Skip to content

PictureSharp 0.6.5

Choose a tag to compare

@art-ix art-ix released this 16 Oct 09:48
· 86 commits to main since this release

HDR photos from iPhone (and Android Ultra HDR 1.1): JPEG files with a gain map are read and reconstructed in linear
light. Loading is unchanged unless DecoderOptions.LoadGainMap is set.

  • Loading: with LoadGainMap = true, Image.Load keeps the same SDR base pixels and attaches the compressed gain
    map; image.Hdr (HdrInfo) and image.HasHdr describe it, Image.Identify fills ImageInfo.Hdr from the
    headers. ISO 21496-1 (iOS 18+, Ultra HDR 1.1) and Apple HDR Gain Map (MakerNote headroom, iOS 14–17) are
    supported; ISO wins when both are present. Damaged or unsupported HDR data never fails the load.
  • Reconstruction: image.ToHdr(HdrOptions?) returns an HdrImage with Rgba128 pixels (four floats, SDR white
    = 1, nothing clipped) in linear Display P3 or sRGB, with an optional linear TargetHeadroom. ISO reconstruction
    follows libultrahdr v2.0.2 (inverse-distance sampling, gamma, offsets, headroom weight, base or alternate color
    space, gray or RGB maps). ApplyExifOrientation and TargetSize work without a full-size map.
  • Tone mapping: hdr.ToneMap(ToneMappingOptions?) returns an 8-bit sRGB image with a documented extended
    Reinhard operator (exposure, white point).
  • Rules: Mutate and frame changes remove the map, Clone() keeps it; encoders never write HDR.
  • Metadata: ReadMetadata and EditMetadata now accept JPEG files with a valid gain map: both compressed images
    are kept byte for byte and MPF offsets are updated; changing the ICC profile or removing data the map needs throws
    InvalidOperationException. Other MPO files are still rejected.
  • CLI: info prints the gain map, convert -hdr tone-map [-exposure] [-white] renders HDR.

Validation: Release build with zero warnings; 5594/5594 tests on linux-x64 (including Oracle, Slow and local iPhone
samples); on ARM64 (Orange Pi 5) the full suite, Slow, Local, Oracle without libultrahdr and adapter tests pass, the
package consumer with trimming and Native AOT runs, and tone-mapped iPhone photos are byte-identical to x64. The
synthetic corpus is checked against an independent double-precision model (1e-5 relative); libultrahdr v2.0.2's
functions on identical pixels agree within the same tolerance, its full decoder within 0.42 % median. Plain JPEG
loading does no extra work without LoadGainMap.

Not yet verified: reconstruction against Apple ImageIO/Core Image on macOS (planned for the next release), BenchmarkDotNet
measurements. Maps whose aspect ratio differs by more than 1 % from the base use per-axis scaling instead of
libultrahdr's resize. See HDR photos.