Geometry

Every channel wants a product in its own shape: a marketplace a white square with the product filling most of it, a story a tall 9:16 frame, a newsletter a strip of 600 by 200, a link preview 1200 by 628. The same photo cannot go everywhere as it is, and each change of shape asks one question: when the picture and the frame do not have the same proportions, what gives way — part of the picture, the proportions, or empty space around it?

The geometry nodes answer that question explicitly, one mode at a time. They run on Madoo’s workers, cost no credits and give the same result every time. This page explains the modes, the nodes that cut, frame and turn a picture, and the units they accept; the lab makes one product ready for five channels and shows what each choice keeps and what it loses. When the missing space must be invented rather than filled with a colour — a landscape from a portrait photo — that is a job for the AI nodes described in AI edits and utilities.

Resize: four ways to meet a frame

Resize (image/resize) takes a width, a height or both. With one of them it scales in proportion. With both, its mode decides what happens when the proportions differ:

mode	Size of the result	What gives way
fit (default)	Inside the box: 1600×1600 on a landscape photo gives 1600×1195	Nothing — but the result is not the box size
fill	Exactly the box	The parts outside it, cut from the centre
stretch	Exactly the box	The proportions: the picture is distorted
pad	Exactly the box	Nothing is cut: the picture fits, and bands of backgroundColor fill the rest (transparent by default, which makes a PNG)

Two rules apply to every mode. The crop of fill is always centred — it does not know where the product is; if the subject is off-centre, crop it yourself first. And a picture smaller than the box is enlarged, which blurs it: to make a small photo bigger with real detail, use Upscale (AI, paid) before resizing.

Turn on Never Enlarge (never_enlarge) to only ever shrink: a picture that already fits is returned exactly as it was stored — not re-encoded, so a JPEG loses nothing — and in the other modes nothing is scaled up (fill still cuts to the box where the picture is larger, pad still adds the canvas). It is the setting for a batch of mixed sizes: the large photos come down to the size, the small ones stay sharp.

Platform sizes

Resize for Platform (image/resize_for_platform) knows ten sizes and writes the result as JPEG, PNG or WebP (format, quality):

platform	Size		platform	Size
instagram_square	1080 × 1080		linkedin	1200 × 627
instagram_story	1080 × 1920		banner_web	1920 × 480
tiktok	1080 × 1920		billboard	1920 × 1080
facebook_post	1200 × 630		email_header	600 × 200
twitter_post	1200 × 675		pinterest	1000 × 1500

It always works like fill: exact size, centre crop. That is right for a photo with room around the subject, and wrong for a product that fills its frame — in the lab the story cuts the teapot’s spout and the email header cuts its top and bottom. When nothing may be lost, use Resize with mode: pad and the platform’s size, or Place on Canvas below.

Cut

Node	What it does	Good to know
Crop (image/crop)	Keeps the rectangle x, y, width, height — or the one wired into Region	In pixels of the upright picture, from the top-left corner. Connect Region to cut what another step found: the measurements of Measure Subject (the subject’s box), the best_region of Image Detail (the area to look at closely), the instances of Instance Segment (the box of the first selected person or object), the crop_region of Face Detect (a head shot framed around a face, in the shape you choose), or any {x, y, width, height}. With Region, a rectangle beyond the picture fails the step (REGION_OUT_OF_BOUNDS), or is cut to the picture with On Out Of Bounds clip; the output region says what was cut. To compute a rectangle from the picture’s size, read it with Image Info and calculate
Auto Crop (image/auto_crop)	Trims a uniform border	With autoDetect the border colour is the top-left pixel — transparency on a cut-out, white on a studio shot; tolerance (0–100 %) accepts near colours; margin adds space back in marginColor

Frame

Node	What it does	Good to know
Place on Canvas (image/place_on_canvas)	Fits the picture, centred, on a canvas of canvasWidth × canvasHeight in backgroundColor, with padding	Padding in pixels is per side; in percent it is the total per axis — 16 leaves 8 % on each side, so the product fills 84 % of the width or of the height. paddingTop/Right/Bottom/Left override one side. Always a PNG
Extend Canvas (image/extend_canvas)	Adds top, right, bottom, left space without resizing	In % each side is a share of the picture’s own width or height. Transparent by default, which makes a PNG
Round Corners (image/round_corners)	Rounds the corners by radius	In % of the shorter side, at most half of it. Transparent corners make a PNG; give a backgroundColor to keep the format

A marketplace listing is a two-step recipe: Auto Crop to the product’s edges, then Place on Canvas on white with a percentage padding. Because the product is trimmed first, the margin is the same whatever empty space the supplier left around it.

Turn and straighten

Node	What it does	Good to know
Rotate (image/rotate)	Turns by angle degrees	The canvas grows to hold the turned picture; the corners are filled with backgroundColor. Phone photos are already turned upright when any image node reads them, so autoOrient is not needed
Flip (image/flip)	Mirrors horizontal, vertical or both	Mirroring a product with text or a logo reverses them
Deskew (image/deskew)	Straightens a small tilt — a scanned page, a photographed label	Automatic; threshold (0–100 %) sets how strongly it reacts
Perspective (image/perspective)	Moves the four corners to new places	Each corner is a fraction of the width and height (0–1); the defaults leave the picture unchanged

Units

Resize, Extend Canvas, Place on Canvas and Round Corners accept px, %, in, cm and mm (unit, canvasUnit, paddingUnit). Inches, centimetres and millimetres are turned into pixels with the density written in the image, or 72 DPI when it has none — so “10 cm” is 283 pixels on a web image and 1181 on a 300-DPI scan. When a print size in centimetres must come out at a given resolution, Perfect Size (ai/perfect_size) takes the size and the DPI together.

The lab

The geometry lab makes one product ready for five channels, from two files: the cut-out vase of Image files and formats and a product photo of a red teapot on a grey studio background, 1286×1600.

• Trim the empty border takes the vase from 1280×1600 to its edges, 1001×1297.
• Marketplace square: 2000×2000, white, padding: 16 %: the vase fills 84 % of the height, whatever empty space the cut-out had. Rounded tile rounds it by 6 % into a PNG with transparent corners.
• Email header, padded: 600×200 in the brand’s cream, the whole vase — small, because a tall product in a wide strip always is.
• The story and email header presets crop the teapot photo from the centre: the spout goes in the first, the top and bottom in the second.
• The link preview twice at 1200×628: fill crops the photo to the frame; pad keeps it all and adds grey bands on the sides — visible, because a flat colour does not match a studio gradient.

The run costs 0 credits. When the bands of a padded picture must look like the photo rather than a flat colour, the space has to be generated: AI edits and utilities extends a picture with outpainting.

node madoo-install-example.mjs lab-image-geometry --api <your Madoo API URL> --key <key.json> --run --save channels

Limits

• Every resize mode enlarges pictures smaller than the target, without adding detail, unless Never Enlarge is on.
• The crops of fill and of the platform sizes are always centred; there is no gravity or subject detection. To crop around the subject, wire Measure Subject’s measurements into Crop’s Region.
• Place on Canvas always centres; for another position, extend the canvas side by side or use Overlay on a canvas of your own.
• Platform sizes are the ten above; any other size is Resize with a width, a height and a mode.

Photos: “Vase”, Louis C. Tiffany, 1893–96, Favrile glass, and “Teapot”, George E. Ohr, about 1897–1900, earthenware, The Metropolitan Museum of Art, New York (Open Access, public domain).