v1.4.0

[1.4.0] — 2026-06-10

Feature release: high bit depth and additional chroma samplings, plus the
fixes from a full audit. 8-bit 4:2:0 output with default parameters is
unchanged (all 198 pre-existing golden hashes reproduce bit-identically);
diMode=1 scene-change frames and diMode=2 edge columns intentionally
changed toward Avisynth parity (see Fixed).

Added

• 10/12/16-bit integer YUV input (u16 storage). IVTC decisions
  (field matching, decimation, scene logic) are bit-depth-deterministic:
  the decision kernels round at the 8-bit map level (pavgbScore), so a
  shifted 8-bit clip decides identically at any depth, while output
  pixels keep full precision. No external oracle exists for HBD
  (upstream is YV12-only); guarded by cross-bit-depth consistency tests.
• 4:2:2 and 4:4:4 chroma sampling across all kernels (4:1:1/4:4:0
  remain rejected).
• Real-content robustness + bit-depth-consistency harness
  (scripts/test_real_content.py).
• threshold parameter is now actually wired (upstream accepted it but
  hardcoded AdjPara(20); identical behaviour at the default 20).
• Golden coverage for the previously oracle-free axes: ref
  BOTTOM/ALL/NONE, diMode 0/1/2, blend 0/1 (new motion_flicker
  fixture — the first whose blocks actually pass the blend motion gate);
  compCn unit tests incl. the upstream uppercase-'N' quirk; the oracle
  parameter grid now lives in one module (scripts/param_grid.py).

Fixed

• 4:2:0 heights with height % 4 == 2 (e.g. 720×486) are now
  rejected. The field-interleaved chroma row mapping requires an even
  chroma-plane height; such clips previously caused out-of-bounds chroma
  reads on the default path, an out-of-bounds write in the blend path,
  and an uninitialized final chroma row (visible as cross-plane pixel
  contamination). Upstream has the identical latent bug, so nothing
  bit-compatible is lost.
• diMode=1 now always runs the full deinterlacer (Avisynth parity:
  only SIMPLE_BLUR/ONE_FIELD have the DrawPrevFrame scene-change
  shortcut). Previously a scene-change frame emitted a copy of the
  previous frame's matched fields instead.
• blend=1 is access-order independent: the blend kernel's reach into
  the previous block (base-1, weight 37/256) is now decided before
  rendering, so seeking produces the same pixels as linear playback.
• diMode=2 edge columns match upstream: the 3-tap motion test now
  replicates upstream's cross-row pointer wraps at x=0/x=w−1 instead of
  substituting 0.
• The 1.3.2 "Known issue" (diMode=1 SIMD pavgb vs scalar truncated
  average) was already fixed in the 4:2:2/4:4:4 work; the SIMD and
  scalar paths are verified identical across all kernels.
• Hardening: threshold/pthreshold validated to [0, 100000]
  (previously overflowed i32 internally in release builds); VFR clips
  and fps=24 on <2-frame clips rejected with clear errors; height capped
  at 8192; out-of-nominal-range 10/12-bit samples saturate instead of
  invoking UB; allocation-failure paths now set a proper filter error
  instead of returning garbage frames.

CI / packaging

• The integration suite (golden hashes, upstream comparison, determinism,
  bit depth) is now gating — it previously ran with a job-level
  continue-on-error and could never fail CI; a missing plugin now fails
  the run instead of silently skipping every test.
• Releases run the full test gate (Debug + ReleaseFast unit tests +
  integration suite against the ReleaseFast artifact) before anything is
  published; unit tests also run under ReleaseFast in CI.
• PyPI wheels now carry the full manylinux compatibility alias tags in
  WHEEL (the comment claimed it; the code wrote only the first alias).
• GitHub release bodies use the latest changelog section, not the whole
  file; pytest collection pinned to tests/integration.

v1.3.3

Changelog

All notable changes to this project are documented here. Format loosely
follows Keep a Changelog, versioning is
SemVer.

[1.3.3] — 2026-05-24

Packaging-only release. Same Binaries.

Fixed

• PyPI wheel should install into site-packages/vapoursynth/plugins/

[1.3.2] — 2026-05-24

Patch release. Pixel output unchanged from 1.3.1 on every supported
parameter combination — verified bit-identical against the
vapoursynth-cpp-api4 reference across the full upstream-compare grid
(8 fixtures × parameter combos) and across the 3-mode smoke
(default / ref=ALL / blend=1, 150 frames each, sha256 prefix
63e6227182560887 unchanged).

Performance

The hot algorithm functions (motion.makeMotionMap2{Min,Max},
output.{copyCPNField, deinterlace, simpleBlur, deintOneField}) now
take their plane-view arguments as *const plane.PlaneView instead of
by value. Each PlaneView is 48 bytes; the 24fps path threads several
through several calls per output frame, so the reduction in stack
copies shows up clearly:

Pipeline	1.3.1	1.3.2	Δ
fps=24 (default)	2395 fps	2701 fps	+12.8%
fps=24 + blend	2391 fps	2699 fps	+12.9%
fps=24 ref=ALL	2490 fps	2795 fps	+12.3%
fps=30 diMode=1	716 fps	715 fps	—
fps=30 diMode=2	3409 fps	3447 fps	+1.1%

(720x480 NTSC, 10 000 measured frames per run × 3 runs, taskset -c 1,
ReleaseFast.) The 30fps modes don't move because they do one
makeOutput per output frame; the 24fps path runs getFrameSub 5×
per output frame in the decimation block, which is where the saved
copies add up.

Internal restructuring (no behaviour change)

• Bindings: dropped the hand-rolled @cImport layer in src/c.zig
  and the three vendored VS4 headers in favour of the upstream
  vapoursynth-zig
  package. Filter + plugin entry now use ZAPI / ZMap helpers with
  type-safe enums and typed prop setters (setFieldBased,
  setCombed, setDuration*). Net diff ~−1450 lines.
• PlaneView / PlaneViewMut introduced in plane.zig — one
  canonical 6-field struct replaces the 18–28 separate
  [*]const u8 + usize parameters the algorithm functions used to
  take.
• scalar.zig module for tiny u8-math helpers (absDiff,
  pavgb, subSat) that were previously redefined per-file (6
  inline copies across 5 files). Vector versions stay in simd.zig.
• Deinterlace dedupe: extracted
  output.deinterlacePixelScalar — one canonical inline kernel for
  the per-pixel scoring + pick + motion-override that was previously
  written twice in output.deinterlace (~75 lines inside the SIMD
  body's chroma loop + ~109 lines in the scalar tail).
• Audit cleanup: dead fields (CFrameInfo.{out, matchAcc},
  CallState.{realFrame, iUsePrev, iUseNext}) removed; redundant
  per-frame @memset calls in CallState.resetForFrame dropped
  (consumers either fully overwrite or pair partial writes with
  matching partial reads); 12 var ... = undefined plane-row locals
  in simpleBlur paths became const expressions.
• Centralised constants: MAX_WIDTH (3 copies) and
  CHROMA_LANES (2 copies) consolidated into plane.zig.

Tooling

• scripts/bench.py — wall-clock fps benchmark across the
  relevant filter configurations. Used to validate the
  PlaneView-by-pointer change above; sub-1% spread on SIMD-bound
  paths once pinned with taskset -c 1.
• CI workflow (.github/workflows/ci.yml) — consolidated three
  duplicated curl | tar Zig installs into mlugg/setup-zig@v2;
  collapsed the lint/unit/cross jobs into a single build job;
  replaced the PPA-based VapourSynth install on the integration job
  with pip install VapourSynth (the PyPI wheel ships
  libvapoursynth.so + vspipe on Linux). −50 lines.
• Release workflow (.github/workflows/release.yml) — same
  mlugg/setup-zig@v2 swap; tag-resolve collapsed to a one-liner via
  inputs.tag || github.ref_name; the
  gh release view || create; upload --clobber; edit choreography
  became a single softprops/action-gh-release@v2 step. −20+ lines.

Known issue (carried forward, not new)

The SIMD body of output.deinterlace (diMode=1) uses pavgb (rounded
average) for the motion-override luma override, while the scalar tail
and the Avisynth C upstream both use truncated (t+b)>>1. At pixels
where the motion-override fires and (pT[x]+pB[x]) is odd, our SIMD
path differs by ±1 from the scalar path. The vapoursynth-cpp-api4
reference plugin hardcodes one_field (no diMode parameter), so
this can't yet be validated against ground truth — a TODO marker in
the source records the one-liner fix for whoever builds the Avisynth
oracle (or hand-computes reference fixtures).

[1.3.1] — 2026-05-23

Patch release. Performance-only — pixel output unchanged from 1.3.0
on every supported parameter combination.

Changes

• SIMD coverage extended to the three output-stage functions in
  src/output.zig:
  • simpleBlur (diMode=2): 32-lane motion-hit count + 16-lane luma
    body with overlap-loaded 3-tap motion neighbours.
  • deintOneField (diMode=3, default): 16-lane luma body using
    cross-row field-map OR-mask + pavgb + @select; chroma
    unconditional via tight LANES/2 pass.
  • deinterlace (diMode=1): 32-lane luma with all five IV scores +
    @select-chain min-pick + motion override.
• ARM64 binaries included by default (Linux aarch64 + macOS
  aarch64) since the 1.3.0 release-workflow rework.

Speed delta vs 1.3.0 (720x480 NTSC, ReleaseFast, best-of-3)

Pipeline	1.3.0	1.3.1	Δ
fps=30	3217 fps	3175 fps	~0 (noise)
fps=24	3148 fps	3023 fps	~0
fps=24 diMode=1	3041 fps	3215 fps	+5.7%
fps=24 blend=1	2353 fps	2301 fps	~0

The default fps=24/fps=30 paths don't move measurably because most
frames on telecined NTSC content classify as ip='P' and skip the
deinterlacer entirely. Workloads that hit ip='I' frequently
(interlaced-heavy content, or pthreshold=1-forced runs) see the
diMode=1 gain on every frame.

Bit-exactness

Default-path 720-frame regression against the API-4 upstream port
still passes byte-for-byte. diMode=1 SIMD was independently
verified against the pre-SIMD scalar build (80 frames, all md5s
identical).

[1.3.0] — 2026-05-23

Initial release. A from-scratch Zig port of the VapourSynth-IT
inverse-telecine filter, bit-exact to the upstream C reference path
and with the original Avisynth-only parameters restored.

Version numbering picks up the IT lineage:

• Avisynth IT_YV12 v0.1.03 (minamina, 2003)
• VapourSynth VS_IT.dll v0103.1.2 (msg7086, 2014)
• This Zig port: v1.3.0.

Features

• Filter core.zit.IT(clip, fps=24, threshold=20, pthreshold=75, ref="TOP", blend=0, diMode=3) registered for VapourSynth API 4
  (R55+). Accepts YUV420P8 input with width % 16 == 0, height % 2 == 0,
  width ≤ 8192.
• Decimation modes:
  • fps=24 — 3:2-pulldown removal, output 24000/1001 fps (5→4
    frames per cycle).
  • fps=30 — field-matching only, input rate preserved.
• ref parameter ("TOP"/"BOTTOM"/"ALL"/"NONE"): the
  Avisynth original's field-match-direction switch, fully reimplemented.
  The VapourSynth upstream had stripped this down to a hardcoded
  "TOP".
• blend parameter: pure-Zig port of Avisynth's BlendFrame_YV12,
  triangular kernel weighted blend of post-matched frames, motion-gated
  on the same minD/avgD heuristic as the original. Only active when
  fps=24.
• diMode parameter with all four Avisynth deinterlace modes:
  • 0 DI_MODE_NONE — straight field copy, no deinterlace.
  • 1 DI_MODE_DEINTERLACE — full per-pixel five-candidate scorer
    with motion-gated vertical-average fallback (~500 LoC of Avisynth
    MMX re-expressed as pure scalar Zig).
  • 2 DI_MODE_SIMPLE_BLUR — vertical (T+2C+B)/4 blur on
    motion-flagged pixels.
  • 3 DI_MODE_ONE_FIELD — default; field-interpolation using the
    simple-blur and motion2max maps.

Frame properties

• Standard, always set on output: _FieldBased=0, _Combed per
  frame, _DurationNum/_DurationDen derived from output rate. All
  source props (_Matrix, _Transfer, _Primaries,
  _ChromaLocation, _SARNum/Den, _Range/_ColorRange, …)
  inherited via propSrc in newVideoFrame.
• Diagnostic (per-frame inspection for power-user scripts):
  ITMatch, ITMflag, ITIpFlag (utf8 1-char each); ITIvC/P/N/M,
  ITDiffP0/P1/S0/S1, ITBlended (ints).

Verification

• Bit-exact to a mechanically API-3 → API-4-ported build of the
  upstream C reference (reference/vapoursynth-cpp-api4/). 198 frames
  across 10 fixture × parameter combinations, plus 720 frames of two
  real telecined NTSC VOB samples in both fps=24 and fps=30 mode,
  byte-for-byte identical.
• 35 Zig unit tests for the algorithm primitives + SIMD helpers.
• 58 Python/VapourSynth integration tests — invariants,
  property checks, error paths, the regression-pinned md5 hashes, the
  upstream-compare matrix, and the new parameter sweeps.

Performance

500 frames of 720×480 NTSC, ReleaseFast build, single-threaded
fmParallelRequests:

Pipeline	fps
core.zit.IT(clip, fps=30)	~2700
core.zit.IT(clip, fps=24)	~2200
core.zit.IT(clip, fps=24, diMode=1)	~2350
core.zit.IT(clip, fps=24, blend=1)	~1700
core.vivtc.VFM(clip) (reference)	~600
core.vivtc.VFM + VDecimate (ref)	~1270

~4× faster than vivtc.VFM a...

v1.3.2

Changelog

All notable changes to this project are documented here. Format loosely
follows Keep a Changelog, versioning is
SemVer.

[1.3.2] — 2026-05-24

Performance

The hot algorithm functions (motion.makeMotionMap2{Min,Max},
output.{copyCPNField, deinterlace, simpleBlur, deintOneField}) now
take their plane-view arguments as *const plane.PlaneView instead of
by value. Each PlaneView is 48 bytes; the 24fps path threads several
through several calls per output frame, so the reduction in stack
copies shows up clearly:

Pipeline	1.3.1	1.3.2	Δ
fps=24 (default)	2395 fps	2701 fps	+12.8%
fps=24 + blend	2391 fps	2699 fps	+12.9%
fps=24 ref=ALL	2490 fps	2795 fps	+12.3%
fps=30 diMode=1	716 fps	715 fps	—
fps=30 diMode=2	3409 fps	3447 fps	+1.1%

(720x480 NTSC, 10 000 measured frames per run × 3 runs, taskset -c 1,
ReleaseFast.) The 30fps modes don't move because they do one
makeOutput per output frame; the 24fps path runs getFrameSub 5×
per output frame in the decimation block, which is where the saved
copies add up.

Internal restructuring (no behaviour change)

• Bindings: dropped the hand-rolled @cImport layer in src/c.zig
  and the three vendored VS4 headers in favour of the upstream
  vapoursynth-zig
  package. Filter + plugin entry now use ZAPI / ZMap helpers with
  type-safe enums and typed prop setters (setFieldBased,
  setCombed, setDuration*). Net diff ~−1450 lines.
• PlaneView / PlaneViewMut introduced in plane.zig — one
  canonical 6-field struct replaces the 18–28 separate
  [*]const u8 + usize parameters the algorithm functions used to
  take.
• scalar.zig module for tiny u8-math helpers (absDiff,
  pavgb, subSat) that were previously redefined per-file (6
  inline copies across 5 files). Vector versions stay in simd.zig.
• Deinterlace dedupe: extracted
  output.deinterlacePixelScalar — one canonical inline kernel for
  the per-pixel scoring + pick + motion-override that was previously
  written twice in output.deinterlace (~75 lines inside the SIMD
  body's chroma loop + ~109 lines in the scalar tail).
• Audit cleanup: dead fields (CFrameInfo.{out, matchAcc},
  CallState.{realFrame, iUsePrev, iUseNext}) removed; redundant
  per-frame @memset calls in CallState.resetForFrame dropped
  (consumers either fully overwrite or pair partial writes with
  matching partial reads); 12 var ... = undefined plane-row locals
  in simpleBlur paths became const expressions.
• Centralised constants: MAX_WIDTH (3 copies) and
  CHROMA_LANES (2 copies) consolidated into plane.zig.

Known issue (carried forward, not new)

The SIMD body of output.deinterlace (diMode=1) uses pavgb (rounded
average) for the motion-override luma override, while the scalar tail
and the Avisynth C upstream both use truncated (t+b)>>1. At pixels
where the motion-override fires and (pT[x]+pB[x]) is odd, our SIMD
path differs by ±1 from the scalar path. The vapoursynth-cpp-api4
reference plugin hardcodes one_field (no diMode parameter), so
this can't yet be validated against ground truth — a TODO marker in
the source records the one-liner fix for whoever builds the Avisynth
oracle (or hand-computes reference fixtures).

v1.3.1

Changelog

All notable changes to this project are documented here. Format loosely
follows Keep a Changelog, versioning is
SemVer.

[1.3.1] — 2026-05-23

Patch release. Performance-only — pixel output unchanged from 1.3.0
on every supported parameter combination.

Changes

• SIMD coverage extended to the three output-stage functions in
  src/output.zig:
  • simpleBlur (diMode=2): 32-lane motion-hit count + 16-lane luma
    body with overlap-loaded 3-tap motion neighbours.
  • deintOneField (diMode=3, default): 16-lane luma body using
    cross-row field-map OR-mask + pavgb + @select; chroma
    unconditional via tight LANES/2 pass.
  • deinterlace (diMode=1): 32-lane luma with all five IV scores +
    @select-chain min-pick + motion override.
• ARM64 binaries included by default (Linux aarch64 + macOS
  aarch64) since the 1.3.0 release-workflow rework.

Speed delta vs 1.3.0 (720x480 NTSC, ReleaseFast, best-of-3)

Pipeline	1.3.0	1.3.1	Δ
fps=30	3217 fps	3175 fps	~0 (noise)
fps=24	3148 fps	3023 fps	~0
fps=24 diMode=1	3041 fps	3215 fps	+5.7%
fps=24 blend=1	2353 fps	2301 fps	~0

The default fps=24/fps=30 paths don't move measurably because most
frames on telecined NTSC content classify as ip='P' and skip the
deinterlacer entirely. Workloads that hit ip='I' frequently
(interlaced-heavy content, or pthreshold=1-forced runs) see the
diMode=1 gain on every frame.

Bit-exactness

Default-path 720-frame regression against the API-4 upstream port
still passes byte-for-byte. diMode=1 SIMD was independently
verified against the pre-SIMD scalar build (80 frames, all md5s
identical).

[1.3.0] — 2026-05-23

Initial release. A from-scratch Zig port of the VapourSynth-IT
inverse-telecine filter, bit-exact to the upstream C reference path
and with the original Avisynth-only parameters restored.

Version numbering picks up the IT lineage:

• Avisynth IT_YV12 v0.1.03 (minamina, 2003)
• VapourSynth VS_IT.dll v0103.1.2 (msg7086, 2014)
• This Zig port: v1.3.0.

Features

• Filter core.zit.IT(clip, fps=24, threshold=20, pthreshold=75, ref="TOP", blend=0, diMode=3) registered for VapourSynth API 4
  (R55+). Accepts YUV420P8 input with width % 16 == 0, height % 2 == 0,
  width ≤ 8192.
• Decimation modes:
  • fps=24 — 3:2-pulldown removal, output 24000/1001 fps (5→4
    frames per cycle).
  • fps=30 — field-matching only, input rate preserved.
• ref parameter ("TOP"/"BOTTOM"/"ALL"/"NONE"): the
  Avisynth original's field-match-direction switch, fully reimplemented.
  The VapourSynth upstream had stripped this down to a hardcoded
  "TOP".
• blend parameter: pure-Zig port of Avisynth's BlendFrame_YV12,
  triangular kernel weighted blend of post-matched frames, motion-gated
  on the same minD/avgD heuristic as the original. Only active when
  fps=24.
• diMode parameter with all four Avisynth deinterlace modes:
  • 0 DI_MODE_NONE — straight field copy, no deinterlace.
  • 1 DI_MODE_DEINTERLACE — full per-pixel five-candidate scorer
    with motion-gated vertical-average fallback (~500 LoC of Avisynth
    MMX re-expressed as pure scalar Zig).
  • 2 DI_MODE_SIMPLE_BLUR — vertical (T+2C+B)/4 blur on
    motion-flagged pixels.
  • 3 DI_MODE_ONE_FIELD — default; field-interpolation using the
    simple-blur and motion2max maps.

Frame properties

• Standard, always set on output: _FieldBased=0, _Combed per
  frame, _DurationNum/_DurationDen derived from output rate. All
  source props (_Matrix, _Transfer, _Primaries,
  _ChromaLocation, _SARNum/Den, _Range/_ColorRange, …)
  inherited via propSrc in newVideoFrame.
• Diagnostic (per-frame inspection for power-user scripts):
  ITMatch, ITMflag, ITIpFlag (utf8 1-char each); ITIvC/P/N/M,
  ITDiffP0/P1/S0/S1, ITBlended (ints).

Verification

• Bit-exact to a mechanically API-3 → API-4-ported build of the
  upstream C reference (reference/vapoursynth-cpp-api4/). 198 frames
  across 10 fixture × parameter combinations, plus 720 frames of two
  real telecined NTSC VOB samples in both fps=24 and fps=30 mode,
  byte-for-byte identical.
• 35 Zig unit tests for the algorithm primitives + SIMD helpers.
• 58 Python/VapourSynth integration tests — invariants,
  property checks, error paths, the regression-pinned md5 hashes, the
  upstream-compare matrix, and the new parameter sweeps.

Performance

500 frames of 720×480 NTSC, ReleaseFast build, single-threaded
fmParallelRequests:

Pipeline	fps
core.zit.IT(clip, fps=30)	~2700
core.zit.IT(clip, fps=24)	~2200
core.zit.IT(clip, fps=24, diMode=1)	~2350
core.zit.IT(clip, fps=24, blend=1)	~1700
core.vivtc.VFM(clip) (reference)	~600
core.vivtc.VFM + VDecimate (ref)	~1270

~4× faster than vivtc.VFM at field-matching, ~2× faster than
vivtc.VFM + VDecimate at IVTC, using the pure-C path + portable
Zig @Vector(N, u8) SIMD (pavgb, psubusb, pmaxub/pminub
equivalents).

Differences from the VapourSynth upstream

The upstream
VapourSynth-IT
plugin (msg7086, 2014) was a partial port of the Avisynth original.
This port reintroduces what was stripped and fixes four framework-level
issues:

1. ref, blend, diMode=0/1/2 reinstated (upstream hardcoded
   ref="TOP", blend=false, diMode=3).
2. Frame properties: upstream passes propSrc=null to
   newVideoFrame, so output frames carry no metadata at all. We
   inherit source props and set _FieldBased/_Combed/_Duration*
   explicitly.
3. Threading mode: fmParallelRequests. Upstream registered as
   fmParallel despite sharing mutable per-instance state — a latent
   race condition the modern VS scheduler hits more often.
4. Frame-request range widened to [base-2, base+6] (fps=24) /
   [n-2, n+2] (fps=30) so all reads through getFrameFilter are
   satisfiable. Upstream relied on API 3's sync getFrame to pull
   neighbours from cache; under API 4 that is not allowed inside a
   filter callback.
5. Edge clamping: getFrameFilter(n, …) is called with n clipped
   to [0, numFrames-1]. Under API 3 the core silently clamped; under
   API 4 it returns null and a deref crashes.

Build & packaging

• zig build --release=fast produces zig-out/lib/libzit.so on the
  host platform.
• zig build cross produces release binaries for all three targets
  under zig-out/{linux,macos,windows}/:
  • libzit-linux-x86_64.so
  • libzit-macos-x86_64.dylib
  • zit-windows-x86_64.dll
• GitHub Actions workflow under .github/workflows/ci.yml runs
  lint + unit + cross-build + integration on workflow_dispatch only;
  push/pull-request triggers can be enabled by appending to the on:
  block.

Credits

Algorithm: thejam79 (IT 0.051, 2002), minamina (IT_YV12 0.1.03, 2003),
poodle (64-bit / 8k mod), msg7086 (VapourSynth port, 2014). Zig port:
this repo.

v1.3.0 — initial Zig port

Changelog

All notable changes to this project are documented here. Format loosely
follows Keep a Changelog, versioning is
SemVer.

[1.3.0] — 2026-05-23

Initial release. A from-scratch Zig port of the VapourSynth-IT
inverse-telecine filter, bit-exact to the upstream C reference path
and with the original Avisynth-only parameters restored.

Version numbering picks up the IT lineage:

• Avisynth IT_YV12 v0.1.03 (minamina, 2003)
• VapourSynth VS_IT.dll v0103.1.2 (msg7086, 2014)
• This Zig port: v1.3.0.

Features

• Filter core.zit.IT(clip, fps=24, threshold=20, pthreshold=75, ref="TOP", blend=0, diMode=3) registered for VapourSynth API 4
  (R55+). Accepts YUV420P8 input with width % 16 == 0, height % 2 == 0,
  width ≤ 8192.
• Decimation modes:
  • fps=24 — 3:2-pulldown removal, output 24000/1001 fps (5→4
    frames per cycle).
  • fps=30 — field-matching only, input rate preserved.
• ref parameter ("TOP"/"BOTTOM"/"ALL"/"NONE"): the
  Avisynth original's field-match-direction switch, fully reimplemented.
  The VapourSynth upstream had stripped this down to a hardcoded
  "TOP".
• blend parameter: pure-Zig port of Avisynth's BlendFrame_YV12,
  triangular kernel weighted blend of post-matched frames, motion-gated
  on the same minD/avgD heuristic as the original. Only active when
  fps=24.
• diMode parameter with all four Avisynth deinterlace modes:
  • 0 DI_MODE_NONE — straight field copy, no deinterlace.
  • 1 DI_MODE_DEINTERLACE — full per-pixel five-candidate scorer
    with motion-gated vertical-average fallback (~500 LoC of Avisynth
    MMX re-expressed as pure scalar Zig).
  • 2 DI_MODE_SIMPLE_BLUR — vertical (T+2C+B)/4 blur on
    motion-flagged pixels.
  • 3 DI_MODE_ONE_FIELD — default; field-interpolation using the
    simple-blur and motion2max maps.

Frame properties

• Standard, always set on output: _FieldBased=0, _Combed per
  frame, _DurationNum/_DurationDen derived from output rate. All
  source props (_Matrix, _Transfer, _Primaries,
  _ChromaLocation, _SARNum/Den, _Range/_ColorRange, …)
  inherited via propSrc in newVideoFrame.
• Diagnostic (per-frame inspection for power-user scripts):
  ITMatch, ITMflag, ITIpFlag (utf8 1-char each); ITIvC/P/N/M,
  ITDiffP0/P1/S0/S1, ITBlended (ints).

Verification

• Bit-exact to a mechanically API-3 → API-4-ported build of the
  upstream C reference (reference/vapoursynth-cpp-api4/). 198 frames
  across 10 fixture × parameter combinations, plus 720 frames of two
  real telecined NTSC VOB samples in both fps=24 and fps=30 mode,
  byte-for-byte identical.
• 35 Zig unit tests for the algorithm primitives + SIMD helpers.
• 58 Python/VapourSynth integration tests — invariants,
  property checks, error paths, the regression-pinned md5 hashes, the
  upstream-compare matrix, and the new parameter sweeps.

Performance

500 frames of 720×480 NTSC, ReleaseFast build, single-threaded
fmParallelRequests:

Pipeline	fps
core.zit.IT(clip, fps=30)	~2700
core.zit.IT(clip, fps=24)	~2200
core.zit.IT(clip, fps=24, diMode=1)	~2350
core.zit.IT(clip, fps=24, blend=1)	~1700
core.vivtc.VFM(clip) (reference)	~600
core.vivtc.VFM + VDecimate (ref)	~1270

~4× faster than vivtc.VFM at field-matching, ~2× faster than
vivtc.VFM + VDecimate at IVTC, using the pure-C path + portable
Zig @Vector(N, u8) SIMD (pavgb, psubusb, pmaxub/pminub
equivalents).

Differences from the VapourSynth upstream

The upstream
VapourSynth-IT
plugin (msg7086, 2014) was a partial port of the Avisynth original.
This port reintroduces what was stripped and fixes four framework-level
issues:

1. ref, blend, diMode=0/1/2 reinstated (upstream hardcoded
   ref="TOP", blend=false, diMode=3).
2. Frame properties: upstream passes propSrc=null to
   newVideoFrame, so output frames carry no metadata at all. We
   inherit source props and set _FieldBased/_Combed/_Duration*
   explicitly.
3. Threading mode: fmParallelRequests. Upstream registered as
   fmParallel despite sharing mutable per-instance state — a latent
   race condition the modern VS scheduler hits more often.
4. Frame-request range widened to [base-2, base+6] (fps=24) /
   [n-2, n+2] (fps=30) so all reads through getFrameFilter are
   satisfiable. Upstream relied on API 3's sync getFrame to pull
   neighbours from cache; under API 4 that is not allowed inside a
   filter callback.
5. Edge clamping: getFrameFilter(n, …) is called with n clipped
   to [0, numFrames-1]. Under API 3 the core silently clamped; under
   API 4 it returns null and a deref crashes.

Build & packaging

• zig build --release=fast produces zig-out/lib/libzit.so on the
  host platform.
• zig build cross produces release binaries for all three targets
  under zig-out/{linux,macos,windows}/:
  • libzit-linux-x86_64.so
  • libzit-macos-x86_64.dylib
  • zit-windows-x86_64.dll
• GitHub Actions workflow under .github/workflows/ci.yml runs
  lint + unit + cross-build + integration on workflow_dispatch only;
  push/pull-request triggers can be enabled by appending to the on:
  block.

Credits

Algorithm: thejam79 (IT 0.051, 2002), minamina (IT_YV12 0.1.03, 2003),
poodle (64-bit / 8k mod), msg7086 (VapourSynth port, 2014). Zig port:
this repo.