20 KiB
20 KiB
File format parity
Last reviewed: 2026-10-10 · Last updated: 2026-10-12 · Change: minor (
.figexport: written files checked by opening them in other software; experimental.figexport; export and inspect behaviour suite and fixes; .fig import: editable vector networks and more layer kinds; embedded image colour profiles;.figimport, best effort; animated GIF fills and TIFF images; last major: created, formats from the installed Figma's Info.plist and help center, ours from source) · Target: Figma Design (desktop app 126.10.7)
Every format Figma reads or writes, with FigCraft's support. Sources: Info.plist of
/Applications/Figma.app (document types only; nothing inside the bundle's code was read), Figma's
help center (image and video types, export formats), and FigCraft's source.
Summary (estimated): ~60% ready. The core interchange formats (SVG, PNG, JPG, PDF) are solid
and tested; Figma's own document, .fig, opens (best effort, not yet checked against real-world
files) and can be written, experimentally: written files round trip through FigCraft's own
importer and were checked by opening them in other software (layers, fills, gradients, strokes,
shadows, vectors, components and text draw as in FigCraft); effects such as glass and noise, images
and prototype interactions are still unchecked there.
Documents
| Format | Figma | FigCraft read | FigCraft write | Fidelity / testing | Hours |
|---|---|---|---|---|---|
.fig (Figma Design, com.figma.document, application/x-figma) |
native (cloud; "Save local copy") | yes, best effort (file.open, file.import, drop, CLI, MCP) |
experimental (below) | .fig files open (best effort): the zip, its deflate/zstd chunks and the kiwi schema are decoded generically and mapped by field name onto the REST importer, so pages, the layer tree, geometry, paints (styles resolved), strokes, effects, text with styled runs, vectors as editable vector networks (points, curved segments, filled areas with their winding, per-point corner radius and handle mirroring, end caps; outlines when the network is missing or damaged), auto layout (grid included), constraints, ellipse arcs, boolean operations, components, sets (with their descriptions), instances (expanded from their main with overrides, derived layout and property values), styles, variables with modes and bindings, prototyping (set-variable, set-variable-mode and conditional actions included) and image fills from the archive come across; sticky notes, shapes with text, connectors, tables, code blocks, slides and similar layers come in as the closest frames, text layers and vectors, with a warning; anything else is a warning. Not read: connector labels, table cells beyond their own layers, filled areas with their own fills, per-point stroke joins, variables bound to fields the editor can't bind, fixed or content-sized grid tracks, video fills. Tested on fixtures generated in code (crates/format/src/fig/tests.rs, tail_tests.rs, fuzzed) and by reading back what the export writes; no real-world files yet |
read tail 5–10, write: checks against real files 5–15 |
.jam (FigJam), .deck (Slides) |
native to other Figma products | no | no | out of scope | — |
REST API file JSON (GET /v1/files/:key) + local variables JSON + images/ folder |
the API's output | yes (file.open, file.import, CLI) |
no | pages, layer tree, geometry, paints, strokes, effects, text runs, vectors, auto layout, constraints, grids, components, sets, instances with overrides, component properties, styles, prototyping (incl. conditionals and variable modes), variables with modes, aliases, scopes and bindings; warnings for what's skipped (video triggers and actions, modifier keys). Unit tests in crates/format/src/rest/*_tests.rs |
— |
.figcraft (ours: zip of document.json + images, versions inside) |
— | yes | yes | streamed JSON, 50k layers save ~0.09 s / load ~0.16 s (cargo xtask bench); old files keep opening (large_tests.rs) |
— |
| Local library files | team libraries in the cloud | yes (enable a .figcraft as a library) |
yes (publish) | per-asset versions, updates with overrides kept | see collaboration-parity |
Import (place / paste / drop)
| Format | Figma | FigCraft | Notes | Hours |
|---|---|---|---|---|
| SVG | editable vectors | yes, editable | corpus of 21 synthetic files with hand-computed expected pixels, structure and round-trip checks; gradients, patterns as image fills, clip paths, masks, shadow/blur filters, blend modes, text as text layers; other filters dropped with a warning; text on one path as text on a path (on several paths as outlines); fuzzed | 2–4 (filters) |
| PNG, JPG | yes | yes | embedded ICC profiles (PNG iCCP, JPG APP2) and PNG sRGB / gAMA / cHRM chunks convert the image into the file's colour space; no profile (or CMYK, damaged): taken as sRGB |
— |
| GIF | yes (animated) | yes, animated | the file is kept as is; the canvas and PNG, JPG and PDF exports show the first frame (SVG exports embed the file, so viewers that animate GIFs animate it), the presenter and the preview play it (frames decoded on first play: at most 500 frames and 128 MB, past that the rest don't play; always loops) | — |
| WebP | yes | yes | embedded ICC profile honoured | — |
| HEIC | yes | no | shortcut: no Rust decoder; the usable ones (libheif, libde265) are C libraries, and HEVC carries patents. Revisit when a Rust decoder exists | 2–3 |
| TIFF | Safari only | yes | open, import, place, drop (the image crate's TIFF decoder; its embedded ICC profile is honoured); SVG exports embed it (and BMP) as PNG, which SVG viewers show |
— |
| BMP | — | yes | — | |
| MP4, MOV, WebM (video fills, paid plans) | yes | no | shortcut: decoding H.264/HEVC/VP9 needs non-Rust libraries (FFmpeg or each platform's media framework); presenter playback would reuse the animated GIF clock | 4–8 |
| Images over 4096 px | downscaled to 4096 | kept | — | |
| Pasted content from other apps | images, SVG, text, Figma clipboard | images, SVG, text (newest clipboard content wins) | no Figma clipboard (it carries Figma's own binary) | with .fig |
Export
| Format | Figma | FigCraft | Notes | Hours |
|---|---|---|---|---|
| PNG | yes, scales/suffixes | yes | export settings per layer, presets (2x, 512w, 512h), preview, "Export N layers"; contents only (off: every layer showing in the area) and include bounding box (the layer's own box); slashes in layer names export into folders. The exported area is what the layer paints: shadows towards their offset, outside and centre strokes, unclipped children, a mask group's mask, nothing for hidden layers; slices export everything under them; sides round up to whole pixels |
— |
| JPG | yes | yes | white where transparent; fixed quality (90) | — |
| SVG | yes (outline text option, ids) | yes | alpha/luminance masks kept; noise, texture, progressive and background blur, glass embedded as images; outline text (off: <text> with one <tspan> per line, one style per layer), unique id attributes from layer names, simplified strokes (stroke attributes for centred strokes and for inside and outside strokes on boxes and ellipses) |
— |
| yes | yes | one page per frame, real selectable text in subset fonts, links, vectors and gradients as vectors, effects as images; bleed, crop/bleed marks, page boxes, naive CMYK; checked by rendering the PDF back against PNG exports | CMYK output intent 2–4 | |
| WebP | — | yes (ours) | — | |
| Copy as CSS / SVG / PNG / text / link | yes | yes | Copy as PNG: the whole selection at 2x | — |
| Dev Mode code (CSS, iOS SwiftUI, Android Compose) | yes | yes, plus HTML and React | see dev-mode-parity | — |
.fig ("Save local copy") |
yes | experimental | the whole file; see Export to .fig. Round trips through our own importer; written files checked by opening them in other software: layers, fills, gradients, strokes, shadows, vectors, components and text draw as in FigCraft |
5–15 (checks of effects, images and prototypes, and against real files) |
Export to .fig (experimental)
file.export {format: "fig"} (or a path ending in .fig), File ▸ Export… ▸ .fig (experimental),
figcraft-cli export <file> -o <out>.fig (and convert) and MCP's export {format: "fig"} write
the whole document. The writer (crates/format/src/fig/export.rs) works from the same public format
notes as the import and inverts its mapping: the zip holds canvas.fig (the fig-kiwi header, then a
raw-deflate binary schema and a raw-deflate message), meta.json, thumbnail.png and
images/<hash>; meta.json names the file and gives the page colour and the thumbnail size. The
schema written into the file describes only the fields written, under the names, kinds (struct or
message), field ids and enum values the format gives them. The document and each page say they are
visible and opaque (a page also its background opacity): a reader takes what is left out as off, and
draws nothing below a hidden page. The message
lists one node change per layer, parents first, each with its GUID, its parent and a position
string that orders it among its siblings; vector networks and outlines go in blobs, written once
each.
- Written: pages (name, background); frames, groups, sections, rectangles (as rounded rectangles, with their four radii), ellipses (arcs), polygons, stars, lines, vectors (editable network blobs with per-point corner radius, handle mirroring, caps and joins, plus the outline of their filled areas), text, boolean groups, slices; the fill outline of frames and shapes and the stroke outline of plain strokes (a ring around boxes and whole ellipses, the widened path of other centred strokes); solid, gradient (handles as the paint transform) and image paints (the image bytes, scale mode, rotation, tile scale, crop); strokes (weight, align, join, caps, dashes, miter limit, per-side weights); shadows, layer and background blurs (progressive too), noise and texture; corners and smoothing, constraints, clip content, masks with their type; auto layout (direction, wrap, gaps, padding, alignment, hug, strokes in layout, stacking order, grids with their tracks and cells) and how layers sit in it (fill, absolute, min/max sizes); layout grids; export settings; text style (font, with styles under their standard names such as "Semi Bold"; size; line height, a share of the font size as a raw multiplier; letter spacing, case, decoration, links), alignment, auto resize, truncation, paragraph spacing and indent, list types and styled runs with their fills (indexed by UTF-16 unit); components (description, boolean, text and instance swap properties and the layers bound to them), component sets, instances (their main, property values, and the fields their layers change from the main's, by id path, through nested instances and swaps); prototype interactions (clicks, hovers, presses, drags, mouse triggers, delays, keys with modifiers and controller buttons; navigate, overlays with their position, outside-click setting and background, swap, change to, scroll to, back, close, open link; transitions and easing, custom curves included), scrolling and fixed layers, flow starting points.
- Left out, with a warning each: shared styles (layers keep the values), variables and their bindings, comments, Dev Mode annotations, statuses, resources and code links, plugin data, ruler guides, prototype device settings, slot properties and slot content, variable-width and brush strokes (written plain), pattern fills, WebP export settings (written as PNG), image adjustments, glass settings other than frost, springs other than Gentle (written as ease out), set-variable, set-variable-mode and conditional actions, scroll-to offsets, text decoration looks, variable font axes, text on a path, nested list levels, list spacing, hanging punctuation, leading trim, OpenType features, per-layer operations inside boolean groups, grid tracks of fixed or content size, different stroke ends on lines. Not written at all: stroke outlines of dashed and per-side strokes and of inside or outside strokes on polygons, stars and vectors, glyph outlines of text (a reader lays text out with its own fonts), version history and branches.
- Checked by opening written files in other software: the example documents and a small test file open there, and layers, fills, gradients, strokes, shadows, vectors, components and text draw as in FigCraft. A file saved again there and read back keeps the layer tree and properties; two differences it showed were fixed (font styles are written under their standard names, and a line height that is a share of the font size is written as a raw multiplier, since percent units count from the font's own line height); the import now reads per-side stroke weights that leave the sides without a border out. Still unchecked there: effects such as glass and noise, images, prototype interactions; variables are not written. Text boxes sized to their content can come out a pixel wider or narrower there.
- Round trips through our own importer: 60 generated documents (every layer
kind above, components, nested instances with overrides, a grid inside a component,
interactions) are exported, read back and compared layer by layer: structure, geometry within
32-bit float tolerance, paints, text and runs, auto layout, components and prototypes. The written
file decodes with the generic kiwi decoder and lists every layer after its parent; the encoder
agrees with the decoder on random messages and refuses values the schema can't hold; junk
documents never panic (
crates/format/src/fig/export_tests.rs,crates/engine/src/fig_export_tests.rs). What other readers need (visible document and pages, outlines, text settings, schema numbering, font style names, line height units, mask types) has its own tests (crates/format/src/fig/interop_tests.rs).
Fonts
| Figma | FigCraft | |
|---|---|---|
| Local fonts | yes (font helper / desktop app) | yes (system scan) |
| Bundled/shared fonts | Google Fonts in the cloud | optional CRAFT_FONTS_DIR build input (storytold/craft-fonts) |
| Variable fonts | yes | yes (axes, named instances, per-range axis values) |
| Missing-font replacement on open | yes | warnings only (REST import) |
Revision history
| Date | Change | Summary |
|---|---|---|
| 2026-10-12 | minor | .fig export made to open elsewhere (gap 1): the document and its pages say they are visible and opaque (left out, nothing was drawn), the schema in the file uses the format's own names, kinds, field ids and enum values, rectangles are rounded rectangles with four radii, frames and shapes carry their fill outline and plain strokes their stroke outline, text carries its layout settings, font styles use their standard names, line heights that are a share of the font size are raw multipliers, masks keep their type (read back too); the import reads per-side stroke weights whose zero sides are left out. Written files were checked by opening them in other software: layers, fills, gradients, strokes, shadows, vectors, components and text draw as in FigCraft; effects such as glass and noise, images and prototype interactions are still unchecked there, and variables are not written. Numbers unchanged |
| 2026-10-12 | minor | Experimental .fig export (gap 1): file.export {format: "fig"}, File ▸ Export… ▸ .fig (experimental), the CLI and MCP write the whole document (layer tree, geometry, paints with image bytes, strokes, effects, text with styled runs, vector networks, auto layout, constraints, components, sets, instances with their overrides and property values, prototype interactions) with a kiwi encoder that mirrors the decoder and the schema of the fields written; what's left out is a warning. The import gained ellipse arcs, boolean operation types and component set descriptions, and reads durations as whole milliseconds. Verified only by round trips through our own importer. Summary ~59% → ~60%; .fig write 15–30 h → 5–15 h (checks against real files) |
| 2026-10-11 | minor | Export and inspect behaviour suite written from the public documentation (139 cases in 19 tests; 47 failed before the fixes) and its fixes: an export covers what the layer paints, so a drop shadow adds room only where it falls, strokes on boxes and ellipses reach their real width (they added up to four times as much), hidden layers take no room and a mask group exports the mask's area; a slice exports everything under it; slashes in layer names export into folders; Copy as PNG copies the whole selection. Added: the export settings' contents only and include bounding box, and for SVG outline text (off writes <text>), unique id attributes and simplified strokes (stroke attributes where exact); Inspect colour formats (hex, RGB, CSS, HSL, HSB), CSS in rem, a Blend row, positions measured from the frame around a layer (groups aren't frames), CSS for rotation, blend modes, gradient directions and kinds, text alignment, case and decoration; icons and images in the asset list (file.exportLayers {assets}). Summary ~55% → ~56%; SVG options 1–2 h → done |
| 2026-10-11 | minor | .fig import's tail (gap 1): vectors open as editable vector networks (points, curved segments, filled areas with their winding, per-point corner radius and handle mirroring, end caps from the end points), falling back to outlines when the network is missing or damaged; sticky notes, shapes with text, connectors, tables, code blocks, slides and similar layers import as the closest frames, text layers and vectors with a warning instead of being dropped; set-variable, set-variable-mode and conditional prototype actions come across. New parsers fuzzed on fixtures generated in code. File formats 55% → 58%, 8–15 h → 5–10 h; import/export breadth 80% → 82%, ready 50% → 54%; feature and area totals and ready for real work (~49%) unchanged; beta ~173–310 h → ~170–305 h; full parity ~321–618 h → ~318–613 h |
| 2026-10-11 | minor | Embedded image colour profiles: placed and opened images are converted once, when decoded, from their own colour space into the file's (sRGB or Display P3): ICC profiles in PNG (iCCP), JPG (APP2), WebP and TIFF, matrix/curve and table profiles in RGB or grey (moxcms, already built for the image crate), and a PNG's sRGB, gAMA and cHRM chunks; images without a profile, or with one that can't be used (CMYK, damaged, over 4 MB), are taken as sRGB as before. The original file is kept (PDF exports still embed it); pixel exports carry the converted pixels with the file's profile. The image fill's details show the source profile (Unknown — treated as sRGB when there's none) and image.place reports it as imageProfile. |
| 2026-10-11 | minor | .fig import, best effort (gap 1): .fig files open (best effort): the zip, its deflate/zstd chunks and the kiwi schema are decoded generically and mapped by field name onto the REST importer, so pages, the layer tree, geometry, paints (styles resolved), strokes, effects, text with styled runs, vectors (as outlines), auto layout (grid included), constraints, components, sets, instances (expanded from their main with overrides, derived layout and property values), styles, variables with modes and bindings, prototyping and image fills from the archive come across; anything else is a warning. Summary ~40% → ~55%; .fig read 30–60 h → 8–15 h (write 15–30 h unchanged) |
| 2026-10-11 | minor | Animated GIF fills and TIFF images (gap 12, without video and HEIC): GIF fills keep the original file, show the first frame on the canvas and in exports and play in the presenter and the preview (frames decoded on first play, at most 500 frames and 128 MB; always looping; the window wakes only for a GIF's next frame); TIFF images open, import, place and paste. Paint breadth 85% → 87%, 5–9 h → 4–7 h; gap 12 6–12 h → 5–10 h (video playback and HEIC left, both needing non-Rust decoders); feature breadth (~79%), ready for real work (~47%) and full parity hours unchanged at this rounding |
| 2026-10-10 | major | Created from the 2026-10-10 re-measure |