Photos in. Out comes the object: its 3D model and its metadata.

thicken turns photos of a real object into a GLB model at its real size and its metadata as JSON, read off the object by vision models and looked up in public catalogs. What it knows about each kind of object lives in a profile: a few come with thicken, and you can add your own.

Photos of the front and back of a CGC-graded Giannis Antetokounmpo autograph card on a wooden table, beside the clear 3D slab thicken made from them, turning all the way round

_{In each, the photos going in and the model coming out, with its metadata, like the one below. The logo above is a thicken model too, made by its lettering profile: thicken lettering thicken.}

The other half: its metadata. Part of the first one's, a graded Giannis Antetokounmpo card:

json

{
  "profile": "slab",
  "holder": "cgc",
  "grader": "cgc",
  "cert": "6072896005",
  "widthMm": 80.5, "heightMm": 135, "depthMm": 7.5,
  "read": {
    "grade": "9", "gradeName": "MINT",
    "autoGrade": "10", "autoGradeName": "AUTO",
    "signed": "G. ANTETOKOUNMPO, 8/2/25",
    "name": "Giannis Antetokounmpo", "year": "2024-25", "maker": "Panini",
    "set": "Court Kings Basketball", "number": "#5", "variant": "Blank Slate"
  },
  "readBy": { "grade": "agreed", "autoGrade": "agreed", "set": "disputed" },
  "inside": {
    "card": { "widthMm": 63.5, "heightMm": 88.9, "corners": "square", "foil": "none" },
    "label": { "widthMm": 66.4, "heightMm": 26.9 }
  }
}

The grader and certificate number come from the label's QR code; the vision models read the rest off the slab. What each profile records is in the table of profiles, and every field is described in USAGE.md.

Try it · From an AI agent · How it works · The thicken object · What it makes
In detail (USAGE.md): batch · photos · vision · video · training mode · review · the sidecar · previews · inputs · sizes
What leaves your machine · Adding a profile · Security · Built on · Development · License

Try it

npm install -g thicken                      # Node 22.18 or later; or npx thicken …
thicken photo.jpg                           # it works out what it is: a code on it, or the vision models
thicken book cover.heic back.heic spine.heic
thicken --isbn 9780061120084                # no photo: the cover and size from Open Library
thicken slab front.jpg back.jpg
thicken card front.jpg back.jpg --pt 130    # a thick patch card
thicken badge badge.png --width 50          # cut-out artwork, extruded
thicken lettering thicken                   # the logo
thicken batch ~/Pictures/books --preview    # a model per photo, PDF, video or folder

Each run writes <name>.glb and <name>.json beside the first picture, and with --preview a picture of both sides. Photos straight on and cropped work as they are; uncropped photos, angled shots and video need a vision roster, any OpenAI-compatible endpoint (Vision). As a library: import { thicken } from 'thicken', then thicken(files, options) gives the GLB's bytes and the sidecar.

From an AI agent

thicken is an MCP server too (thicken mcp), with an Agent Skill that tells the agent when and how to use it: ask for a model of the book on your desk, and the agent makes it. In Claude Code:

/plugin marketplace add bikemeardsley/thicken
/plugin install thicken@thicken

Or for any MCP client, npx -y thicken mcp on stdio. Codex, Copilot, Cursor, Gemini CLI and the rest are in USAGE.md.

How it works

Photos, a video, words or an ISBN go in; thicken matches them to a profile, processes them (the vision models find each face, cut it out and straighten it, and read what's printed), builds the object at its real size, and writes a GLB model and a JSON sidecar.

What thicken knows about a kind of object lives in its profile: how to recognize it, its faces, what to read off it, and how to build it at its real size. A code on the object picks the profile (an ISBN barcode, a grader's QR code), or the vision models say which it is. thicken is not a photogrammetry tool: everything it handles is close to a box or a flat outline, so it builds the geometry from known sizes and maps your photos onto it.

The thicken object

What thicken makes is more than a 3D shape. Each run writes a thicken object: the model (a GLB at the object's real size, in millimeters) and, beside it, its metadata (the JSON sidecar), which knows what the object is, in its profile's own terms. For the profiles that come with thicken, for example:

A book: its ISBN, its catalog record (title, authors, publishers, dates, subjects, Dewey and Library of Congress classes, pages), and how sure the match is when it was found by title.
A graded slab: its grader, cert number and verification link, its holder, and the card and label inside it, each sized and placed. With vision, the label read: the card's grade (its number and words, "9 MINT", any qualifier and subgrades), an autograph's own grade when the label grades one apart from the card ("AUTO 10"), who signed it and when, and the card as the grader names it.
A raw card: its size, its thickness in points for a thick card, its corners and its foil. With vision, what's printed on it: player or character, team, year, maker, set, number, parallel.
A box: with vision, its name, brand, edition, rating and year.
A badge: with vision, what it names, who issued it, its level, subject, date, any credential number, and every other word on it.
Every object: what the vision models read off it, and how sure; which faces are your photos and which are stand-ins, how each fact was found, what is still worth photographing, and what each vision model answered.

An app shows the model and reads the sidecar: a library shelf of your books, a vault of your slabs. The fields are under What the sidecar records.

What it makes

thicken makes the 3D object; what it knows about kinds of object lives in profiles, one file each in src/profiles. Anyone can add one (CONTRIBUTING.md). Every profile there today:

Profile	What it is	From	Makes	With vision, reads
`badge`	cut-out digital artwork: a certification badge, logo or sticker design	Artwork with transparency around its shape (`--width`, `--depth` in mm)	The artwork's own outline, extruded: the art on its face, its edge color on the sides and back	name, issuer, level, subject, date, id, text
`book`	a book, hardcover or paperback	Front cover, back, spine and page sides, in any order and any of them; or just an ISBN	The book at its real size, from its ISBN's catalog record: a paperback softly rounded, a hardcover with boards standing past its pages	title, subtitle, authors, series, publisher, edition
`card`	a single trading card on its own, not sealed in a plastic holder	Front, and its back	A thin plate, 63.5 × 88.9 mm unless `--size` says, 0.35 mm unless `--pt` gives a thick card's points; its corners cut and its foil shimmering as the vision models see them	name, team, year, maker, set, number, variant
`lettering`	words made into puffy 3D bubble letters	Words, typed where pictures would go (`--width` mm across, `--depth` mm thick)	Puffy gold bubble letters, round in section like a balloon, in the Titan One font	–
`slab`	a graded trading card sealed in a clear plastic holder with a label	Front, and its back	A clear plastic holder at its grader's size, with the card and the grade label inside it, each cut from the photos	grader, grade, gradeName, qualifier, subgrades, autoGrade, autoGradeName, autograph, signed, cert, name, team, year, maker, set, number, variant
`box`	a box-shaped object (a game case, a phone, a deck of cards, packaging)	Front, and any of its back and sides; its size (`--size`), or the vision models' estimate	A softly rounded box with the photos on its faces	title, brand, edition, rating, year

What each one records in its sidecar, beside what every model has (its size and where it came from, its faces and pictures, its codes, what the vision models read and how sure, what's worth a second look: all in USAGE.md):

badge: edge, regions

Field	What it holds
`edge`	the color just inside its outline, used for its sides and back (#rrggbb)
`regions`	how many separate shapes the artwork is (a medal and its ribbon are two)

book: isbn, catalog, match, pages, binding, bindingBy, measuredBy, depthEstimates, coverUrl, sticker, lookups

Field	What it holds
`isbn`	its ISBN: given, read off its barcode, or from a title match with one edition
`catalog`	Open Library's record of the edition and its work: title, authors, publishers, dates, edition, format, series, languages, subjects, Dewey and LC classes, contents, every id, links, cover
`match`	when found by the title read off it: what was read, how sure the match is, the work and its editions
`pages`	its page count
`binding`	hardcover or paperback
`bindingBy`	how that was found: vision, catalog or assumed
`measuredBy`	where each of its sizes came from: measured, openlibrary, loc, googlebooks, pages, photos, cover or trim
`depthEstimates`	every thickness estimate made: from the photos and from the page count
`coverUrl`	Open Library's cover, when it stands in for a front not photographed
`sticker`	where a foil seal or sticker is on its front cover
`lookups`	every catalog request: host, path, status and time

card: corners, cornersBy, foil, foilBy, pictureBox, pictureBoxBy, pt, measuredBy

Field	What it holds
`corners`	how its corners are cut: square, slight or round
`cornersBy`	how that was found: vision or assumed
`foil`	its foil: none, holo (the picture shines), reverse (all but the picture) or full
`foilBy`	how that was found: vision or assumed
`pictureBox`	where its picture box is, in fractions of the front, for holo and reverse foil
`pictureBoxBy`	how that was found: vision or the standard layout
`pt`	a thick card's thickness in points, when given
`measuredBy`	where its size came from: standard or given; its thickness: standard, given or pt

lettering: text, font, fontLicense, color, measuredBy

Field	What it holds
`text`	the words, as lettered
`font`	the font they were drawn in
`fontLicense`	its license (OFL-1.1)
`color`	their color
`measuredBy`	where its width and depth came from: given, the default, or round in section

slab: grader, cert, verifyUrl, holder, inside, measuredBy

Field	What it holds
`grader`	who graded it: psa, bgs, cgc, sgc or tag, from its QR code or its label
`cert`	its certification number
`verifyUrl`	the grader's page for this cert, from its QR code
`holder`	the holder it was built as (psa, cgc-tall…)
`inside`	the card and the label inside: each one's size, middle and what placed it; the card's corners and foil; where the label's hologram is
`measuredBy`	where its size came from: the holder table, or measured

box: barcode, measuredBy

Field	What it holds
`barcode`	its product barcode (UPC or EAN) and its kind, when a photo shows one
`measuredBy`	where each of its sizes came from: measured, or the vision models' estimate

Name the profile, or leave it out and thicken picks one:

An ISBN barcode on any picture, or --isbn, means a book.
A QR code pointing at a grader's site (psacard.com, beckett.com, cgccards.com, gosgc.com, taggrading.com) means a slab, and says whose. The sidecar keeps the grader, the cert number from the link, and the link itself (verifyUrl).
So does an older label's Code 128 cert barcode, but only on a photo shaped like a holder. The grader is then the one the vision models read off the label; with no vision, a PSA-size holder is assumed, with a warning.
Artwork with transparency around a shape that isn't just a rectangle means a badge.
Otherwise, with vision, the models are shown the profiles and say which it is, and roughly how big: two agreeing settle it, as with faces, and box takes anything box-shaped the others don't cover.
Anything still unknown is refused with a request to name the profile. Other barcodes don't count: books carry UPCs, QR codes and library stickers.

The sidecar's identifiedBy.profile says which way it was chosen: given, recognized (codes or the artwork), or vision.

Adding a profile is one file in src/profiles, found by itself, nothing else to change: its name; a few words on what it is and on each of its faces, which is all the vision models are told; what to read off it (more is better: the sidecar is for apps to take what they want from); optionally how its codes or pictures give it away unaided; and how to build it from the photos placed on its faces. Generic profiles are named for the shape they build (box, and later perhaps flat or cylinder), not "unknown": a shape thicken can't build is refused rather than forced into one it can. Contributions welcome: CONTRIBUTING.md says how.

Pictures are placed on faces by what they show (see Photos). With no front, Open Library's cover stands in, marked for review, so one photo of the back is enough to start a book.

What leaves your machine

With THICKEN_VISION_URL set, the photos themselves (re-encoded, so without their EXIF) go to that host. A LAN server you control still forwards each photo to the model's provider; a hosted endpoint sees every photo. The only other calls out are a book's catalog lookups (Open Library, the Library of Congress, Google Books), which send its ISBN, or the title and author the models read off it; never a photo.

Nothing else is sent anywhere, and nothing is fetched at run time but those lookups: previews draw with model-viewer served from your own disk.

Adding a profile

A profile is one file, found by itself: what the object is, its faces, what to read off it, and how to build it. Make your own while you use thicken, or improve one of its own, and offer it back:

thicken profile new vinyl          # from the template, in ~/.thicken/profiles: used at once
thicken profile check vinyl        # what thicken's CI will check
thicken profile contribute vinyl   # forks, branches, checks and opens the pull request

The template, templates/profile.ts, works as it is (a box at the size you give) and says on every field what goes there, the records it writes included. CONTRIBUTING.md has the rest. The table above is made from the profiles themselves.

Security

Report a vulnerability privately, as SECURITY.md says. It also lists what thicken touches: files, the network, and the programs it runs.

Built on

Part	How	Choice
Triangulating badge outlines	Dependency	mapbox/earcut (ISC)
Writing GLB	Dependency	donmccurdy/glTF-Transform (MIT)
Decoding HEIC	Dependency	catdad-experiments/libheif-js (LGPL-3.0, used unmodified). sharp's binaries leave out HEVC.
Reading barcodes	Dependency	Sec-ant/zxing-wasm (MIT). Its `.wasm` is loaded from `node_modules`, not its default CDN.
Pictures from scanner PDFs	Dependency	unjs/unpdf (MIT, pdf.js inside)
Frames from video	Outside program	FFmpeg, run from the `PATH`, never bundled: it is GPL, thicken MIT. eugeneware/ffmpeg-static (GPL-3.0) would put a GPL binary in every install, video or not; ffmpegwasm/ffmpeg.wasm is built for browsers. Choosing the frames is ours.
Reading camera details (EXIF)	Dependency	devongovett/exif-reader (MIT): parses the EXIF block sharp returns
Drawing previews	Dependency	google/model-viewer (Apache-2.0): frames and lights a GLB; three.js inside
Running the preview browser	Dependency	microsoft/playwright (`playwright-core`, Apache-2.0): drives an installed Chrome or Edge, no browser download
Book dimension sources	Write	Read palmerabollo/node-isbn's provider fallback. Ours is per dimension, not first-provider-wins. The Library of Congress's one field (MARC 300) is read with a regular expression, not a MARC library.
Hardcovers	Write	Four rounded boxes from `src/glb.ts`, and the binding asked of the vision models like a card's corners.
Holofoil	Write	Which areas shine for each kind of foil read from simeydotme/pokemon-cards-css (GPL-3.0: facts read, nothing copied); the film values are `KHR_materials_iridescence`'s defaults, as KhronosGroup/glTF-Sample-Assets uses them; glTF-Transform, already a dependency, writes the extension.
Profiles, box, classifier, CLI	Write
Slab holder, card and label	Write	glTF-Transform's materials (already a dependency) for the clear holder; the card and label plates, where they sit, and putting them together are ours. Khronos's glTF-Sample-Assets were the reference for clear materials.
Finding faces in photos	Write	Vision models over any OpenAI-compatible endpoint, primary and peer (src/vision.ts): one POST each, no SDK.
Straightening faces	Adapt, don't depend	jlouthan/perspective-transform (MIT): the 4-point homography; the resampling over sharp's pixels is ours. sharp has no perspective warp, and OpenCV.js is megabytes for one.
Snapping corners to edges	Write	`snap` in src/image.ts moves a model's corners (about 1% off) onto the real edges. It samples brightness steps along each side, takes the line they support nearest the model's, and fits it by least squares, in about 70 ms a face. puffinsoft/jscanify's corner ranking was read; a line fit near the model's corners was enough.
Lettering's font	Dependency	opentypejs/opentype.js (MIT) reads it: the words' outlines, kerned. The font is Titan One by Rodrigo Fuenzalida, shipped unmodified under the SIL Open Font License 1.1 (`fonts/OFL.txt`; "Titan" is its Reserved Font Name).
Inflating letters	Adapt, don't depend	Each pixel's distance from the outline is mapbox/tiny-sdf's (BSD-2-Clause) exact distance transform, ported; it places the outline to a fraction of a pixel. The rise is ours: √u, where ∇²u = −2 inside the letters, round in section across any stroke.

For refractor and holo effects in a viewer, simeydotme/pokemon-cards-css is a reference only. It is GPL-3.0, so read it and don't copy it.

Development

TypeScript on Node 22.18+, with sharp for images. From a clone it runs directly (Node strips the types). The npm package is JavaScript built into dist (npm run build, run by npm pack and npm publish), because Node won't strip types inside node_modules. Every push installs the packed package on Windows, macOS and Linux and runs it.

From a clone: npm install, then node src/cli.ts … in place of thicken …; npm test, npm run check, and npm run readme after changing a profile.