Project status: Alpha as of May 31, 2026. APIs are actively being stabilized.
Raw image processing library focused on compatibility, correctness, and interoperability.
Image processing is messy and no single library could do it all (e.g., accurate colour, complete metadata, consistent format support). Rawshift seeks to stand out in at least some of the following ways:
- Compatibility: A single library processes all forms of image formats including all popular compressed and RAW image formats used by both consumers and creative professionals.
- Correctness: While decoding implementations should remain flexible (e.g. to slightly non-conformant image), it should strictly conform to open standards and retain maximum metadata across formats.
- Interoperability: This library compiles and is optimized for several standard desktop and mobile platforms. This is possible because the majority of the library is written in pure Rust and non-Rust dependencies are encapsulated with best practices.
rawshift is both a wrapper for image and video encoding/decoding needs, as well as implementations for various (often proprietary) RAW formats validated against specific camera bodies.
The key priorities in order are:
- Porting and stabilizing capabilities into Rust: A number of formats (e.g. HEIF, AV1) still depend on C/C++ libraries that are much more mature and battle-tested. It is unreasonable to port them in the short term and it is equally important to actively avoid implementations that are vibe-coded or that have dubious code licensing. The goal is to eventually support portable Rust equivalents that have the same features and performance characteristics of the benchmark implementations.
- Output Quality: Porting to Rust expands our test coverage and allows us to contribute to dependent libraries for feature parity. We validate these improvements by testing decoding against our camera database and encoding against our user base.
- Performance: Given mature and feature-complete libraries, we optimize the cost of common operations for various tasks end-to-end (e.g. format transcoding, metadata modification)
This library is still in active development. See
masterbranch for latest improvements (noting potential instability). Alpha packages may be published to crates.io occasionally.
Rawshift targets both still image and video formats. Image decoding is the current focus; video support is planned but not yet implemented (see Video).
Features and performance are constantly improving. Most functionality is implemented from scratch to meet project goals, so expect progressive format-support improvements over time. We aim to be liberal in what we accept (decode) and strict in what we give (encode).
Decode support spans the common compressed formats (JPEG, PNG, WebP, JXL, AVIF, HEIC, GIF, TIFF, SVG, PPM) plus a growing set of RAW formats, prioritised as:
- Sony ARW (all variations at least up to v5.0.1)
- Adobe DNG (up to v1.7, including what is necessary for Apple ProRAW)
- Standard TIFF
- Canon CR3
- Canon CR2
The full per-format decode/encode support table — backend libraries, encode
availability, and maturity status, plus notes on encoding and default
implementations — lives in the
rawshift-image README.
Video is parked for v1 — rawshift v1 ships image only. No video code ships
today, rawshift-video is unpublished (publish = false), and it is not a
dependency of the rawshift facade, so there is no video feature to enable.
The roadmap of formats (XAVC HS/S, Apple ProRes, HEVC, H.264) and the
container-first implementation plan live in the
rawshift-video README; the crate is
re-added to the publish set and the facade when it has an implementation.
rawshift is a Cargo workspace:
| Crate | Purpose |
|---|---|
rawshift |
Facade. Re-exports rawshift-image behind the coarse image feature. Most consumers depend on this. |
rawshift-image |
Still-image decoding, RAW processing, and encoding. Carries the full per-format feature system. |
rawshift-image-* format crates |
Independently usable format leaves (jpeg, png, dng, arw, and others) behind shared rawshift-image-core contracts. |
rawshift-image-{metadata,ifd,ljpeg} |
Focused support crates shared by format leaves; applications normally use them transitively. |
rawshift-video |
Video support — parked and unpublished for v1 (see Video). |
rawshift-core |
Shared types — geometry, codec descriptors, the metadata model. Charter is documented on the crate. |
The rawshift facade deliberately exposes only coarse features:
image(default) — still-image support (rawshift-imagewith its own default formats).serde—Serialize/Deserializefor metadata and option types.hw,hw-videotoolbox,hw-vaapi,hw-mediacodec— hardware still-frame decode (HEVC for HEIC, AV1 for AVIF) viarawshift-hwdec;hwpicks the native backend for the compile target, thehw-*flags pin one explicitly and fail the compile elsewhere (see docs/SUPPORT.md).full— every image format,serde, andhw.
There is no video feature: video is parked for v1 (see Video).
The facade does not re-export per-format flags. Cargo cannot forward a child
crate's features, so re-listing them would be duplicated, rot-prone state — and
a build that wants only video should never have to reason about image flags. For
fine-grained control (individual formats, alternative codec backends) depend
on rawshift-image directly.
The image and video libraries each carry their own feature systems — restating them here would duplicate state that drifts. See each crate's README:
rawshift-imagefeature flags — a tiered per-format system (bundles → formats → directions → infrastructure; gamut is the backend, plus four retained implementation aliases for the permanent exceptions and blocked TIFF migration).rawshift-videofeature flags — mirrors the image tiers; currently gates no code (video is unimplemented).
A device is officially supported if we have thoroughly tested compatibility for it.
Compatibility is verified against the default decoder implementation for each format (the named library for that format in the
rawshift-imagesupport table). Non-default implementations selected via implementation feature flags are not covered by this list.
The Image Formats column lists formats with verified decode support. The Video Formats column lists formats produced by the device that are on the video roadmap — these are not yet implemented or verified.
| Device | Image Formats | Video Formats | Notes |
|---|---|---|---|
| Sony A7RV (ILCE-7RM5) | ARW, JPEG, HEIC | XAVC HS, XAVC S | |
| Sony A7IV (ILCE-7M4) | ARW, JPEG, HEIC | XAVC HS, XAVC S | |
| Sony a6700 (ILCE-6700) | ARW, JPEG, HEIC | XAVC HS, XAVC S | |
| iPhone 13 Pro (Max) | DNG, HEIC, JPEG | HEVC, H.264, ProRes | |
| iPhone 16 Pro (Max) | DNG, HEIC, JXL | HEVC, H.264, ProRes |
The minimum supported Rust version (MSRV) tracks the minimum required by our upstream dependencies (currently 1.92.0, set by gamut) and will remain as low as the upstream dependencies require — it is never raised independently.
rawshift consumes the published gamut crates for image primitives, color, metadata, container parsing, and codecs. Their versions are managed centrally in the workspace Cargo.toml; git dependencies are not permitted because they prevent publishing rawshift. See AGENTS.md for the upstream-first policy that governs when rawshift may change in response to a gamut gap.
The workspace dependency table is the single point where upstream behaviour enters rawshift. Updating it is a deliberate, reviewed change, not a drive-by edit.
- One commit, one concern. The commit updates only the gamut requirements in
[workspace.dependencies](plusCargo.lock). No code changes ride along — migrations that depend on the update land in follow-up commits or a separate PR. - Confirm the gate shipped. Every gamut issue that blocked rawshift work must be closed and merged to gamut
master, and the affected crates must be published to crates.io. - Full test run.
cargo test --workspaceandcargo test -p rawshift-image --features fullmust pass, including fixture-driven tests. Do not use workspace-wide--all-features: it deliberately enables mutually exclusive platform hardware backends. - Full benchmark run. Run the criterion benches (
decode,demosaic,pipeline) against the pre-update baseline. Unexplained regressions block the update. - CHANGELOG note. Any behavioural change — decoder output bytes, error variants, metadata round-trip fidelity — gets a
CHANGELOG.mdentry. A pure no-op update is recorded as such.
Crates are added to [workspace.dependencies] lazily, as each migration issue starts consuming one, so the dependency list stays an accurate record of what rawshift actually uses.
It is important that development velocity is maintained regardless of project complexity. Unit tests for all contributions are expected, especially for platform-specific behaviours!
Install lefthook and activate the pre-commit and pre-push hooks:
# macOS / Homebrew
brew install lefthook
# Linux (Homebrew on Linux)
brew install lefthook
# via cargo
cargo install lefthook
# then install the hooks
lefthook installThe pre-commit hook runs cargo fmt --check and cargo clippy.
The pre-push hook runs the full test suite.
# whole workspace, default features
cargo test --workspace
# everything, all image formats
just test-allFixture-based integration tests need test data — just setup-test-data fetches
real fixtures and generates synthetic ones. See the justfile for the full set
of recipes (just build-image, just build-video, just test-features, …).
It is my intention (as developer and maintainer) to ensure rawshift remains open permissively to all.
While many open-source implementations historically used LGPL or similar licenses, rawshift prefers a more permissive license (MPL-2.0; see LICENSE). You are free to link to any software although we welcome contributions in any way.