Files
Dylan M. Taylor 0984f412a6 craft-affinity: bounded reader for native Affinity documents, from VectorCraft and PhotoCraft (#914)
* craft-affinity: bounded reader for native Affinity documents, moved from VectorCraft

The reader of .af, .afdesign, .afphoto and .afpub files (archive, object stream and document
model) as it is in storytold/vectorcraft at 6e873b80 (crates/affinity), under a neutral name.
Its corpus tests read the pinned public documents scripts/fetch-affinity-corpus.sh fetches.

* craft-affinity: fold in what only PhotoCraft's copy of the reader had

From storytold/photocraft at 63b641b5 (crates/affinity), onto the VectorCraft copy:

- Special shapes: clouds, hearts, cogs, callouts, arrows, double stars, tears, crescents,
  diamonds and segments come in as their own outlines, not their bounding ellipse.
- Null slots of object lists are charged to the document's work before a list is scanned,
  on every visit of a shared object (children, attachments, effects, compound operands,
  page lists, paint lists, gradient colours, text blocks and runs, table cells, tile lists).
- A colour or gradient position record is decoded only as far as its components are read.
- 16-bit and 8-bit planes are interleaved through tables, not an f64 for each pixel.
- Limits::max_pixel_bytes: the caller sets what the pixel layers of one import may hold.
- Pixels::cmyka8_to_rgba8 for a caller without colour management.

The crate README and docs describe the model on its own terms.

* craft-affinity: a callout with a right-to-left box no longer panics; an oversized shape imports as its box

Ported from storytold/photocraft#10473 by DJHoltkamp. A callout rectangle clamped its tail
with clamp(x0, x1), which panics when the box is stored right to left. A box with finite
corners whose size overflows gave infinite or NaN outline points; it now imports as its box
with a warning. Also drops the inflated doc.dat once it is parsed.
2026-11-11 17:42:11 -05:00

14 KiB
Raw Permalink Blame History

craft-affinity

An independent, bounded Rust reader for native Affinity documents: .af from Affinity 3 and .afdesign, .afphoto and .afpub from Affinity 1 and 2 (container versions 8 to 12). It reads the archive, the tagged object stream in doc.dat and the part of the document model listed below, and returns a neutral Document in pixels with every transform composed. Mapping that document onto an application's own model stays in the application. It does not write Affinity files.

match craft_affinity::read(bytes, craft_affinity::Limits::default()) {
    Ok(document) => { /* spreads, nodes, warnings */ }
    // The flattened PNG preview every document embeds, when the native document can't be read.
    Err(reason) => { let preview = craft_affinity::preview(bytes); }
}

is_affinity recognises the container by its signature. Document::warnings lists what was left out or approximated, one line per kind with a count.

What is read

Affinity Model
Pages, Publisher spreads (one or two pages) Document::spreads, each with its page rectangles
Master pages (MpCh) shown on pages through master page instances (MPIN: master, page offset, master page, count) Document::masters and Kind::MasterInstance nodes
Artboards (legacy ABEn or Affinity 3 phrp/aprp), with name, paint and any transform Kind::Artboard: export rectangle, the actual clipping outline, background and strokes
Layers (Scop), groups (Grup), pass-through or isolated Kind::Layer, Kind::Group, Blend::PassThrough
Curves (PCrv): cubic subpaths, closed flags, live corners (CnrD) Kind::Shape paths (multi-subpath curves fill even-odd)
Rectangles with corner radii (relative or absolute, locked or per corner) and round or straight corners, ellipses, polygons (smooth too), stars, square stars, pies, triangles, trapezoids, diamonds Kind::Shape paths
Clouds, hearts, cogs (with their hole), rectangular and elliptical callouts, arrows, double stars, tears, crescents and segments Kind::Shape paths drawn from their parameters (approximate outlines, not Affinity's exact curves)
Compound shapes (Comp, add and subtract) one even-odd path
A shape, curve or text with children the node keeps its children: it clips them
Vector masks (AdCh curves and shapes) Node::mask
Pixel masks (MRst, attached or as a layer in a group) Node::pixel_mask
Solid fills: RGBA, HSLA, CMYK, grey, Lab (D50) Paint::Solid in the colour's own model
Linear, elliptical and radial gradients with Affinity's midpoint bias Paint::Gradient (each stop's bias is the midpoint where the blend is half way)
Strokes: weight, scale with object, dash pattern and phase, alignment, miter limit, behind the fill; a stroke whose style is None (every shape stores one, black and 1 px) is left out paint::Stroke
A stroke's pressure curve (LDeP: position along the path against width) Stroke::profile, the curve sampled at 2 to 8 points per segment (256 segments at most), widths up to the full weight
Fill layers (FRst) Kind::Fill
Artistic and frame text: Unicode scalar-indexed runs, stored fallback font, family/weight/style, boolean OpenType features, all caps, underline, strikethrough, size, tracking, horizontal scale, fixed leading, colour, outlined characters, first baseline; each paragraph's alignment, indents, space before and after, fixed and automatic leading; frame text sized by its text scale (FTxS), not its frame's transform Kind::Text: runs with their character and paragraph formats, anchor, frame and the transforms of the node and of its type
Tables (TxtT): grid, cell text, merged cells, cell insets and vertical alignment, the line on every cell edge Kind::Table
Placed images (ImgN): the original JPEG/PNG embedded in the file Pixels::Encoded
Pixel layers (Rstr): RGBA 8/16-bit and CMYK 8-bit tiles, cropped to their content; source-backed RGBA8 JPEG/PNG tiles with matching dimensions and zero origin; the document's CMYK colour profile (CMYP) Pixels::Rgba8; CMYK layers as Pixels::Cmyka8 (inks kept) with Document::cmyk_profile to show them through, or Pixels::cmyka8_to_rgba8 without colour management
Embedded Affinity documents (EmbN with an EmDc entry), four deep at most their first page's content as the node's children (their artboards only clip; their master pages are left out); other embedded documents and symbols as the picture Affinity cached of them
Layer effects: drop shadows, outer glows, inner shadows; outlines; bevels; colour and gradient overlays in any blend mode Node::effects; overlay::apply bakes colour and gradient overlays into fill, stroke, text colours and pixels for a caller without such an effect
Opacity, visibility, lock, names, blend modes the same on every Node

Everything else is reported in the warnings, never dropped silently: blur and other layer effects, adjustment layers and live filters, brush strokes, pressure curves that can't be read or that go past the million points one document may hold, transparency gradients, fill opacity, bitmap fills, master pages based on other master pages, conical gradients (as radial), parametric shapes of an unknown class (as their bounding ellipse), corner types other than round or straight, stars with rounded points, several fills or strokes on one object (the active one is used), skewed or flipped text, characters of different transparency in one text, text fields such as page numbers, text threading, frames with columns or a curved outline, table cell fills, unsupported font selectors, grey/Lab/32-bit pixels and blend modes outside Blend.

Where it came from

The reader was written in VectorCraft and PhotoCraft from one common start (AF.1 to AF.3 by callmeBalloch, 2026-10) and then grew apart in both. This crate is VectorCraft's copy (crates/affinity at 6e873b80: text and paragraph formats, tables, master pages, effects and overlays, embedded documents, CMYK and large pixel layers, pressure curves, one budget for the whole read) with what only PhotoCraft's copy had (crates/affinity at 63b641b5): the special shapes, null array slots charged before a list is scanned, colour records decoded only as far as they are read, the table-based 16-bit plane conversion, and a caller-set pixel allowance (Limits::max_pixel_bytes).

Provenance and clean-room

Affinity has no published file-format specification, so this reader is built only from a public description of the format and from public files, never from Affinity itself.

  • No Affinity software. Affinity was never downloaded, installed, run, scripted, screenshotted or disassembled for this work, and nothing from an Affinity installation (program code, resources, presets, fonts, colour profiles) was read. No file was made with Affinity for it: every sample is a document its author published.
  • A public, permissively licensed description. The container and object-stream layout was learned from VMDevCpp/afread (MIT, at 04b672334a43e3e37ded6b5ffc57af231d589774, written for container versions 7–11) and re-described in our own words before this Rust code was written; no code was translated. No GPL/AGPL code (such as Inkscape's Affinity extension) was read.
  • Every structure checked against public files. afread doesn't say how its author learned the format, so nothing here rests on it alone: each structure the reader relies on is confirmed by the public documents. Every archive entry carries a CRC-32 that must match, and every field in the object stream names its own type, so a misread layout fails loudly instead of producing plausible wrong data.
  • Meaning fitted to Affinity's own pictures. What shapes, paints, text and pixel data mean was worked out by comparing our render with the thumbnail every Affinity document embeds (Affinity's render of itself), on 176 public documents saved by Affinity 1.x, 2.x and 3.0/3.1 on Windows, macOS and iPad. The collection has since grown to 189 distinct documents, and the reader opens all of them. Most are published under CC0, MIT, BSD, Apache-2.0 or CC BY (-SA); the others (no licence stated, or GPL or non-commercial terms) were only opened locally to compare pictures, never committed or redistributed. The corpus test keeps the comparison for the pinned files, which are CC0, MIT or Apache-2.0 only.
  • Nothing of Affinity's is in this repository: no program code, assets or documents. The test files are fetched at pinned commits and sha256-verified, never committed.
  • Read-only. Nothing is written in Affinity's format; export waits until someone can check written files in Affinity itself.

Tables, paragraph formats, text scale, master pages, layer effects, overlays, embedded documents, large pixel layers and HEIC photos were added in DesignCraft (storytold/designcraft#291) and brought into VectorCraft's copy, from the same public fixtures and the thumbnails of documents their owners made, used only locally. Automatic leading (the font size plus the font's line gap, for artistic and frame text) was fitted on the thumbnails of one such local document with 56 artboards and on the public Patchy text-frame fixture.

Validation

  • Unit tests on synthetic containers built by the synth feature (stored, zlib and zstd entries, checksums, budgets, cycles, hostile streams, every truncation) and property tests of random mutations. tests/it/import_limits.rs and tests/it/outline_limits.rs build hostile documents against the limits below: grids of 16 million cells, grids and stories shared by many layers, 200 000 paragraphs, a document embedded hundreds of times, compounds that fan out or contain themselves, lists of a million null slots linked many times.

  • scripts/fetch-affinity-corpus.sh fetches 51 public CC0/MIT/Apache-2.0 documents at pinned commits, SHA-256 verified against tests/support/corpus.tsv (which names each file's source and licence), with the upstream licence notices. Of these, 33 are current .af: four CC0 vector illustrations and 28 original Affinity 3.2.3 feature probes plus one separately identified derived regression input from SethRobinson/Patchy (MIT). Only document data and provenance notices were inspected, never that project's importer or authoring scripts. The files are never committed.

    scripts/fetch-affinity-corpus.sh
    AFFINITY_CORPUS=target/fixtures/affinity cargo test -p craft-affinity --all-features
    

    Without AFFINITY_CORPUS the corpus tests are skipped. AFFINITY_SAMPLES names any other folder of documents to parse.

  • tests/it/real_files.rs parses every file and freezes the modelled features of all 33 current .af files (tests/support/affinity3-signatures.txt: counts, fonts, artboards, warnings and a CRC of the whole model). Native raster tests compare decoded pixels with independent embedded images; text tests assert Unicode run boundaries, fallback fonts and feature retention.

  • fuzz/ has cargo-fuzz targets for the whole reader (container) and the object stream (stream):

    cd crates/affinity
    cargo +nightly fuzz run stream -- -max_total_time=600
    

What has not been verified: files from Affinity builds or platforms outside the public set, files written by other applications, rotated or skewed images against Affinity's render, mask polarity against a render with and without the mask, the special shapes against Affinity's own outlines, and anything listed as a warning above. The synth builders write only what this reader accepts; Affinity has never opened their output, so they are test fixtures, not an exporter.

Safety limits

Every size is checked before it is allocated: 256 MiB per archive entry and 1 GiB per import by default (Limits), a 64 MiB zstd window, 4096 saved revisions, array lengths no longer than the bytes left, 16 777 216 decoded values per document stream (fields and array elements together, counted before anything is allocated: a value takes about 40 bytes in memory however few it took in the file; the largest of those 189 public documents uses 3.3 million), 384 levels of object nesting, 128 levels of layers, 500 000 layers, four million curve nodes per curve; compound outlines also cap recursion at 128 levels, operand visits at 100 000 and cumulative geometry work at four million, rejecting cycles and partial results on limits. Pie sweeps normalize finite angles in constant time. Pixel layers are decoded at no more than 64 megapixels each (cropped to their content first; larger ones are read at a half, a quarter… of their resolution) and Limits::max_pixel_bytes for the whole document (2 GiB as RGBA, 512 megapixels, by default); pixel masks are read only over what they mask (all of it under artistic text, whose glyphs only the layout places). One budget covers the whole read: layers, pixels, 16 million characters of stories and 262 144 table cells, so content that many layers link, or a document embedded many times, can't multiply them. Embedded documents nest four deep at most; both of their readings are charged to that budget and to the importing document's archive limits, also when they fail part of the way. Tables have at most 4096 rows and columns and 65 536 cells, and 64 layer effects per node are read. Gradients are limited to 1024 stops. Text collection is bounded to a million characters per node before allocation, with a cumulative 16 MiB budget for owned font names (256 bytes at most each), features and gradient stops; paragraph formats are matched to runs by binary search. Exhaustion warns and uses default attributes for remaining text, preserving its characters. Colour and gradient overlays share one work budget per import (overlay::MAX_WORK: the nodes and pixels each is baked into); past it they are left out, and said. Null slots of object lists are charged to the document's work before a list is scanned, on every visit of a shared object. Every archive entry's CRC-32 and size must match. Malformed input returns an Error; the crate has no panics outside tests.

Privacy

Affinity documents can hold the folder they were saved in, user names, original image paths and XMP metadata. This reader never imports those fields.