Krita documents

Picto opens Krita's native documents (.kra, and the .krz archive variant) with their layers. File › Open, recent files, drops, the CLI (convert, info, batch) and automation (doc.open, MCP) all use the reader in picto-io; the browser build accepts them too. A file is recognised by its mimetype entry (application/x-krita). Krita files are import-only: Save suggests OpenRaster (.ora), which Krita opens with its layers, and .pcraft keeps everything.

What is read

Krita	Picto
Paint layers, 8 and 16-bit RGBA, 8 and 16-bit gray	pixel layers, at the document's depth (a gray layer in an RGB document is converted)
Layer position (`x`, `y`), tiles anywhere, also off-canvas	kept
Groups, nested; pass-through groups; collapsed state	groups; pass-through; expanded state
Names, visibility, opacity, lock	kept
Blend modes	Krita's ids, mapped as for OpenRaster (`docs/ora.md`); a few are approximations, reported
Transparency masks	the layer mask (the first one; more are reported and left out)
Selection masks	ignored (a layer's local selection, not part of its look)
The image's ICC profile, resolution	kept

Left out with a warning per layer: vector, filter, fill, file and clone layers, filter, transform and colorize masks, layers in another colour space than those above, Inherit Alpha, per-channel blending flags (channel locks), and animation beyond the frame saved in the layer. Drawing assistants and reference images are reported as left out. Krita's merged image still shows the original look, so the result can differ where those were used.

Documents whose image colour space we don't read yet (CMYK, Lab, XYZ, YCbCr, 16 and 32-bit float) open as Krita's merged image (mergedimage.png), flagged as a preview, with a warning.

The format, as observed

A ZIP whose first, stored entry is mimetype; maindoc.xml holds the node tree (topmost first) and points each paint layer and mask at a file under <image name>/layers/. Those files are a text header (VERSION 2, TILEWIDTH, TILEHEIGHT, PIXELSIZE, DATA <tiles>) and per tile a x,y,LZF,<bytes> line followed by a compression flag and the tile's pixels in byte planes, LZF compressed. RGB pixels are B, G, R, A; 16-bit samples are little-endian. A .defaultpixel file next to each holds the colour outside its tiles. A group's own x/y does not move its children (checked against Krita's merged image).

Limits and safety

The reader uses the bounded ZIP reader of picto-format (no ZIP64, CRC-checked), parses maindoc.xml with roxmltree (32 MiB cap; the DTD is declared by public id and not fetched), and decodes tiles with a bounds-checked LZF decoder that must produce exactly one tile. Tiles are at most 1024 px a side; canvases at most 300,000 px a side and 1 GiB of decoded pixels; 8,192 nodes; nesting at most MAX_GROUP_DEPTH. Version 1 tile files (Krita 1.x) are refused. Malformed archives, XML, numbers, tile headers and tile data return errors; truncated and corrupted files are covered by tests.

Verification

Unit tests (crates/io/src/kra/tests.rs) build Krita archives in code (raw and LZF tiles) and check layers, groups, pass-through, offsets, opacity, visibility, locks, masks, default pixels, 16-bit and gray conversion, the merged-image fallback, LZF edge cases and malformed input. crates/automation/tests/it/krita.rs opens a .kra through the agent surface, edits it and saves it as OpenRaster. crates/engine/tests/it/kra.rs (opt-in) opens a Krita-authored 16-bit file in a session, checks it is read-only, edits with undo and saves as .ora and .pcraft.

Real-file checks use the Krita-authored oracles in picto-corpus (krita/, made by tools/krita-oracles/generate.py, which removes Krita's embedded third-party ICC profiles so the files are sRGB):

KRITA_TEST_DIR=/path/to/picto-corpus/krita cargo test -p picto-io --test it -- krita_corpus:: --ignored --nocapture
KRITA_TEST_DIR=/path/to/picto-corpus/krita cargo test -p picto-engine --test it -- kra:: --ignored

2026-10-09, Krita 5.2.9: 31 of 35 .kra files render within 3/255 of Krita's merged image; the others differ for known reasons (the three blend-mode tie-breaks as for OpenRaster, and the file with a vector layer, which is left out). Before the profiles were stripped, a linear-light 16-bit document with its embedded profile matched Krita's merged image within 0.5/255 after flattening and converting to sRGB.

The reader is written from files Krita writes; no Krita code was used.