Specifications

01 System requirements

Halidra runs as a Windows desktop app with a WebGL2 develop pipeline and a pure-JavaScript RAW reader. No native modules, no Adobe dependency.

Operating system

Windows 10 or 11, 64-bit (x64)

GPU

WebGL2-capable — the whole develop pipeline runs on the GPU

Download size

~95 MB (portable single .exe or installer)

Disk

Download size, plus a thumbnail cache

macOS

Coming soon

Internet

Not required to edit

Note: Minimum RAM and CPU figures have not been published yet. Tier 2 sensor decode is the heaviest task and runs on a single thread.

02 Distribution

Two Windows x64 executables, both built from the same source.

Packaged builds are checked with a self-test before being handed over: it boots the app, confirms WebGL initialises, confirms the decode worker loads as an ES module from inside the packaged app, and confirms a real image renders to non-black pixels. Builds are signed with Microsoft Artifact Signing (Azure Trusted Signing), and the signature is checked by asking Windows directly — not by reading the build log.

03 Camera & file support

Every RAW file is read in two tiers, and which tier you get decides how much latitude you have — not whether the file opens.

Tier 1 — the embedded preview

Every RAW the camera wrote a JPEG preview into opens, culls, rates and exports. That is effectively every camera: it's the camera's own full-resolution rendering, which is also what makes culling quick.

Tier 2 — the sensor data

Demosaicing the actual sensor plane gives real highlight headroom and the full tonal range. Press Ctrl+Alt+R to switch a photo to it.

Containers & other input

Containers parsed directly: CR2, NEF, ARW, DNG, ORF, RW2, PEF, SRW, IIQ, 3FR (TIFF/IFD), CR3 (ISO-BMFF) and RAF (Fujifilm header). Anything with an embedded preview opens via a brute-force scan even if the container itself is unrecognised, so unknown formats still open at Tier 1. JPEG, PNG and TIFF files can also be imported directly alongside RAW.

Tier 2 is opt-in — Ctrl+Alt+R — pure JavaScript, and takes about 13 seconds for a 26 MP frame on one thread, which is why it isn't the default. If a body ever looks wrong, the same shortcut returns you to the embedded preview.

04 Colour pipeline

Colour needs no per-camera table. Instead, the camera-RGB → sRGB transform is fitted per-photo against the camera's own embedded preview of that same frame.

Matrix fitting

Full 3×3 matrix fitted against the embedded preview

Row constraint

Each row sums to 1, which keeps neutrals neutral

Fallback

Per-channel gains when the scene can't constrain a full matrix

Fallback rate

3 of 12 reference files take the fallback

Black level

Measured from the optically shielded sensor border

White balance

As-shot, read from the camera's MakerNote

A full matrix needs six free parameters, which needs a scene with colour in it; per-channel gains need only two, so they work on a near-monochrome scene. Two of the twelve reference files are near-monochrome, and one has ample blue-green variation but almost no red-green, leaving its red row unconstrained — all three take the gains fallback. Gains fix the colour cast, which is what makes an uncorrected render obviously wrong, but leave saturation where the sensor put it. Which tier a file lands in is reported in the sensor-decode line and asserted by the test suite.

Canon files state none of the DNG metadata a decoder needs. The MakerNote is read for the sensor borders (SensorInfo) and the as-shot white balance (ColorData, whose layout is identified by record length); the white level comes from the real bit depth of the compressed stream rather than an assumption.

05 Develop controls

100 controls across 10 groups, plus Auto Tone, Auto White Balance, a before/after split view and clipping warnings. Every control takes the mouse wheel — Shift for fine steps, Ctrl for coarse — and this works identically on a favourited copy of a control.

06 Masking spec

Add a mask and every control in the app — tone, curves, colour mixer, colour grading, detail, effects, calibration — applies inside it and nowhere else.

Per-mask controls

Feather, opacity, invert, show/hide

Stack limit

Up to 8 masks per photo

Coordinates

Source-image coordinates — stay glued through crop, straighten, 90° rotation, perspective and lens distortion

Rendering

Each mask gets a complete pipeline pass of its own, composited over the global result by coverage

Select a mask and the develop panels turn blue and start writing into it; deselect and they go back to the whole photo. A mask is seeded from the photo's settings at the moment it's created, so its sliders start where the picture already is — later changes to the global settings do not flow into an existing mask.

The test suite asserts confinement exactly: at the centre of a mask the result equals the fully-adjusted value, and a sample outside it is identical to the untouched original — not close, identical.

07 Healing spec

Click a blemish and it is repaired; drag to choose where the repair comes from. Q opens the tool, Shift+Q toggles Heal ↔ Clone.

Heal

Blend with background — keeps the source's texture, replaces colour and brightness with the destination's

Clone

Replace with similar — the source area is copied verbatim

Formula

source − blur(source) + blur(destination)

Auto source

Ring matching around the spot against candidate positions, rejecting overlaps with the spot, other spots, or the frame edge

Spot limit

Up to 32 spots per photo, all applied on the GPU, all fully reversible

Coordinates

Source-image coordinates — stay stuck through crop, straighten, 90° rotation, perspective and lens distortion

The difference between Heal and Clone is easy to see on a gradient: healing from a much darker area onto a bright background reads 196 against a 198 background (invisible), while cloning the same source reads 42 — an obvious dark blotch. That comparison is an assertion in the test suite.

Brush size follows the mouse wheel or [ / ]; feather and opacity are set per spot; Del removes the selected spot.

08 Auto adjust

The Auto ▾ button in the toolbar, or its shortcuts, offer three different solves.

The neutral-area solve is a true solve, not a guess: on a blue-cast test frame it moves the blue-minus-red difference from +59 to −4.

09 Styles

29 built-in looks in the Styles panel, grouped into seven categories.

A Strength slider scales the whole look: numeric settings interpolate from their defaults toward the style's values, so 50% really is half the look rather than a different one. Switching styles cleans up whatever the previous one touched, and Clear returns those controls to their defaults — a style is never a trap. Styles are separate from presets: presets capture your photo's exact settings, styles are reusable starting points.

10 Library & culling

• Import a folder (optionally including subfolders) of RAW and JPEG/PNG/TIFF files
• Grid and filmstrip with lazily-decoded, disk-cached thumbnails
• Rate 0–5 stars, flag pick/reject, colour-label red→purple
• Auto-advance mode — a rating keystroke moves straight to the next frame
• Filter by minimum rating, flag, label, text; sort by name/date/rating/flag/size
• RAW+JPEG pairs shot together are detected and grouped

11 Crop, geometry & zoom

Crop — C

Drag the rectangle or its eight handles, lock to an aspect ratio (1:1, 4:5, 3:2, 16:9, 21:9, original, free), straighten with the angle slider, rotate in 90° steps, flip orientation, and choose an overlay (thirds, golden ratio, grid, diagonal, centre cross). Enter applies, Esc cancels. The crop travels with the photo through rotation, copy/paste and export.

Zoom

Fit / Fill / 1:1 / 2:1 / 4:1 buttons, + / - stepping, Z for 100%, Space to toggle fit↔100%, cursor-anchored scroll-wheel zoom, drag to pan, and a navigator thumbnail showing (and letting you move) the visible region. The renderer only processes the visible region at screen resolution, so zooming to 1:1 costs the same as fit-to-window.

12 Display calibration

Ctrl+Shift+K. A view transform, never an edit.

Set your panel's white point (or dial R/G/B gains by eye against neutral strips), correct any shadow cast with black-point lifts, and match gamma against a striped reference. Halidra then compensates the preview so what you judge as neutral on your screen is neutral in the actual file.

Exported pixels are never compensated — asserted in the test suite, which renders the same photo through both paths and checks the export is bit-identical with calibration on and off while the preview measurably changes. This is calibration by eye, not measurement; if you own a colorimeter, profile the display with that and leave this switched off.

13 Copy / paste / sync / presets / history

Presets save the whole look for reuse; a per-photo history list lets you step back through edits. Favourites: drag any control's label out of a develop panel into a Favourites box in the right panel — reorder within a box, drag between boxes, add and rename boxes. A favourited slider is a live copy, in sync with the original, and the layout persists between sessions.

14 Keyboard

Ctrl+K opens the shortcut editor: 576 bindable commands, searchable and filtered by category. Every control is assignable — each slider gets Increase / Decrease / Reset / Focus / Add to Favorites; toggles get Toggle; dropdowns get Next/Previous Option. Record a combination with + Add. Conflicts show in red and the new binding takes the key. Multiple keys per command are allowed, and the keymap survives upgrades.

15 Export formats

Ctrl+E exports the selected photos, rendered through the same GPU pipeline at full resolution, with quality and long-edge resize. The originals are untouched in every case.

16 Data & storage

Catalog

.halidra/catalog.json

Location

Beside the photos

Originals

Opened read-only, never written to

Legacy catalogs

.aperture/

Ratings, flags, labels, crops, healing spots and every develop setting live in the catalog and are baked into pixels only when you export. Catalogs written by earlier versions under .aperture/ are still read, and favourites, shortcuts, presets and calibration are migrated across on first launch, so nothing is lost in the rename.

17 Quality assurance

76 end-to-end checks boot the real renderer headlessly and cover WebGL initialisation, folder import, thumbnail and full-image decode, that exposure / white balance / saturation / tone curve / clarity / sharpening / vignette / grain actually change rendered pixels, rating and catalog round-trips, keyboard dispatch and rebinding, favourites drag targets and cross-copy sync, crop geometry, 90° rotation with crop remapping, copy/paste, undo/redo, zoom and pan clamping, image orientation, mouse-wheel control of sliders/toggles/dropdowns, healing, styles and strength scaling, display calibration reaching the preview but never the export, TIFF structure and bit depth, masking confinement, all three auto-adjust modes, and export in all five formats. Screenshots are written to test/screenshots/ on every run.

Per-vendor RAW suite

A separate suite runs against ~500 MB of CC0 RAW samples from raw.pixls.us — the same database darktable and RawTherapee regression-test against. Opening, reading metadata and rendering the preview are pass/fail; sensor decode for vendors not yet covered is reported as a named gap, not a failure, and a gap becomes a pass once the codec lands.

LibRaw reference comparison

LibRaw's unprocessed_raw writes the mosaiced sensor plane with nothing else applied, giving an exact pixel-level answer to compare against — catching errors a preview-correlation check alone would miss (this is what caught a Sony plane sitting two bits low, which still correlated at 0.98 against the preview).

Correctness without reference renders

A new decoder is correlated against the camera's own preview — a different rendering of the same frame — over a 24×16 grid, deliberately coarse enough to be fair (a correct decoder scores 0.98) but fine enough to catch structural errors (a decoder with its rows shifted by one byte scores 0.79 at this grid, versus 0.96 at a too-coarse 4×4).

Packaged executables are verified with a self-test before release, and their signatures are verified rather than assumed.

18 Functional requirements

A representative slice of the product requirements, with an acceptance criterion the README states is asserted or observable for each.