes is the whole resolution job for those lines. There is no pixel count because there are no pixels. How to make an image higher resolution, for letterpress that already exists as outlines, is to stop using the image.

I exported a reference PNG from the PDF at 2400 pixels wide, press-tombstone-ref.png, only so I could put it beside the screenshot and show the intern the counters. We did not place the PNG. We placed live type. The RIP checklist's 300 ppi rule does not apply to live type. I wrote that in the ticket so nobody would rasterize the caption "just to be safe" and hand me another screenshot in September.

The chair photograph was already solved and I am glad to say so in one paragraph. chair-press.tif, from the same press packet, is 4000 by 5000, a gray sweep, the chair in focus, scuffs on the front leg visible. I did not upscale it. I did not enhance it. The catalog designer placed it at 6 inches and still had pixels to spare. The sick file was the label, because somebody in the gallery had treated a wall card as a thing you text.

If you are hunting a camera original or a print you can rescan, that path is closer to how to get a high resolution image. This label did not have a negative. It had a PDF and a graphite scribble.

The graphite line the PDF did not know about

Under the letterpress, in pencil, someone had written seat restored 1978, N. Pell. It is not in the press release. It is on the wall card, small, slightly slanted, the pencil softer than the type. The editor wanted it in the caption because the restoration date is the point of the essay. A retype of a pencil line is allowed, and I offered it. She wanted the hand, the actual marks, because Pell's note is a conservator's hand and the essay is about that hand.

The screenshot's version of the line was about 400 pixels wide and nine pixels tall. Useless. The counters problem is bad. A pencil at nine pixels is a gray bar.

I asked for the camera roll, not the thread. IMG_8841.HEIC was still there, 4032 by 3024, a wide shot of the gallery with the chair left of center and the label a small rectangle on the wall. Nobody had cropped it in the museum. They had zoomed with their fingers and texted the zoom, which is why the screenshot looked like a label photo and the HEIC looked like a room. The label in the HEIC is about 900 pixels wide. The pencil line, cropped as pell-line.png, is 740 by 110. Still humble. Much more pencil than the bubble had.

You can see the tooth of the card in the pencil, a broken cadence in the stroke of the 1, and the 9 starting to close because the pencil doubled back. N. Pell is readable. The P has a clean bowl. This is the master for that one line. It is not a master for a 7 inch reproduction. 740 pixels at 7 inches is about 106 pixels per inch, and pencil at 106 looks like a photocopy of a photocopy. I said no to 7 inches of pencil.

Two letters, redrawn, and then I stopped

I upscaled the pencil crop 2x only. pell-line-2x.png is 1480 by 220. The enhance photo quality tool is the one I opened next, and then I mostly did not use it. A contrast push on pencil makes every tooth of the paper into a black speck and makes the doubled 9 look intentionally gothic. I lifted the midtones a little so the card was not a muddy gray, amount small enough that the pencil did not break into dots, and I stopped. Enhancement, on this line, was a way to see the strokes. It was not a way to invent stroke.

At 2x the 9 in 1978 closed. The pencil's double pass had been ambiguous, and the model chose a closed bowl. On the HEIC crop, zoomed, the bowl is open by a hair of card color. I redrew the opening. One path, color sampled from the pencil at 70, 68, 64, not from the model's darker guess. The repair is maybe fifteen pixels in the 2x file. I did it on a layer named redraw-9 and left the model's pixels underneath, hidden, so a later person can turn the lie back on and see why I hated it.

The P in Pell grew a spur at the top of the stem, a little flag the model adds to degraded verticals. The HEIC stem is blunt. I cloned the spur out with card color from immediately beside it, redraw-p, twelve pixels. Then I sat on my hands. The rest of the letters are soft and they are the pencil. Soft is the truth of a note written on a wall card and photographed from across a bench. If I redrew Pell into a clean signature I would be forging the conservator.

I did not run 4x on the pencil. The letterpress 4x had already taught me what this model does when it is bored and confident. A forged 1978 with a prettier 9 would be a worse sin than a thick Oak, because the restoration date is the claim of the essay.

Where it sat on the page

The designer gave the pencil line 4.5 inches of width in the caption column, under the retyped tombstone, with a printed credit in the catalog face that says wall note, N. Pell. 1480 pixels across 4.5 inches is about 329 pixels per inch. The checklist's 300 is met without pretending the line could survive 7. At 4.5 inches you can see that it is pencil. You cannot see my fifteen-pixel opening in the 9 unless you know to look, and the opening matches the HEIC, which is the standard I care about.

The full-card picture, the screenshot crop, did not go in the catalog. We talked about a small documentary frame, 3 inches, of the whole card including the crooked plaster. At 3 inches the 2x file is 520 pixels per inch and the thickened type is still the wrong weight, and the keystone makes the card look like a mistake. The designer used the chair photograph and the retyped tombstone and the pencil strip. The screenshot stayed in the job folder as evidence of what not to place.

I straightened nothing on the placed strip. The pencil slants by about two degrees. Straightening it would have made it look set, not written. The slant is doing documentary work.

Color on press: the pencil strip is warm gray. I converted a copy to the catalog's gray profile and the card went greenish. I pulled back to sRGB and let their RIP handle it, then I checked the contract proof. The card was warm again, the pencil a step lighter than I wanted, still clearly pencil. We did not chase it. Chasing gray card into a perfect neutral is how you posterize the tooth.

What I measure before I agree to a checklist

A production checklist that says 300 at 7 inches is a pixel budget. It is not a sharpness budget and it is not a truth budget. I answer it with a crop and a cap height, or with a pencil height, before I answer it with a model.

• If the strokes are under about 8 pixels and the words exist as text, retype.
• If a blue or gray UI edge is in the crop, the crop is wrong. Recrop. Do not upscale the interface.
• Read every proper name and every digit at 100 percent after any scale factor. Circle them on a JPEG and make someone else read the circle.
• Keep the camera original even when it is a wide gallery frame. The zoomed text is a worse master.
• Place the file at a size the pixels can support. 4.5 inches of pencil beat 7 inches of a closed 9.
• Do not enhance type into looking more like type. You will enhance the halo.

The cousin of this question, when someone says the photo is low resolution and means the tag or the print, sits in improve photo resolution. Whether a model can increase detail in a photograph at all, the general case, is can you increase the resolution of a photo. I am not repeating those essays. Jordan's mole and this chair's date are different kinds of facts. I will not move either of them to satisfy a robot.

The folder, so the next catalog does not grab the 4x

File	Pixels	Decision
ScreenShot-label.png	1170 by 2532	Thread capture. Chrome included. Not a master.
label-bad-crop.png	800 by 560	Six pixels of bubble blue. Rejected.
label-bubble-crop.png	780 by 540	Tight crop. Cap height 14. Source for tests.
label-sr2x.png	1560 by 1080	Thickened type, still readable. Not placed.
label-sr4x.png	3120 by 2160	192A, Eilis, filled counters, grain in Oak. Rejected.
press-release-chair.pdf	vectors	Letterpress retyped from this.
IMG_8841.HEIC	4032 by 3024	Wide gallery frame. Holds the pencil.
pell-line.png	740 by 110	Pencil crop from the HEIC.
pell-line-2x.png	1480 by 220	Placed at 4.5 inches after two letter repairs.
chair-press.tif	4000 by 5000	Chair photo. Untouched.

The placed row is the pencil. The tombstone is not in the table as pixels because it stopped being an image. That absence is the success.

This sample is not the label. It is a generated room, a teal panel, a sun in a square, a simple figure, enlarged by repeating pixels until the window and the panel edges become blocks. I show it to editors who think their screenshot is "basically high resolution" because the PNG dimensions are 1170 by 2532. The PNG is a high-resolution picture of a low-resolution photograph sitting in a thread. The blocks in the sample are honest about repetition. The blocks in the gray card of ScreenShot-label.png are the same idea, wearing a museum's colors.

The sentence I sent instead of a file

The editor asked me, at the end, for a single export she could drop in the frame and forget. I did not send one. I sent the retype instruction, the pencil strip, and a note that said the screenshot cannot be made into the wall card. How to make an image higher resolution, in the one case where I will say it plainly, was to stop asking the screenshot to be the card.

If the press PDF had not existed, I would have transcribed the letterpress myself from IMG_8841.HEIC, marked the 4 as confirmed against the open counter I can still see at the HEIC's size, and sent the transcription to the museum for a check. I would not have shipped label-sr4x.png with a shrug about "AI text." A chair dated 1929 in a book about 1924 is a correction the internet will repeat. The pencil repair I will defend. The wooden Oak I will not.