vectorcraft/crates/affinity/README.md

8.4 KiB
Raw Permalink Blame History

vectorcraft-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 hands the engine a document in pixels with every transform composed. It does not write Affinity files.

What is imported

Affinity VectorCraft
Pages, Publisher spreads (one or two pages) artboards, spreads side by side
Artboards (ShpN with ABEn), with name, background fill, any transform artboards plus a clip group with the background
Layers (Scop), groups (Grup), pass-through or isolated layers and sublayers, groups
Curves (PCrv): cubic subpaths, closed flags, live corners (CnrD) paths (multi-subpath curves fill even-odd)
Rectangles with corner radii (relative or absolute), ellipses, polygons (smooth too), stars, square stars, pies, triangles, trapezoids paths
Compound shapes (Comp, add and subtract) one even-odd path
A shape or curve with children clips them clip group whose clipping path keeps its fill (behind) and stroke (over)
Vector masks (AdCh curves and shapes) clip group
Pixel masks (MRst, attached or as a layer in a group) opacity mask
Solid fills: RGBA, HSLA, CMYK, grey, Lab (D50) fill appearance (alpha as fill opacity)
Linear, elliptical and radial gradients with Affinity's midpoint bias gradients (the bias as the midpoint where the blend is half way)
Strokes: weight, scale with object, dash pattern and phase, alignment, miter limit, behind the fill stroke appearance
Fill layers (FRst) filled rectangles
Artistic and frame text: characters, font family/weight/style, size, tracking, fixed leading, colour, paragraph alignment, first baseline point type and area type
Placed images (ImgN): the original JPEG/PNG embedded in the file embedded images
Pixel layers (Rstr): RGBA 8/16-bit and CMYK 8-bit tiles, cropped to their content embedded PNG images
Embedded documents and symbols (EmbN) the picture Affinity cached of them
Opacity, visibility, lock, names, the blend modes VectorCraft has the same

Everything else is reported in the import warnings, one line per kind with a count, never dropped silently: layer effects, adjustment layers and live filters, brush and pressure strokes, transparency gradients, fill opacity, bitmap fills, master pages, conical gradients (as radial), special shapes (cloud, heart, cog, callouts, arrows…, imported as their bounding ellipse), corner types other than round, stars with rounded points, several fills or strokes on one object (the active one is used), outlined or scaled text, text fields such as page numbers, frames with columns or a curved outline, grey/Lab/32-bit pixels, CMYK pixels (converted to RGB without the document's profile) and Affinity-only blend modes (Add, Linear Light… as Normal).

When the native document can't be read at all, the engine opens the embedded PNG preview instead, with a warning that names the reason; that preview can't be placed, used as a template or mined for swatches and styles. Save never writes back over the Affinity file.

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. This is how the sibling app PhotoCraft already treats formats that are only partly documented: its PSD reader fills the gaps in Adobe's specification from the MIT-licensed psd-tools and ag-psd, and its camera raw decoder recovered Nikon's compression tables by black-box analysis of CC0 sample files.

  • 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.

The same reader is shared, as an independent copy, with PhotoCraft's photocraft-affinity.

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.

  • cargo xtask corpus --affinity fetches 22 public CC0/MIT/Apache-2.0 documents, including the four Affinity 3 .af files of samuel-etver/vector-art (CC0), at pinned commits, each checked against xtask/affinity-corpus.sha256. tests/real_files.rs parses them; engine/tests/affinity_corpus.rs renders each one and compares it with the thumbnail Affinity saved in it (mean difference 0–4.4 of 255 at the thumbnail's size, with a ceiling per file).

  • engine/tests/import_fuzz.rs mutates a synthetic native document's stream and archive and checks that whatever opens also renders and exports; 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, 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, 64 megapixels per pixel layer (cropped to its content first) and 1024 gradient stops. 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.