Light, the way your screen can show it.

SDR to HDRA Metal pass on the native display path

Pipeline

Y′CbCr or BGRA → R′G′B′ → linear light, with the curve macOS applies to each buffer type (BT.1886 or sRGB, measured) → expansion → BT.2020 primaries → 64-bit half-float linear buffer (1.0 = SDR white), handed to AVSampleBufferDisplayLayer without synthetic mastering metadata.

Peak

min(setting, current EDR headroom); skipped for PQ and HLG sources and for the presentations that ask for SDR.

Fast

The curve above, per pixel.

Balanced

Lifts only what lies above diffuse white, by an amount that shrinks when large areas are bright, with scene statistics, a guided base layer, an area limiter, debanding, and protection for burnt-in subtitles and logos.

High, Maximum

Balanced’s analysis plus a bilateral grid of gains predicted by a small network from a 256×256 thumbnail, run with MPSGraph from .maclcnn files whose every layer shape is checked on load. Gains are read as a share of log(10); each channel stays within 0.2 log of the luminance-weighted gain (0.06° of colour accuracy on held-out scenes).

Model

sdr2hdr-high-v1, sdr2hdr-max-v1: trained by Hazen Studio on Poly Haven CC0 photographs, LGPL like the module that loads them.

Auto

Chooses from power source, thermal state, GPU class and pixel rate; on GPU-time overshoot it steps down as far as Fast.

Measured (M3 Max)

Model prediction 0.45 ms (High) and 0.53 ms (Maximum) per frame. When the expander was introduced: 0.53 ms at 1080p and 0.91 ms at 4K including the synchronous wait; agreement to the code value with an independent CPU reference over an eight-step ramp; a 60-second 4K run in the same wall-clock time as with the switch off, with flat memory. In-app results match the offline model within 1–3 % on HDR captures.

Prior art

Inverse tone mapping is not new: libplacebo, mpv and some TVs do versions of it. What MacLC adds is doing it on macOS’s default display path, where libplacebo cannot reach, as one switch and one slider, with a model trained for it and the numbers above.

HDR filmsFormats, choices and picture modes

Formats

Dolby Vision profiles 5, 7 (base layer and RPU; enhancement layer ignored), 8.1 and 8.4; HDR10+ (ST 2094-40, kept until the stream replaces it, as TVs do); HDR10 (ST 2084 with ST 2086 mastering and CTA-861.3 MaxCLL/MaxFALL); HLG. Matroska’s Dolby Vision configuration record (dvcC, dvvC, dvwC) is read before the first frame.

Advisor

Reads transfer, primaries, mastering volume, MaxCLL/MaxFALL and the Dolby Vision configuration of the track, and the display’s potential, current and reference EDR headroom, battery and Low Power Mode. If the master fits under the display’s peak, the three formats render the same image and HDR10 is recommended (macOS’s own pipeline, lower power); otherwise Dolby Vision, then HDR10+, then HDR10. Automatic track restarts are capped at two per file; switching formats seeks precisely and takes about a second.

Picture modes

Accurate reproduces every level up to 90 % of the display’s peak and folds the rest into a soft PQ-domain exponential shoulder; Balanced is the ITU-R BT.2390 EETF; Bright lifts by up to 1.6× before the BT.2390 roll-off. When the master fits, the mapping is the identity (or a plain gain clamped at the peak). Checked for ST 2084 round trip, monotonicity from 0 to 10,000 cd/m², the peak bound and continuity at the knee.

Reference white

PQ anchored at 100 cd/m² at SDR white, the convention macOS applies to PQ surfaces (measured with HDR captures on macOS 26.2). HLG goes through the BT.2100 OOTF for the 1,000-nit reference display (reference) or for 100 × the current headroom (display).

Engines

Native (AVSampleBufferDisplayLayer, EDR) for HDR10, HLG and SDR, re-mapped on the GPU with the picture-mode curves. For Dolby Vision RPUs and HDR10+: MacLC’s Metal renderer (Dolby Vision decoder ported from libplacebo 7.351, 0.05 % mean error against the C reference on a 4K profile 8.1 frame), with OpenGL kept as a fallback.

Fixed in the inherited HDR pathAn audit of the macOS HDR path of VLC 4.0-dev, on which MacLC builds, found six defects and nine missing capabilities on Apple Silicon

Also addedBeyond the fixes

AV1 in hardware

VideoToolbox on M3 and newer (checked with VTIsHardwareDecodeSupported), dav1d elsewhere.

Dynamic metadata

HDR10+ carried through the filter chain; Dolby Vision extracted from the VideoToolbox bitstream with hardened parsers, including a fix for Dolby Vision lost to a stray bit on the first frame.

Colour

ST 2086 and CTA-861.3 metadata forwarded to the display layer; ICC profile taken from the window’s own screen.

Software decoding

Hand-written ARM64 NEON chroma conversions (NV12/P010 ↔ I420/I420_10L).

OpenGL

Core profile 4.1/3.2 with 64-bit colour, which lets libplacebo’s shaders run (VLC failed with “OpenGL version too old”).

Advanced settingsSettings › HDR & Colour › Advanced

Tone-mapping curve

Automatic (MacLC’s picture modes), hard clip, ITU-R BT.2390 EETF, Reinhard, Mobius, Hable, gamma-power, linear stretch, ITU-R BT.2446A, single-pivot spline

Tone-mapping parameter

The chosen curve’s roll-off; 0 keeps its default

Gamut mapping

Automatic, clip, perceptual, relative colorimetric, saturation preserving, absolute colorimetric, desaturate (BT.2020 sources)

Scaling filters

Built-in bilinear, bicubic Mitchell-Netravali and Catmull-Rom, Lanczos, Spline16/36/64, EWA Lanczos (Jinc), for upscaling and downscaling

Dithering

Blue noise, white noise, Bayer 16×16 or off (OpenGL engine; Metal renders in 16-bit float)

Video engine

Automatic, Native or Metal; OpenGL (legacy) as a fallback

Decoding

Hardware decoders only; decoder pixel format (P010 10-bit by default, 8-bit formats available)

SDR to HDR, advanced

Mid-tones (as graded → lifted), highlight colour (softer → richer), smooth banding (off, normal, strong), protect subtitles and logos

OptionsFor the command line and vlcrc