Both sides of an ID card on one sheet, at exact size, lined up back to back.
A single HTML file. No build step, no dependencies, no server, no uploads.
Photocopying an ID card onto one sheet is a guessing game. You put the card on the glass, print, flip the card, flip the paper, print again — and the second side lands somewhere you didn't want, upside-down, or a different size from the first. Repeat until the paper runs out.
Two things go wrong, and they're easy to confuse:
- Registration — the back doesn't land behind the front, because you can't predict what your printer does with a re-fed sheet.
- Scale — the two sides print at different sizes, because they were cropped differently.
Duplex fixes the first by construction. It fixes the second at the source, with a four-corner crop, and gives you the tools to see and fix whatever is left.
Registration is computed, not guessed. Place the front 18 mm from the left edge of an A4
sheet and the back goes to 210 − 85.6 − 18 = 106.4 mm. Flip the paper and the two cards are
exactly on top of each other. The mirror is its own inverse, so you can drag either face and the
other follows. Change the paper and the arithmetic changes with it — the sheet is a variable,
not a constant baked into the layout.
Each side is cropped by its corners, not by a box. A card photographed at a slight angle, or squashed by a scanner app, is no longer 85.6 : 54 in the photo. An aspect-locked crop box can only grow evenly, so fitting it to that card overshoots one axis or cuts the other, and that side prints at a different size from its partner. Instead you drag each of four corners onto the card's own corner, and those four points are warped onto exactly 85.6 × 54 mm. A tilted, angled or squashed photo comes out the same size as its other side, because both are defined by the card's real edges.
A back whose card is squashed to 1011 × 578 px in the photo. Each corner sits on the card's corner, and the magnifier shows the one being placed.
Scale is measured, not eyeballed. The Align panel scans both cropped sides for where printed content begins, draws each as an outline — front in red, back in blue — and reports the difference in millimetres, per edge. If the outlines don't sit on each other, your two crops are at different scales and the print cannot line up. Match back to front corrects it in one click; the size slider and 0.25 mm nudge are there for the last fraction.
| A mis-scaled back, caught | After Match back to front |
|---|---|
![]() |
![]() |
The back here was cropped ~2.5 mm loose top and bottom — the exact mistake that started this project. The panel reports it as "off by 0.0 × 5.1 mm" and locates it: top 2.54, bottom 2.54.
The outline follows ink, not plastic, so a card that is white right up to its border reads as slightly inset. It's presented as a measurement with the numbers attached, not as ground truth.
Exact millimetres. The PDF is written by hand, byte by byte, with the card placed via an explicit transformation matrix. No library, no rounding through a rasteriser. Verified at 600 dpi: cards land within 0.04 mm of target.
- Two-sided mode — a 2-page PDF, page 2 mirrored so it registers behind page 1
- One-face mode — front and back on a single side, for offices that want it that way
- Six paper sizes — A4, US Letter, US Legal, A5, A3, or a custom sheet in millimetres
- Up to 18 cards per sheet (A3), 8 on A4, so one card never wastes a whole page
- Alignment measurement — content-edge detection with per-edge millimetre readout, a one-click match, and overlay / difference / A-B flip views
- Printer calibration sheet — one scrap page tells you your printer's flip behaviour, for good
- Four-corner crop with perspective correction: drag each corner onto the card's corner, with a magnifier while you drag. Angled or squashed photos come out exactly 85.6 × 54 mm
- Scan or phone photo — EXIF-aware, with brightness/contrast/grayscale for a clean copy
- Follows your system theme, with a toggle that persists — and releases back to automatic when your choice matches the OS again
- Nothing leaves the browser — no network calls anywhere in the source
Open index.html. That's the whole installation.
git clone https://github.com/hafeee44-lab/duplex.git
cd duplex
open index.html # or just double-click it- Capture both sides. A flatbed scan is best; a phone photo works if you crop carefully.
- Put each corner on the card's corner. Drag the four handles onto the card's actual corners, not roughly around it. If the card already fills the photo, drag each handle past the edge and it pins to the image corner.
- Check the overlay. Switch to Edges. Front is red, back is blue, shared edges go dark. If the outlines don't sit on each other, fix it with the size slider before printing.
- Calibrate once. Hit Test sheet, print page 1, reinsert the paper the way you intend to, print page 2. Hold it to the light — the boxes should overlap and both arrows point the same way. Arrows opposing → tick "back came out upside-down". Boxes not overlapping → switch the flip direction. You only ever do this once per printer.
- Print at 100%. Never "Fit to page". Margins: none. This is the one remaining way to get the wrong size.
Many cheap laser printers — the HP LaserJet M11xx/P11xx family among them — have no page description language. They can't interpret PostScript or PCL; the computer must rasterise every page and send raw bitmap data. It's why those printers often have no working driver on current macOS, and why a USB print-server dongle can't put them on your network.
None of that affects Duplex, which produces a normal PDF. It's worth knowing if you're trying to print from a phone to one of them — you'll need a computer with the driver in between.
The PDF writer (buildPdf) emits objects and computes the cross-reference table by tracking
byte offsets as it builds. Images are embedded as JPEG XObjects with DCTDecode — no
re-encoding, no decompression — and shared across both pages rather than duplicated.
Placement (cmFor) returns a 6-element transformation matrix. PDF image space puts sample
(0,0) at unit-square (0,1), so the rotation matrices are written to match what the on-screen
preview shows — an easy thing to get 180° wrong, and invisible in a duplex proof because both
sides rotate together.
The crop pipeline maps the output card onto the four crop corners with a projective transform
(squareToQuad) and fills every output pixel by inverse mapping with bilinear sampling. When the card
spans far more pixels than the 300 dpi output, it samples a smoothed, downscaled copy so fine print
doesn't alias. The Align panel's size and nudge are a similarity in output space, and folding that into
the projective map gives another projective map, so applyAdj carries them as four new corners. Anything
outside the photo stays white instead of silently distorting the card.
N-up registration only works if the set of occupied grid cells is symmetric about the page centre — otherwise mirroring puts backs where no front was printed. Column count depends on both the paper and the card orientation (A4 landscape gives 2, A4 portrait 3, A3 portrait 4), so the app recomputes the offered copy counts whenever either changes and only lists ones that fill whole rows.
Eight two-sided copies on one A4 sheet, page 2 mirrored so it registers behind page 1:
| Dark | Light |
|---|---|
![]() |
![]() |
All screenshots use a fictional specimen card generated by the test suite.
npm install playwright
node test/verify.js84 assertions. The suite drives the real UI in headless Chromium and measures the real PDF with
pdftoppm (poppler-utils) plus Pillow/NumPy. It checks:
- Registration — un-mirroring every back position lands on a front, swept across all 20 reachable combinations of orientation, flip direction and copy count
- Rotation — the PDF's placed corner matches the same corner in the on-screen preview, for all four orientation/flip states. This is the assertion that catches a transposed rotation matrix, which is otherwise invisible in a duplex proof
- Exact size — the card measures 85.6 × 54 mm at the requested millimetre position
- Every paper size — each of the six sheets is selected through the real control, then the PDF's MediaBox, the printed card size, and the registration and bounds of every offered copy count are all checked against it. A custom sheet is driven through the inputs, including an absurd value to confirm it is clamped rather than obeyed
- Bounds — no card outside the 6 mm printable margin, in any configuration
- Four-corner crop — real handle drags move only their own corner, pin to the photo edge, and can't fold the frame. A squashed back cropped by its corners measures the same as the front with no align-panel help, and a card shot at an angle fills the printed box edge to edge
- Alignment controls — the size slider changes only the targeted side, reset restores exactly, nudge steps land on 0.25 and 1 mm
- The screenshots tell the truth — the suite asserts the specimen back really is mis-scaled before capturing the "before" image, and that Match back to front really closes the gap before capturing the "after" one
- Theme resolution — a fresh profile follows the OS in both directions, a manual override
survives a reload, choosing what the OS already prefers releases the override, and everything
still works with
localStoragethrowing - Robustness — 1×1 images, non-image files, removing a side mid-edit, the flip timer stopping when the panel hides, rapid mode switching, no horizontal overflow at 390 px, all four crop corners in view when the editor opens on a phone
It also regenerates every screenshot in docs/ from that specimen card, so the images in
this README cannot drift from what the app actually does.
Any current browser. Uses createImageBitmap with EXIF orientation where available and falls
back to <img> decoding. Canvas, Pointer Events, and Blob are the only other requirements.
MIT — see LICENSE.
Designed and built by Hafi





