RenderPreparation measures (or reuses cached) luminance, colour, region
and camera-path series and turns them into per-frame corrections.
RenderPipeline renders planned frames through exposure, one geometry
resample, flow synthesis, motion blur, stacking and the knee, overlapping
encoding with processing; cancellation and failures leave no output and
no unobserved exceptions. Batch renders share one preparation.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Client renders prefer hardware Media Foundation encoders, fall back from
HEVC/AV1 to H.264, drop to 8-bit and cap size/bit rate with reason keys;
the server reports that MP4 codecs need a render node.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
ExportResolver applies overrides and names the result after the preset
it equals (or custom); ExportNaming expands the name template with
collision suffixes; ExportEstimator predicts size and time from a
calibratable throughput model.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Sessions write to a temporary file or folder that Complete renames into
place and Dispose deletes; constant-rate outputs repeat held frames.
Adds AVI/GIF/APNG readers used to verify rendered outputs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Output frames on the export frame grid with fractional source positions,
exact 90 kHz durations, held/interpolated flags, motion cuts and plan
statistics; FitToSeconds rescales the output axis.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
EncodeRow8/16 reuse one pooled scratch row for RGB-to-gray output
instead of allocating per row; drop unused JPEG marker constants; the
24 MP timing test skips itself when the JIT optimiser is disabled.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Single-strip uncompressed files (Photoshop's default) exceed
TiffReader's bounded read; such strips are now read in pieces, the
uncompressed path skips a copy, and hostile strip/tile counts are
rejected up front.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Two-pass Lanczos3 (downscale, support stretched by the reduction) and
Catmull-Rom (upscale) with normalised, edge-folded weights; exact
area-average box path for thumbnails; pass order chosen by cost,
Vector<float> inner loops, parallel rows. Tests cover constant images,
energy preservation, step ringing against the kernel's exact discrete
bound, Exif orientations and dithering.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
16-bit LinearRaw DNG 1.4 with tiles (lossless JPEG or uncompressed),
ColorMatrix1 = XYZ(D65) to linear sRGB, ForwardMatrix1 to XYZ(D50),
D65 illuminant, neutral AsShotNeutral, WhiteLevel 65535 and optional
Exif. LibRaw renders test patches back to the exact source values.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
TiffImageReader decodes strips and tiles (chunky or planar) with none,
LZW (old-style bit order tolerated), Deflate, PackBits and JPEG
compression (JPEGTables, YCbCr or RGB, lossless SOF3 tiles reshaped as
DNG writes them), predictors 2/3 and the DNG x2/x4 variants, 1-16 bit
integer and 16/32-bit float samples; tiles decode in parallel. Output
is a raw TiffRaster or a linear ImageBuffer. JPEG strips match libtiff
bit for bit.
TiffWriter writes 8/16-bit gray/RGB/RGBA, uncompressed or Deflate with
horizontal predictor, and embeds capture metadata from an Exif block.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
BuildTiff writes a little-endian TIFF with IFD0 (camera, orientation,
software, date, authorship, resolution), the Exif IFD (capture time
with sub-seconds and offset, exposure, ISO with RecommendedExposureIndex
above 65535, focal lengths, bias, body serial, lens) and a GPS IFD when
known. BuildJpegApp1 wraps it in a complete APP1 segment; strings are
capped at 2000 UTF-8 bytes on character boundaries.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
ImageSignature.Detect identifies every BRIEF 8.1 format by magic bytes,
refining TIFF-based RAWs by the CR2 marker, DNGVersion and Make, and
ISO-BMFF files by ftyp brands; the extension only breaks ties.
MetadataReader.Read (path or seekable stream) dispatches to the
container parsers, opens files shared for read/write/delete with a
small buffer, honours cancellation and turns any parser defect into a
warning so malformed input never throws.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
RAF: embedded JPEG Exif and preview, size from the RAF directory
(0x0111 cropped, else 0x0100), bit depth from the FujiIFD. PNG: IHDR,
eXIf and iTXt/tEXt XMP (zlib supported), seeking over image data.
WebP: VP8X canvas or VP8/VP8L headers, EXIF and XMP chunks. GIF and
BMP sizes, the JPEG XR IFD and the JPEG XL size header plus container
Exif/XML boxes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
BmffReader reads box headers only (32/64-bit and to-end sizes, uuid
user types, per-container count and path depth caps). HeifParser
handles hdlr/pitm/iinf (infe v2/v3)/iloc (v0-v2, multi-extent, idat)/
iprp: size from the primary item's ispe, bit depth from pixi, Exif and
XMP items; irot/imir define the orientation and override Exif because
HEIF decoders apply them. Cr3Parser reads CMT1-4 and THMB from the
Canon uuid box, PRVW, the XMP uuid, track 1's full-size JPEG through
the sample tables and the raw size from the CRAW/CMP1 header.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
ExifParser walks IFD0, Exif, GPS and SubIFDs, selects the main raster
(largest full-resolution IFD), and enumerates embedded JPEG previews
(JPEGInterchangeFormat, single-strip/tile Compression 6/7, lossless
raw strips excluded). JpegParser scans APP1 Exif/XMP/extended XMP,
APP2 MPF large thumbnails and SOFn up to SOS, tolerating fill bytes
and garbage. Vendor MakerNotes (Canon, Nikon, Sony, Fujifilm,
Olympus/OM, Panasonic) fill serial, lens, shutter count, stabiliser,
ISO and colour temperature only when standard tags did not.
TiffTags gains GPSVersionID, GPSStatus, YCbCrPositioning,
StandardOutputSensitivity and ISOSpeed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Bounds-checked positional stream reads, an in-memory span reader,
Exif/XMP date parsing, and the MetadataBuilder accumulator that applies
every fallback chain (timestamps, ISO, serial/lens, orientation,
preview kinds) in one place. XMP is read with XmlReader (no DTDs, no
resolver, 4 MB cap) in both attribute and element forms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
The flow median filter runs the odd-even transposition network on
Vector<float> blocks in the interior (shared with the temporal median),
which removes the dominant cost of Best-quality flow. The stabiliser
analyses its nine patches in parallel with one correlator each and
gathers results in patch order, so output stays deterministic. Adds
brute-force reference tests for the median, box and binomial kernels.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Metadata steps are verified against the detrended measured jump; only
received steps are removed, the step-free series deflickered and the
steps restored as smoothstep transitions that lengthen until no frame
step exceeds 0.3 EV.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Horizon from the strongest column-averaged row step (median over
frames), refined per column by a dynamic-programming path and feathered;
Otsu luminance fallback with morphology; confidences for both. Region
deflicker fits sky and land separately and blends through the mask.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Robust LOWESS target curve with intentional exposure offsets removed,
blended by Tukey weight and surviving-window fraction; flicker spectrum
from the autocorrelation of detrended luminance recommends a window of
four periods.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Known-answer fixtures: brightness ramp, seeded global/sky/land flicker,
metadata and pixel exposure steps, horizon split, colour temperature
drift and white-balance jitter, star field, black/white/duplicate
frames, gaps and clock shifts. Covers frame statistics (CCT against
exact blackbody chromaticities) and the light analyzer with it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Cadence, jitter, gaps and sessions (small sessions merge), timestamp
sanity (duplicates, out-of-order frames, clock jumps incl. DST), EV
series and exposure steps, camera/lens identity, resolution and
orientation consistency, and NOAA day phases from GPS. Findings carry
stable resource keys plus positional arguments.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Every colour type and bit depth (1/2/4/8/16), PLTE + tRNS (palette alpha
and colour keys), Adam7, all five filters, zlib via ZLibStream; CRC
mismatches and truncated data are tolerated with warnings. PngEncoder
gains an additive ImageBuffer overload (8/16-bit sRGB with sRGB chunk)
and the Palette colour type for decoding.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Port libjpeg's resync_to_restart decision rules (skip bogus markers and
stale RSTs, account for lost intervals), reject oversubscribed Huffman
tables before filling the lookahead table (found by fuzzing), and keep
triangle chroma upsampling at 1/8 scale for better thumbnails.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
TemporalStacker keeps a sliding window of pooled frame copies and renders
an exact per-pixel median (vectorised odd-even transposition network up
to 32 frames, Hoare selection beyond) or mean. StarTrailStacker keeps a
progressive maximum with exponential decay; trail length in frames maps
to the decay that fades a trail to 1 %.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
FrameInterpolator synthesises a frame at fractional time t from the
bidirectional flow: of the two candidate velocities at each pixel (A's
flow, B's reversed flow) it keeps the one whose path is forward–backward
consistent through both frames, samples both frames with Catmull-Rom and
blends them by (1-t, t) weighted by each sample's agreement, so samples
occluded in the other frame drop out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Baseline, extended 12-bit, progressive and lossless (SOF3) Huffman
decoding; 1/3/4 components with Adobe CMYK/YCCK; all integral sampling
factors with libjpeg fancy upsampling; restart resynchronisation;
DCT-domain scaling. Full-size blocks use a port of libjpeg's ISLOW IDCT,
so output is bit-identical to djpeg on the cjpeg fixtures.
PixelConvert maps interleaved 8/16-bit samples to linear ImageBuffers and
back, gray/RGB, Exif orientation and clamping.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
BezierEasing evaluates CSS-style cubic-bezier timing curves (Newton with
bisection fallback) with linear, ease-in, ease-out and ease-in-out
presets. VirtualCamera interpolates centre, roll and geometric zoom
between keyframes and returns one view-to-source affine composed with
the inverse stabiliser correction.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC
Per frame, nine phase-correlation patches on a 3x3 grid are gated by
noise-compensated gradient energy (Immerkær noise estimate) and peak
sharpness, then fitted with a similarity by exhaustive two-point
hypotheses and weighted least squares on the inliers. Solve smooths the
camera path with Gaussian-weighted local linear regression, derives
corrections, shake RMS, recommendation (> 0.3 px) and the crop needed to
hide borders.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012iLMBNwBE2GnphAsu4cjaC