Cardwright

Website · Download the latest release · Guides (wiki) · Discussions

Desktop app that turns Magic, Pokémon, Yu-Gi-Oh and Riftbound card images into true 1200 DPI print-ready proxies using AI upscaling on your own GPU - free, offline after setup, no upload limits.

Unlike web-based proxy builders (which embed ~300 DPI images into a "1200 DPI" PDF), Cardwright reconstructs real detail with per-card AI model selection, then builds guillotine-ready - or cutting-machine-ready - sheets with print-shop features.

Features

Upscaling

• Real AI upscaling to 2976×4160 px (1200 DPI at 63×88 mm) on your GPU via Vulkan. No GPU? It falls back to high-quality resizing.
• Automatic per-card model choice: paper scans, digital renders and photorealistic art each get the model that suits them, with a manual override per card.
• Already high-res? The AI step is skipped automatically instead of bloating the file.
• Parallel processing, retry of failed cards, status filtering.

Getting cards in

• Type a card name and a gallery opens with every printing of it, thumbnails and all - you see what you are about to upscale before committing to it. Click a picture to see it large, and click up to three to compare them side by side.
• Switch source without leaving the gallery: Scryfall, Gatherer, MPC Autofill, Pokémon (TCGdex), Yu-Gi-Oh (YGOPRODeck) and Riftbound (Riftcodex). MPC bleed edges are trimmed automatically; picking a Yu-Gi-Oh card switches the card size to 59×86 mm. Pokémon and Riftbound cards are 63×88 mm, the same as Magic - Riftbound's landscape Battlefields included, which are turned upright to fit the sheet.
• Best scan picks the sharpest printing of a searched name for you, keeping the same artwork - a weak scan is exactly what upscaling magnifies.
• Card language: fetch printings in any of the 11 languages Scryfall carries. Applies to name lookups and decklists; cards never printed in that language come back in English and are listed, rather than failing.
• Pasting a Scryfall or Gatherer link goes straight to the queue - you already chose the printing, so nothing second-guesses it. Gatherer links use the Gatherer image, never a substitute, in the language the link names - including printings Scryfall does not carry at all.
• One exact printing, by name: type Sol Ring (SLD) 2560 and you get that art, not whichever one a bare name resolves to. Set code plus collector number, the way every deckbuilder writes it. A quantity in front, a lowercase set code and a trailing [matte] or *F* all work.
• Decklist paste: the same line, one card per row, so a list copied out of Archidekt or Moxfield goes in whole. Their deck URLs work too.
• Save and load a project: the queue itself, not just settings. A hand-built list with its quantities, chosen printings and per-card models survives closing the app. Save it after upscaling and it reopens ready to export, without paying for the AI again.
• Tokens: optionally add every token the deck's cards make - taken from Scryfall's own data, so it's exact.
• Import a decklist with the art coming from Gatherer instead of Scryfall, or load an MPC Autofill order .xml to get exactly the art you picked there.
• Card list .json: a batch of cards named by image URL, one entry each. Unlike the decklist, it is not tied to Magic - a Pokemon, Yu-Gi-Oh or Riftbound order arrives in one file instead of being typed into the gallery card by card, and each entry can carry its own back and a note (the paper, the finish) that shows up in the queue. Saying which game an entry is picks the right card back and moves the card size. See cardlist.py for the shape.
• Local files (PNG/JPG/WEBP/AVIF/…) and drag & drop.
• Double-faced cards fetch both faces and stay paired, from Scryfall and from Gatherer alike.

Editable print preview

The preview is a workspace, not a picture - what you see is what prints.

• Scroll through every sheet, not just the first.
• Drag a card to reorder it anywhere in the layout.
• Select several with click and shift+click, and the right-click menu acts on all of them at once, as one undo step.
• Right-click a card to duplicate it, change its art, remove it from the sheet, or delete its file from the output folder.
• Add cards… to pull more in mid-export - or open the dialog with nothing queued and build a sheet entirely from cards already on disk.
• It stays open after exporting, so reprinting a sheet with one card moved does not mean rebuilding the arrangement.
• Magnifier on hover to inspect detail at print resolution.
• Build a PDF straight from picked files, without using the queue.
• Print only selected sheets (e.g. 1 or 1-3,5).
• Export presets - save and reload named configurations.

Print sheets

• Layouts: 3×3 portrait, 4×2 landscape, 7-card Silhouette, plus 3×4 and 4×4 for the bigger sheets - A3 and Tabloid fit 16 cards, and 2×1 landscape for the 4x6 photo print.
• Card sizes: MTG / Pokémon (63×88), Yu-Gi-Oh (59×86), mini (44×68), tarot (70×120), plus any custom size you enter in mm. Pokémon cards are the same size as Magic cards, so they work as-is.
• Pages: A4, Letter, A3, Legal, Tabloid, A5, and 4x6 photo - two cards on a photo print, which in some countries costs a fraction of nine cards on A4 or Letter.
• Output as PDF, PNG or JPEG. Photo labs commonly refuse PDF, so the raster formats are there for them, at 300, 600 or 1200 DPI (4x6 at 300 DPI is 1800×1200). PDFs are lossless or JPEG, and split into one file per N pages; a bitmap has no pages, so each sheet is its own file.
• Cut guides with adjustable style (cross or corner crop marks), length, thickness and offset; margin ticks; optional rounded corners. Thickness shows its width in millimetres, and cannot be set thinner than an inkjet can print.
• Edge bleed with selectable colour, and a page shift on both axes, both directions - move the whole layout off the strip your printer's feed cannot use, with the preview showing how far it will actually go.

Cutting machines (Silhouette / Cricut)

• Registration marks for print & cut: 3-mark standard and 4-mark CAMEO 5a patterns, with adjustable inset, arm length and thickness.
• Marks follow Silhouette Studio's own published geometry (0.394 in inset, 0.350 in length, 0.039 in thickness), so a sheet lands on a template built in Studio.
• Those marks reach onto the grid, so on a small page they cost card slots. On A3 or Tabloid every layout keeps every slot - including 4×4, sixteen cards a sheet - which makes them the best pairings for machine cutting. Legal keeps all nine in 3×3; on A4 use 4×2 (8) or the 7-card layout (7). If a mark would sit on a card, that slot is left empty and the card moves to the next sheet, and a live hint names the mark inset that would keep them all.
• Guides and margin ticks never print inside a mark's clear area, and turning guides off removes all of them.
• 7-card Silhouette layout: one vertically centred card in the left column plus a 3×2 block, clearing both left corners where a Cameo's key marks sit (clearance goes from ~5 mm to ~49 mm on Letter).

Duplex backs

• Mirrored back pages for flip-on-long-edge printing, DFC backs paired automatically, plus a user-supplied or MPC-sourced generic back.
• Back offset and rotation to correct duplex misalignment, and back bleed to survive drift.
• Duplex alignment test: a two-page front/back registration sheet. Print it double-sided, hold it to the light, dial in offset and rotation.

Print quality

• Calibration sheet with 9 colour profiles to match your printer.
• Shadow-lift test sheet - rescues dark details crushed by inkjet plus matte lamination.
• Black border deepening for scans whose border prints as dark grey, in two flavours: Contrast edges (the default) pushes the dark pixels inside a fixed band at the card's edge, so nothing has to guess where the border is; Auto-detect measures the frame and snaps it to black. Both take a per-card override, and you choose which sources get treated at all - MPC art already carries a true black edge, so it is left alone by default.
• Output sharpening. All of it is applied at export time - your upscaled masters stay untouched.

Install (Windows)

1. Download the latest release from Releases - the installer (Cardwright_Setup-x.y.z.exe, per-user, no admin) or the portable Cardwright.exe.
2. Run it. On first launch it downloads the AI engine and models (one time) from their official sources and detects your GPU.
3. The app updates itself from within.

Upscaled cards are about 29 MB each, so a project fills a drive fast. Output folder… > Change moves them to another one; if that drive is not connected when the app starts, it falls back to its own folder and says so.

Something misbehaving? The Log button in the header opens a log file with the full details of any failure - attach it to a bug report.

If your antivirus flags it

Windows Defender may report Trojan:Win32/Wacatac.C!ml. The !ml suffix means a machine-learning guess about the file's shape, not a match against a known virus, and Wacatac is Defender's catch-all for packed executables.

Cardwright is a packed executable that legitimately does three things malware also does: it ships as one self-extracting file (PyInstaller), it downloads and runs another executable on first launch (the Real-ESRGAN engine), and it replaces its own .exe when you accept an update. All three are visible in the source in this repository.

Verify what you downloaded. Every release lists the SHA-256 of both files:

certutil -hashfile Cardwright.exe SHA256

If it matches the release page, you have exactly the file that was published. You can also build it yourself from source, below.

The real fix is a code-signing certificate, which is not yet in place. Reporting a build to Microsoft as a false positive clears only that one file, so it has to be repeated for every release.

Run from source

Runs on Windows and macOS, Python 3.10+. There is no packaged macOS build yet, so on a Mac this is the way to run it.

Windows

pip install -r requirements.txt
python main.py

macOS - Homebrew ships the Tk bindings as a separate formula per Python version (python-tk@3.13 pairs with python@3.13), and a mismatched pair fails with No module named '_tkinter'. Install a matching pair and use that interpreter the whole way through:

brew install python@3.13 python-tk@3.13
python3.13 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python main.py

First launch downloads the macOS build of the AI engine and detects the GPU exactly as it does on Windows - Apple Silicon is native, reaching the GPU through Metal rather than Vulkan. The in-app updater stays Windows-only, since releases only carry Windows builds.

Licence

Cardwright is free to use but it is not open source. The code is published so anyone can read and audit what the app does - you are welcome to study it, build it for your own use, and send bug reports or pull requests. You may not redistribute it, publish a modified or rebranded version, or sell it. See LICENSE.

Legal

• This is an unofficial Fan Content project, not affiliated with or endorsed by Wizards of the Coast, Nintendo, The Pokémon Company, Konami or Riot Games. It ships no publisher assets; card images are fetched at the user's request from public APIs.
• Card data and images courtesy of Scryfall; this app is not affiliated with Scryfall.
• Yu-Gi-Oh card data and images courtesy of YGOPRODeck; this app is not affiliated with them. Images are downloaded once to your machine rather than hotlinked, per their API terms.
• Pokémon card data and images courtesy of TCGdex; this app is not affiliated with them. Images are downloaded to your machine rather than hotlinked.
• Riftbound card data courtesy of Riftcodex, an unofficial fan project; card images are Riot Games' own. This app is not affiliated with either. Images are downloaded to your machine rather than hotlinked.
• AI engine: Real-ESRGAN (BSD-3). Community models UltraSharp (Kim2091) and High Fidelity are fetched from the Upscayl project and carry non-commercial licenses.
• Registration-mark geometry follows the spec used by silhouette-card-maker (MIT); the 7-card arrangement matches the "SevenCard" template from ProxySheet (GPL-3.0). No code from either project is used - only the published geometry.
• The contrast-edges border treatment is a reimplementation of the approach used by Proxxied (MIT). Their code is GLSL; this is an independent numpy implementation.
• Intended for personal playtesting. You are responsible for how you use the output.

Community

Questions, half-formed ideas and "is this supposed to work like this?" go in Discussions. Bugs go in Issues, where the forms ask for what usually has to be guessed at - your printer and its tray for a print problem, the log for anything else. The wiki has the longer guides.

Support

Free for the community. If Cardwright saves you time, donations help keep it maintained: