Can you increase the resolution of a photo

Resolution is a pixel count, and a tag will not add samples. Raise a low-resolution photo only after you have measured the long edge.

A cousin asked can you increase the resolution of a photo, and the photo was a face 80 pixels tall. I measured it twice, once from the chin to the top of the hair and once with the hair excluded, because hair was a soft mass and I did not want to flatter the count. Chin to hairline was 74 pixels. Chin to the top of the hair mass was 80. I am using 80 because that is the number I would have bragged about, and the brag is already too generous.

The file that arrived by email was reunion-crop.jpg, 640 by 480, 86 kilobytes, dated in the metadata as a camera from 2009 and then rewritten by the email save, so the date is a rumor. The face sits left of center. The rest of the frame is a kitchen, a paper banner, and a second person cut off at the shoulder. The request was an 8 by 10 "of her," by which they meant the face, large, sharp, and suitable for a frame they had already bought.

TL;DR. An 80-pixel face can be resampled to 160, 320, or 640 pixels tall. That is a real increase in pixel count and a fictional increase in recorded detail. The file kept a hair mass, a skin mass, two dark eye spots about 22 pixels apart, a nose-shaped shadow, and a closed mouth. Heavy JPEG took what the sensor had already softened. A face restore added a mole and parted the lips. I refused the 8 by 10 headshot, printed the whole scene at a size the samples could carry, and kept the tiny original next to every enlargement so the invention had a witness.

What I actually had, feature by feature

I opened the JPEG and set the view to one file pixel per screen pixel. Fit-to-window had made the face look merely "a bit soft." At 100 percent it was a cluster of tones.

The eye spots were not eyes in any useful sense. Each was a dark patch about 6 pixels wide and 4 tall, and the centers of the patches were 22 pixels apart. I could not see a pupil, an iris boundary, or a catchlight. The left patch was slightly higher than the right, which I wrote down, because later repairs liked to level the gaze and that tilt is part of how I recognize her in this particular frame.

The mouth was a darker notch, about 16 pixels wide, with no separated lips and no teeth. The nose was a lighter wedge, maybe 8 pixels across at the tip, with the nostril side lost in JPEG mush. One ear was visible, a C-shape about 14 pixels tall, and the earlobe was not distinct from the jaw. The hair had a side part that I could only defend because a lighter streak ran for about 20 pixels. Individual hairs were not in the file. Eyebrows were a single darker bar over each eye spot, about 2 pixels thick.

There was a collar, a stripe on the shirt about 3 pixels wide, and a pendant that read as a bright blob 5 pixels across. I do not know if the pendant was a circle or a heart. At 800 percent, nearest neighbor, it was four light pixels and a fringe. I am not going to call it a heart because hearts are a thing models like to complete.

The second person, at the right edge, had a face about 40 pixels tall and was partly out of frame. I told the cousin that this file could not become a portrait of both of them. They only wanted the one face. I still kept the second person in the scene prints, because cropping them out felt like a different photograph than the one that was taken.

The arithmetic of the frame they bought

An 8 by 10 inch print at 300 pixels per inch wants 2400 by 3000 pixels. At 200 pixels per inch, which is a kinder bar for a snapshot viewed in a hallway, it still wants 1600 by 2000.

The head, if it is to read as a portrait and not as a speck in a kitchen, needs to be something like 3 inches tall on that sheet. At 200 ppi that is 600 pixels of head. We have 80. The scale factor is 600 divided by 80, which is 7.5. At 300 ppi a 3-inch head wants 900 pixels, which is more than 11 times the recorded head. The 4× upscale makes an 80-pixel head into 320 pixels. Three hundred and twenty divided by 300 is about 1.1 inches. On an 8 by 10 that is a small head surrounded by a large invented kitchen, or, if I crop to the face first, a 1-inch head blown up to fill the frame with skin that was never sampled.

I wrote this on the job sheet in one line: pixel count is a multiplication, recorded detail is not. The cousin heard "so you can't do anything." That is not the line. The line is that I can make a larger file, and I cannot make a larger memory of her face than the phone stored.

Target	Pixels the head would need	Scale from 80	What those extra pixels are
Head 0.75 inch on a 6 by 4.5 snapshot at about 107 ppi	80	1×	The file itself
Head 1.1 inch after 4×, printed at 300 ppi	320	4×	Mostly proposal
Head 2.1 inch after 8×, printed at 300 ppi	640	8×	Proposal with confidence
Head 3 inches on an 8 by 10 at 200 ppi	600	7.5×	Past the recorded face
Head filling an 8 by 10 crop at 300 ppi	around 2000	about 25×	A different picture

The 1× row is the honest print. A 640 by 480 frame printed 6 inches wide lands near 107 ppi, and the 80-pixel head is about three quarters of an inch tall. That is a snapshot size. People lived with that size for decades. It looks like a photograph of a day, not like a studio sitting. The frame they bought was asking for the studio sitting.

What the phone and the JPEG had already thrown away

A 2009 phone sensor, in a 640 by 480 delivery, was not keeping 80 independent measurements of facial structure. The lens at that focal length, the focus miss I can still infer from the banner being softer than the face, and the noise reduction the phone applied, all happened before the file I received. The face being the sharpest thing in the kitchen only means the plane was nearer her than the banner. It does not mean the plane resolved an iris.

Then the JPEG. Eighty-six kilobytes for 640 by 480 is a harsh setting. I looked at the kitchen wall, which should have been a smooth painted surface. It was tiled. The tiles were 8 by 8 pixels, with a step in luminance at the boundary and a little chroma smear, green in one corner of a block and pink in the opposite corner. The shirt stripe, 3 pixels wide, was already broken where a block boundary crossed it. The pendant's edge rang, a dark pixel then a light pixel, which is the quantizer, not a chain.

So the detail ceiling is below 80 pixels of head. If I had to name the real samples, I would say the file knows perhaps a dozen tonal regions on the face and almost nothing about their boundaries. Anything I add inside an eye spot is fan fiction with a scientific accent.

The AI image upscaler is still the right place to see a 2× proposal, because a proposal can be better than a stair-step if I keep the authority in the original. It is not a time machine back to the sensor.

The enlargements, and the mole that was not there

I ran 2× first. Output 1280 by 960. The head became 160 pixels tall. The shirt stripe continued more cleanly than bicubic, which had turned it into a gray ribbon. The wall blocks softened. The eye spots gained a rounder boundary and a hint of a lighter dot that I could not find in the nearest-neighbor blowup of the original. I flagged the dot and kept going, because a hint is not yet a new person.

At 4× the file was 2560 by 1920 and the head was 320 pixels tall. The eye spots were now eyes, with an iris color that leaned brown. Her eyes are brown, so the prior got lucky, and luck is not a method. The eyebrows became hairs. The pendant became a small heart. I went back to the 800 percent nearest-neighbor view of the original pendant and there is no notch that would force a heart. The model likes hearts and circles, and it picked a heart. The earlobe separated from the jaw in a way that looked plausible and that I cannot verify.

At 8× the head was 640 pixels tall and the skin had pores in a pattern, a fine regular grain across the cheek. The original cheek is four JPEG blocks. Pores cannot be stored in four blocks. The mouth, which had been closed, developed a dark line between the lips that read as a slight part. The gaze leveled. The left eye, which had sat higher, was corrected by the prior toward a passport symmetry. That was the frame where I stopped calling the result her.

Face restore was a separate attempt on the 2× file, because I wanted to see the aggressive path once, on purpose, rather than discover it inside a general upscale. It made a sharp person. The catchlights were crisp and matched each other. A mole appeared on the left cheek, about where a beauty mark often sits in training sets. I overlayed the original, scaled with nearest neighbor so it would not blur, at 50 percent. There is no dark cluster in that cheek in the source. The mouth opened enough to suggest a tooth. She is not smiling in the kitchen. I rejected the face restore in writing, in the email, not only in the folder. Spoken rejections get "cleaned up" later by someone who was not in the room.

How I compared without fooling myself

The failure mode on a tiny face is falling in love with the larger file because it looks finished. I used three views, and I did them in this order so the finished one could not go first.

Nearest neighbor, 800 percent, on the original. This view is cruel and it is the evidence. Blocks, spots, the tilt of the eyes, the closed mouth, the blob of the pendant. I took a screenshot of that view and put it on the second monitor. It stayed there while I judged the others.

Bicubic at 4×. Soft, gray, no new mole, no heart, the eye tilt preserved because a blur does not re-pose a face. It looked disappointing, which made it a useful control. Disappointment is not the same as honesty, but in this case they lined up.

The model outputs, 2× then 4× then 8×, with the overlay. I scaled each output back down in my head by looking at the overlay flicker, original against proposal, on the eyes, the mouth, the pendant, and the shirt stripe. The shirt stripe could improve. The mouth could not change state. A stripe has no biography. A mouth does.

Identity checks I will not skip on a head this small

I wrote the checks as a list because I had already skipped one by running 8× first. The list lives on the job sheet now.

• The eye tilt stays. On this file the left eye spot sits higher. A repair that levels it has re-posed her.
• The mouth stays closed. No tooth line, no new smile, no parting that the 16-pixel notch did not have.
• No new mole, freckle, or earring. If I cannot find the dark pixels in the nearest-neighbor source, the mark is the model's.
• The pendant stays a blob unless the source forces a shape. A heart is a guess.
• Banner and any other lettering are cropped away from the model output. Invented words do not get a reprint.
• The 640 by 480 original stays in the delivery folder under the same base name as the print file.

I also looked at the banner text. In the original it is about 11 pixels tall and I can read "WEL" and then a guess. The 4× file made the banner say a word I will not repeat, because it was a real-looking word and it was not the word on the paper. I checked a second copy of the photo that turned up on a CD the cousin's brother mailed. On the CD the banner is still unreadable, but the letter count does not match the model's word. Invented text is how these jobs embarrass you. I cropped the banner out of any file that had been through a model, and I left it only in the untouched original.

There is a related argument about focus, rather than pixel count, in can you unblur a photo and in is it possible to unblur a photo. This kitchen frame is a little soft and a lot small. Unblur would be the wrong first name for it. The limit is the sample count on the face.

What I printed, and the size I refused

I offered three physical options and I refused one.

The refusal was the 8 by 10 headshot. Filling that frame with this face requires a scale I will not sign. I said that in a short email and I attached the 800 percent crop of the eyes so the refusal had a picture. The cousin understood it once they saw the squares. They had not understood it from the word "resolution."

The option they chose was a 6 by 4.5 inch print of the whole kitchen, made from a 2× file that I had checked against the eye tilt and the closed mouth. The 2× file is 1280 by 960. Printed 6 inches wide, that is about 213 ppi. The head is 160 pixels, which is 160 divided by 213, about three quarters of an inch. Same physical head as the 1× print. I am not going to pretend the 2× made her face larger on the paper. It made the shirt stripe and the door frame a little more continuous, and it slightly over-resolved the hair mass into strands I do not trust. I pulled a bicubic 2× of the same crop and, in the end, I composited nothing. I shipped the model 2× because the eyes survived the overlay, and I told them the hair is a little more specific than the file earns. They preferred that to the gray bicubic. Preference is allowed when the identity checks have passed. Preference is not allowed to open a mouth.

I also made a 4 by 6 of the CD scan, which was the same pixels, so that the family would have a print that had never been through a model. Two prints. One is the day. One is the day with a careful guess on the stripes. Neither is an 8 by 10 portrait.

The high-resolution request, as a phrase, is what I answered in how to make a high resolution image from a low resolution one. On this face the honest translation is: I can raise the pixel count to match a small print. I cannot raise the 2009 measurement.

Mistakes I am keeping in the note

I ran 8× before I had written down the eye tilt. When I noticed the gaze had leveled, I had to go back and prove it, and for a few minutes I doubted my own memory of the face. The screenshot on the second monitor is now a rule, not a preference. Memory is a prior, the same as the model's, and it is not more reliable just because it is mine.

I described the pendant as a heart in a draft email, because I had been staring at the 4× file. I deleted the sentence after I looked at the four source pixels. If I had sent it, I would have taught the family a piece of jewelry she may not have worn. That is a small lie and it is the same species as the mole.

I almost cropped the second person out "to give the main face more pixels." Cropping does not give the face more recorded pixels. It only removes context. The 80 stayed 80. I had started to believe my own sales language. The job sheet line, pixel count is a multiplication, is there because of that moment.

The generated sample I use for this conversation is a soft scene: a figure whose smile survives as a few tones, and a frame whose bars have melted. It is not her, and it is not a photograph of the kitchen. It is a stand-in I can show while the family file stays private. The point of the sample is the melted face. You can still tell there is a person. You cannot name the features you were not given.

Where the honest increase ends

You can increase the resolution of a photo if the sentence means the file gets more pixels along each edge. A 2× file has four times the pixels. A 4× file has sixteen times. The printer will accept the larger file and will not ask whether the iris was recorded.

You cannot increase the resolution of a photo if the sentence means the 2009 afternoon is recorded in finer detail than the phone kept. That information is not hiding in the JPEG, waiting for a stronger model. The wall blocks are evidence of what the quantizer did. The eye spots are evidence of what the lens and the sensor did. A model proposes a face that could have produced those spots. Many faces could have. The family knows which face it was, and the file does not, not below the scale of those spots.

I keep the 640 by 480 original in the same folder as the 2× print file, with the same base name, so a later relative cannot mistake the proposal for the camera. If someone reprints her at poster size in ten years, they will at least have the small witness beside the large guess. That pairing is the repair. The extra pixels are just the part the frame shop can hold.