v0.4.1

The following release notes are automatically generated.

For the complete changelog, visit here.
If you encounter any problems, please report them on the issues page.

Changelog

The following are changelog from 0.4.0 that got yanked, 0.4.1 fix some build issues

added

• core.imgseqs.PNGWrite(..., output_dir=...) names each still after the source file (ImgSeqPath) and each animated frame after the writer's own frame index, creating the directory recursively. output_path and output_dir are mutually exclusive, and start_number applies only to output_path

build

• Bump wpd to 2e1400f00089a68c7939eb9593d5b49b1144b8e0
• Bump libheif to 1.23.5
  • Also, CI now pin the native builds to this version on every platform for consistency.
  • Drop embedded_libheif feature, which was only used for the Linux build and is now unnecessary.
• Enable SIMD stuff on zune-jpeg
• Bump jxl to 0.8.0
  • This changes some internal stuff to use Event-driven API for decoding animation.
• Bump dav1d to 1.5.4
• Replace jpeg2k and the OpenJPEG sources openjpeg-sys vendored with hayro-jpeg2000, a pure-Rust decoder.
  • No C toolchain and no build.rs on any platform, so the one dependency whose build failed without a diagnostic is gone.
  • openjpeg-COPYING.txt and OpenJPEG's entries leave THIRD_PARTY_NOTICES and LICENSES/, and the plugin links no OpenJPEG at all.
• 0.4.1: Actually bump the cargo version to 0.4.x, this was missed in the last release. The plugin manifest version is already 0.4.1, so this is just to keep the cargo version in sync.

changed

• JPEG 2000 frames are decoded by hayro-jpeg2000 now, at the depth, alpha and palette the file states.
  • Every frame of the 35 page corpus and of the jpeg 2000 fixtures is byte for byte what the OpenJPEG build produced: same planes, same formats, same _Primaries/_Transfer/_Matrix/_Range.
  • An sYCC page still comes out as its own yuv planes, and a page whose every component states one subsampling factor is refused by name instead of handed out at a size its pixels do not have.
  • The decode is faster than the OpenJPEG build: 7.8% at the default window and 7% serial over the 35 page corpus, level at prefetch=16. Peak memory is 1.5x at a deep window (3271 against 2139 MB at prefetch=16), because the decoder holds f32 components.

fixed

• Freeing a clip could hang when the lookahead pool was torn down while a worker sat between its stop check and its wait; the stop and its notification are made under the pool's own lock now

---

Following are the files included in this release:

platform	file
windows x64	windows-x86_64-plugin.zip
linux x64 (glibc)	linux-x86_64-plugin.zip
linux x64 (glibc), relinking sources	linux-relink-source.tar.gz
linux x64 (musl)	linux-musl-x86_64-plugin.zip
linux x64 (musl), relinking sources	linux-musl-relink-source.tar.gz
macos arm64	macos-arm64-plugin.zip

There are also the attached wheels, mirrored to here from PyPI, and a source
distribution to build from.

Please make sure to download the correct file for your system.

---

Attestation URL: https://github.com/noaione/vs-imageseqs/attestations/59540011

v0.3.0

The following release notes are automatically generated.

For the complete changelog, visit here.
If you encounter any problems, please report them on the issues page.

Changelog

added

• ImgSeqAnimationIndex names the displayed picture's position within its own file, and is written on frames from an animated file
• core.imgseqs.PNGWrite(clip, output_path=...) writes the frames a graph asks for as separate PNGs: a request writes that frame before it returns, and a frame nobody requests is never written, so creating the node writes nothing
  • automatically convert YUV clip to RGB following the frame's own _Matrix and _Range. Matrices are converted for non-constant-luminance ones, and a constant-luminance, SMPTE 2085 or chromaticity-derived frame is refused by name. The conversion is vectorized where the processor has one to run it on: AVX2 or AVX-512 on x86-64, and NEON on aarch64, which every aarch64 has.
  • a float clip is a value in [0, 1], which becomes a sixteen bit PNG or the word depth names: a value outside the range is clamped and a NaN is written as the smallest sample
  • depth names the word the PNG stores. 8 gives up precision by rounding, 16 widens an eight or ten bit frame exactly, and 1, 2 or 4 pack a gray frame, which writes back the picture a low bit reader expanded. A frame of nine to fifteen bits is stored as a sixteen bit PNG whose high bits hold the source precision, with an sBIT chunk saying how many are meaningful, so the widening is exact rather than a scale
  • output_path is a filename for a one frame clip or a %d/%06d template, numbered by the writer's own frame index plus start_number rather than any index a trim or a splice moved. compression is 0 to 9 and lossless at every level, always_save writes a frame again instead of skipping one this instance already published, and overwrite is the separate permission to replace a destination that is already there
  • a write goes to a temporary file beside the destination and is published with a rename or a linking create, so a failure leaves no partial file and two writers cannot both claim a path. ImgSeqPNGWritePath, ImgSeqPNGWriteSaved and ImgSeqPNGWritePerformed report what happened, and cICP and icc_profile carry the colour the frame states rather than inventing one
  • a consumer that keeps several frames in flight (frames(prefetch=n), open get_frame_async futures, vspipe) gets several encodes at once, and zlib-rs compresses 1.46x faster than flate2's default fallback into 7.5% smaller files, which at the default level is Pillow's own size for the same picture

changed

• The image crate is gone. Every format this plugin reads has a reader of its own under src/formats/, and a file whose bytes and extension both name no format here is refused with no reader here knows its format. The plugin is about 435 KB smaller, and several of those readers are ports of that crate's, so its licence texts and notices stay
  • webp: Now decoded with wpd, a much faster decoder. The still path is 1.2x to 1.3x faster, the Windows library 134 KB smaller, a corrupt file is refused with the decoder's own words, and building on x86 or x86-64 still needs nasm on PATH
  • tiff: a WebP- or JPEG-compressed strip is decoded here, and a compression code this reader does not take is refused by name at the probe
  • tiff: one, two or four bit gray, gray plus alpha, cmyk (with alpha for the first time), palette and ycbcr pages are all read now, and a ycbcr page comes out as its own yuv planes with _Matrix and _Range from its tags
  • tiff: a page that states an orientation, carries an ICC profile or is a BigTIFF now reports all three
  • bmp/ico: the older BITMAPCOREHEADER and a bare .dib are read, and a Windows cursor is read instead of having its type word mistaken for a bit depth
  • tga: a colour map whose entries are fifteen or sixteen bits is read rather than refused, widening by the same round-to-nearest table a direct sixteen bit image uses
  • hdr: a radiance picture whose resolution line states an orientation other than the common -Y ... +X decodes
  • exr: a first part with no colour channel decodes, and a flat grayscale layer is read with Y as colour and A as the alpha clip
  • jp2: a file that states what its components mean is read as what it states -- gray or r,g,b beside an opacity component -- and a palette is expanded here rather than refused
  • png: a file is identified by its bytes on the probe side too, so a name that lies about it gets the same description and the same cICP properties
  • gif: a one frame gif is read as a still image
• Whether a file plays a timeline is decided by its bytes now, like its format already was, so a gif, webp, avif or heic under a name that does not say so plays its timeline. identify::route decides that and whose file it is in one call, so a file is opened once to identify it rather than about fifteen times, at a cost of 3% more per file at clip creation over the fixture set
• An avif or heif is read by the library its container states, through one walk the probe and the decode share
  • An item written as several extents is joined here, a grid of tiles goes to libheif, and an r,g,b page is read by libheif too; one with no alpha now reports ImgSeqOriginalColorType=rgb8
  • A sequence honours the ctts composition offsets and an edit list that ends its media early, clamping a sample composed before zero and refusing by name an edit list that starts partway into the media
  • A twelve bit subsampled page is handed out as YUV420P12 or YUV422P12
• A PAM whose MAXVAL is above a byte but which names no TUPLTYPE is read at the width its MAXVAL states
• Replace libwebp with wpd for WebP decoding.

fixed

• ImgSeqIndex reports the path's position in files rather than the output frame number, which an animation pushed out of step for every file after it
• A signed JPEG 2000 page, or one whose components state two widths, is refused with a message that names the component, its width and what is wrong with it
• The musllinux wheel no longer crashes on a TIFF (or an AVIF with no picture) read on a VapourSynth worker thread: each format's decode and stream is a call of its own, so the dispatcher no longer grows a 100 KiB frame past musl's 128 KiB stack
• Windows builds enable libheif's built-in dav1d decoder, so animated AVIF sequences decode instead of failing with NoMatchingDecoderInstalled; libheif's default features stay disabled
• An animated png's delays are placed exactly, at the lowest common denominator of the delays the file states, rather than rounded to whole milliseconds. An animation whose length is not a whole number of output frames keeps its last frame, so 600 ms at 24 fps is 15 frames rather than 14

performance

• Creating a clip over a long list reads the front of each file where the container states what a description needs there: 35 webp files of 146 MiB went from 99 ms to 6 ms, 35 jpeg xl from 157 ms to 5 ms, 35 jpeg 2000 from 159 ms to 4 ms, and 35 avif plus 35 heic of 362 MiB from 84 and 170 ms to 6 and 13 ms
• Describing an animated png reads the delays its fcTL chunks state instead of rendering every frame: five 1024x1024 files of eight frames each went from 49 ms to 3 ms
• PNG decoding hands each row to the frame it belongs in rather than buffering the whole picture, and a palette page is expanded here rather than by the decoder. Every png set measured is now ahead of Pillow's own decode column, from 0.89x to 0.69x of its time
• A netpbm, a targa and a bitmap are read a row at a time rather than into a buffer the size of the file, holding 48 MiB rather than 74 MiB for a 3000x3000 image and decoding 7 to 14% faster
• A planar tiff is handed to the frame as the planes it already is, which is 43 ms rather than 139 ms on a 3000x3000 eight bit page
• A DirectDraw surface whose width or height is not a multiple of four is read with the pixels that hang over the edge dropped, and a netpbm whose header is longer than the readable window is read to a megabyte

build

• the Linux release includes a musllinux wheel beside the manylinux one, for hosts whose C library is musl rather than glibc, where the manylinux wheel cannot load at all
  • it carries the same three plugin libraries, so an AVX2 or AVX-512 host is served there the same way
  • auditwheel's musl policy promises the host only libc and libz, so it bundles the C++ runtime the embedded libheif needs instead of needing a package Alpine does not install by default
  • it is accompanied by linux-musl-relink-source.tar.gz, the corresponding source archive the manylinux wheel already has
• x86-64 wheels carry one plugin library per microarchitecture level (avx2 and avx512) and VapourSynth loads the widest the CPU supports, which is 6% to 11% faster on the PNG and JPEG sets measured at the cost of a 10.4 MiB Windows wheel instead of 3.8 MiB

---

Following are the files included in this release:

platform	file
windows x64	windows-x86_64-plugin.zip
linux x64 (glibc)	linux-x86_64-plugin.zip
linux x64 (glibc), relinking sources	linux-relink-source.tar.gz
linux x64 (musl)	linux-musl-x86_64-plugin.zip
linux x64 (musl), relinking sources	linux-musl-relink-source.tar.gz
macos arm64	macos-arm64-plugin.zip

There are also the attached wheels, mirrored to here from PyPI, and a source
distribution to build from.

Please make sure to download the correct file for your system.

---

Attestation URL: https://github.com/noaione/vs-imageseqs/attestations/55373925

v0.2.1

The following release notes are automatically generated.

For the complete changelog, visit here.
If you encounter any problems, please report them on the issues page.

Changelog

quick small release to make macOS wheels much easier to use

build

• fix wrong version number in the plugin manifest itself.
• macOS arm64 wheels and standalone plugin ZIPs include dav1d and libde265, so
  users do not need to install those runtime libraries separately; they target
  macOS 11 or later

---

Following are the files included in this release:

platform	file
windows x64	windows-x86_64-plugin.zip
linux x64	linux-x86_64-plugin.zip
linux x64, relinking sources	linux-relink-source.tar.gz
macos arm64	macos-arm64-plugin.zip

There is also an attached wheel, mirrored to here from PyPI, and a source
distribution to build from.

Please make sure to download the correct file for your system.

---

Attestation URL: https://github.com/noaione/vs-imageseqs/attestations/51319314

v0.2.0

The following release notes are automatically generated.

For the complete changelog, visit here.
If you encounter any problems, please report them on the issues page.

Changelog

the animated images support

added

• animated GIF, APNG, WebP, JPEG XL, AVIF and HEIF/HEIC inputs now play their
  displayed timeline instead of contributing one frame
  • each output frame shows the picture the file displays at that instant, sampled
    onto the clip's own fpsnum/fpsden, so a fractional rate and a delay that
    does not divide it both land where the file says
  • a picture with no delay is held for one output tick rather than dropped, and a
    file's loop count is ignored: each listed path plays once
  • a 16-bit APNG stays 16-bit, and a HEIF or AVIF sequence is cropped to the
    picture it presents rather than handed out as coded
  • still images keep their frame count, format, pixels and metadata whether or
    not they share a list with animations
  • a colour-only read of a HEIF sequence still decodes its alpha track, because
    the library gives no way to skip it

fixed

• a frame request on an AVIF whose item holds no coded frame no longer hangs
  forever; it fails with a decode error naming the file
  • the item decoder now runs at low latency, which is what lets "the decoder has
    all of the item and has no picture" be told apart from "the picture is still
    being decoded"
• an AVIF or HEIF/HEIC that states its orientation as the irot and imir item
  properties is now handed out the way the file describes it
  • ImgSeqOrientation reports the equivalent EXIF code instead of 1 for such
    a file, and apply_rotation=False hands the stored picture back at the
    stored size, which neither container could do before
  • a container transform is the normative statement for these two formats, so an
    EXIF tag beside one stays informational and is not applied on top of it
• an AVIF whose item is split over several extents now decodes instead of failing
  every frame request
  • the container reader does not join extents, so such a file is described as the
    format the fallback decoder produces rather than as the YUV its samples are;
    that decoder does join them
• a malformed AVIF container is refused instead of being read past its own bounds
  • an extent the file does not hold is no longer allocated before the read that
    fails on it, an iloc field wider than an address no longer shifts a value
    out of its type, and a box with an extended size header is read with the whole
    header rather than the last eight bytes of it being dropped
• a colour-only read of an AVIF no longer fails on a file whose alpha item is
  broken
  • the alpha item is a coded item of its own and is not part of what Read hands
    out, so it is not read; ReadAlpha still reports the problem on the same file

performance

• AVIF decoding is 15.6% faster on the 35-page sandbox set, and 43% faster on
  sandbox/hitokage-sample: a still image is one frame, so the item is decoded
  directly instead of through a frame-delay pipeline it cannot use
• reading a container's orientation costs 0.006 ms per AVIF file and 0.037 ms per
  HEIF file at clip creation, and nothing per frame
• bounding an AVIF item's extent costs one file metadata call per file at clip
  creation, 0.005 ms per file, and nothing per frame
• a read that hands out no alpha clip no longer decodes an AVIF's alpha item,
  which is a coded item of its own: 11–20% off a colour-only read of an
  alpha-bearing page, with the colour planes byte identical
• a probe no longer keeps a file's embedded ICC profile unless icc_profile=True
  asks for it, which is 36 MiB off a 35-file clip whose files each carry a 1 MiB
  profile; ImgSeqHasICC is unchanged
• a still image reads as fast as it did before, on every sandbox set and at every
  lookahead depth, so animations cost the still path nothing

build

• improvement to linux wheel distribution, manylinux now has dropped down to glibc 2.28
  • linux build now bundles libdav1d.so.7 and libde265.so.0 to avoid runtime dependency issues
• change the plugin layout to become plugins/imageseqs/ with a manifest.vs file, and the plugin loader path to $ORIGIN/lib

---

Following are the files included in this release:

platform	file
windows x64	windows-x86_64-plugin.zip
linux x64	linux-x86_64-plugin.zip
linux x64, relinking sources	linux-relink-source.tar.gz
macos arm64	macos-arm64-plugin.zip

There is also an attached wheel, mirrored to here from PyPI, and a source
distribution to build from.

Please make sure to download the correct file for your system.

---

Attestation URL: https://github.com/noaione/vs-imageseqs/attestations/51298750

v0.1.0

The following release notes are automatically generated.

For the complete changelog, visit here.
If you encounter any problems, please report them on the issues page.

Changelog

the first release of vapoursynth-imageseqs, a native image-sequence source
for VapourSynth.

added

• Read for ordered image sequences and ReadAlpha for separate color and
  alpha clips.
• support for PNG, JPEG, WebP, AVIF, HEIF/HEIC, JPEG XL, JPEG 2000, DDS,
  farbfeld, TIFF, EXR, HDR, PNM, QOI, TGA, BMP, GIF, and ICO.
• native YUV output for compatible WebP, HEIF/HEIC, AVIF, and JPEG 2000
  sources.
• nominal 9–15-bit formats, including 10-bit and 12-bit gray, RGB, alpha,
  and YUV output.
• EXIF and codestream orientation handling through apply_rotation.
• CICP and container color metadata as VapourSynth frame properties.
• optional embedded ICC profile export through the binary ICCProfile frame
  property.
• variable-size and variable-format sequences through mismatch.
• background prefetching with a bounded memory budget.
• debug timing logs for probing, decoding, frame writing, and properties.

performance

• native libwebp decoding and direct lossy WebP YUV output.
• direct HEIF and AVIF plane output where the container metadata permits it.
• JPEG XL and JPEG 2000 header-only probing before frame requests.
• frame construction moved into the prefetch workers to reduce request-thread
  work.

---

Following are the files included in this release:

platform	file
windows x64	windows-x86_64-plugin.zip
linux x64	linux-x86_64-plugin.zip
macos arm64	macos-arm64-plugin.zip

There is also an attached wheels file, mirrored to here. This is the same as the one from PyPI.

Please make sure to download the correct file for your system.

---

Attestation URL: https://github.com/noaione/vs-imageseqs/attestations/48708131