VCG Deinterlacer

A Windows app for deinterlacing VHS, Hi8, MiniDV and other videotape captures with QTGMC, the highest-quality software deinterlacer there is.

It measures your capture, tells you what it found and what it recommends doing about it, and writes an archival master. Every recommendation can be overridden, and every number behind one can be read.

What it is for

A tape capture is not a finished file. The frames are interlaced, so they have to be rebuilt into progressive ones. The capture card digitised more than the picture, so there are black bars and a torn band at the bottom to remove. The pixels are not square. The levels are often outside legal range, the colour drifts, and the tape brought noise with it.

This tool does the structural work properly — once — so that what you keep is a master you can edit, grade or archive from, rather than something you will want to redo.

What you can feed it

The drop zone names the formats a tape capture actually arrives as:

AVI · MPG · MKV · MTS · DV · VOB

That list is a hint, not a gate. The file picker also offers MP4, MOV, M2TS, MOD and WMV, and a dragged file is not filtered at all, so anything FFmpeg and VapourSynth can open will open — including a genuinely interlaced MP4 off a DVR.

MP4 and MOV are left off the hint deliberately. They are delivery formats, and they are this app's own output, so advertising them as inputs invites a finished export being fed back in and deinterlaced a second time.

The three modes

You choose the mode when you pick the file, because it decides which measurements are even worth taking.

Mode	What it does
Just deinterlace	Deinterlace, fix field order, crop the blanking, square the pixels, tag the colour. Nothing else is touched.
Deinterlace + auto-fix	The above, plus whatever the analysis recommends — denoising, dehalo, levels, colour.
Full control	The above, with eleven settings pages so you decide every filter yourself.

Just deinterlace exists for people producing a master to grade elsewhere — in DaVinci Resolve, or Topaz — who want the structural work done and nothing else altered. It is enforced by test, not by convention: a run in that mode cannot enable a single optional filter, including when you explore in Full control first and then switch back down.

It is also much faster to analyse. Noise, halos, colour cast and chroma bleeding are most of the analysis time, and nothing in Just deinterlace acts on them, so those probes are skipped — about 2 seconds instead of 13 on a short file, and proportionally more on a long tape.

What it measures

Everything below is measured from your actual file, with the figures shown beside each verdict:

• Scan type — interlaced video, or film that was telecined and should be inverse-telecined instead. Telecine is one possible answer, not the question.
• Field order — the single most consequential fact about an interlaced file. Fed the wrong one, QTGMC interpolates each field against the neighbour on the wrong side of it in time, and the result looks worse than the source. Most tape captures declare nothing, so it is measured from the picture.
• Frame edges — see automatic cropping, below.
• Levels — whether luma falls outside the legal 16-235 range.
• Colour — a consistent cast in U or V, and overall saturation.
• Noise — how much, and therefore how much denoising (if any) to suggest.
• Edge halos — oversharpening baked in by the deck or the capture chain.
• Chroma bleeding — colour smeared horizontally past its edges.
• Audio — how many real channels there are, so a mono recording captured as dual-mono, or one with a dead channel, is not silently carried through.

A measurement the app does not trust — too few usable frames, say — is never acted on automatically, and says so.

Automatic cropping

This is the part people are most surprised by.

Other tools offer fixed presets: BT.601 crops 8 columns from each side whatever the file looks like. A real capture card does not oblige. The horizontal blanking it digitises lands wherever the timing of that particular card, tape and deck put it, so the black margin is routinely wider on one side than the other. On one capture measured here it was 14 columns on both sides, which means the 8-column preset leaves blanking inside the picture.

So the crop is measured off your own frames. Three different things sit at the edge of a tape capture, and each gets its own test:

1. Blanking — flat and near-black. Brightness alone cannot find it, because a night scene is also near-black. What separates blanking is that it carries no detail, so a line counts only when it is both dark relative to the picture and flat relative to the picture.
2. The analog edge ramp — past the blanking, the picture does not begin abruptly; it ramps up over a few columns of dim, smeared video. A brightness threshold run across this eats into genuinely dark content. What separates a ramp is that it climbs steadily inward.
3. The head-switching band — the torn band in the last few lines of a VHS frame. It is not dark, so neither test above sees it. It is found instead by how unlike the line above each line is, in luma and especially in chroma: the band is torn out of sequence with the picture and its chroma phase is wrecked.

The result is a recommendation, shown with its numbers, on a page where you can nudge each edge and watch the preview. Presets and a manual crop are still there if you would rather set it yourself.

DV defaults to no crop, because DV uses all 720 samples as active picture.

Deinterlacing

QTGMC runs at a deliberately least-destructive profile. Its only job here is to rebuild the fields faithfully — not to clean, sharpen or stabilise, which are separate, optional, visible steps.

Setting	Value	Why
Preset	Slower	Quality baseline; higher presets give negligible visible gain
TR2	1	Light final temporal smoothing — the default of 2 acts as mild denoising
SourceMatch	3	Corrects the output back toward the original fields: the highest-fidelity mode
Lossless	2	Puts the real source fields back after resharpening
Sharpness	0.1	Effectively off; sharpening is not the deinterlacer's job
Sbb	0	No sharpness-limiting backblend
MatchPreset	slow	Accuracy for the source-match stage
NoiseProcess	0	QTGMC does no denoising at all

Output is double-rate: 50 or 59.94 progressive frames a second, one per field, so no motion is thrown away. A film transfer is the exception — it is not deinterlaced at all, and comes out at the film's own rate. See below.

Two QTGMC implementations ship, and you can pick either:

• havsfunc — havsfunc.QTGMC, the long-standing reference implementation.
• vs-jetpack — vsdeinterlace.QTempGaussMC, the modern rewrite.

Both are driven from the same profile above, so the choice is between the two libraries rather than between two sets of settings. Measured here they take about the same time, and the rewrite carries slightly less fine vertical detail. Neither benefits from a faster GPU, because the motion estimation underneath QTGMC is CPU-only.

Film transfers (3:2 pulldown)

Material shot on film and transferred to NTSC tape carries 3:2 pulldown: four film frames spread over five video frames, two of them holding fields from two different film frames. Deinterlacing that would be the wrong operation — those fields are not a moment apart, they are whole film frames taken to pieces.

The scan reads the picture itself at several points across the whole running time rather than the opening seconds, so a studio logo rendered at the video rate cannot make a film transfer read as video. When it finds the pattern, the original frames are put back together with vivtc.VFM and the native 23.976fps restored. A PAL 2:2 transfer is paired back the same way and stays at 25fps, with no decimation — there are no duplicate frames to drop.

A film transfer is rarely film all the way through. Titles, credits, effects shots and concert inserts are routinely finished at the video rate and cut into the same tape, and a scene change can land mid-cadence. Every frame the field match cannot rebuild is deinterlaced on its own, before the duplicate frames are dropped, so a part-film tape comes out whole rather than with combing left in the video-sourced scenes. Only the flagged frames pay the cost.

And you can overrule it. Film or video is a measurement with a confidence attached, not a certainty: a film transfer where every scene is busy can read as video, and a mostly-static interlaced tape can read as film. The Check video page shows which it decided and lets you change it — including when it found nothing, behind an "Override detection" link.

Optional corrections and enhancements

In Full control each of these is a page of its own, with a before/after preview:

Trim · Levels · Colour tint · Noise · Halos · Colour bleeding · Upscale · Audio · Watermark · Dithering · Film grain

A note on denoising, since it is the one most often asked about: QTGMC performs none. When the Noise step is enabled it runs after the deinterlacer, on the progressive frames. BM3D is the tool, and the script picks the fastest build present on your machine — CUDA on an NVIDIA card, otherwise an AVX2 build about four times quicker than the original, then the original, then KNLMeansCL, then SMDegrain. Luma and chroma are both denoised, in two passes: luma with a temporal radius, chroma spatially only, at half again the luma strength. Chroma noise is a first-class VHS defect — colour is recorded as a "colour-under" subcarrier down at 629 kHz, and it arrives soft and noisy. Those are independent facts, and the softness is a reason to filter chroma harder rather than leave it alone: there is little detail there to lose. The chroma pass skips temporal matching, which is what smears colour.

Output formats

Figures measured on two SD captures at 640x480 double-rate.

	Codec	Bit rate	Per hour	Use it for
Best for editing	ProRes 422 HQ, 10-bit 4:2:2, in .mov	103-110 Mbps	about 50 GB	An edit master. What editing software expects, and fast to scrub. Visually lossless, but a DCT codec.
Highest quality	FFV1 version 3, 8-bit 4:2:2, in .mkv	71-83 Mbps	about 35 GB	Keeping. Mathematically lossless, and the archival community's choice. Smaller than the ProRes, which surprises people.
Best for sharing	H.264 High, 8-bit 4:2:0, in .mp4, x264 at CRF 18	2.7-3.3 Mbps	about 1.5 GB	Sending to someone. Plays anywhere; gives up the most.

All three formats are dithered on the way out. The pipeline processes at 16-bit whatever the source was, so the deinterlacer and the filters have precision to work in. Writing the file reduces that to the codec's depth — 10 bits for ProRes, 8 for FFV1 and H.264 — and error diffusion spreads the rounding error so that a smooth gradient does not come out as visible bands. The 8-bit formats discard more and need it most, but it applies to ProRes as well. Ordered dithering, or none at all, are options on the Dithering page.

Sound is carried through as uncompressed 16-bit PCM in the two master formats, and re-encoded to AAC only for H.264.

Colour is tagged rather than converted — BT.601 primaries, smpte170m for NTSC or bt470bg for PAL. The one exception is upscaling to HD: an HD-sized frame is read as BT.709 by players, so in that case the pixels really are converted, and the tags follow them.

Each format's page carries the full detail behind these summaries, including the trade that actually matters between the two masters — FFV1 is lossless but 8-bit, where ProRes is 10-bit through a DCT.

What travels with the file

Say what the recording actually is

Before you export, you can tell the app what is on the tape, and it writes that into the finished video rather than into a note beside it:

What's on it	"Grandma Jones at the 1998 Christmas party"
When	1998-12-25 — the date of the event, not the date you digitised it
Where	"Toronto, Canada"

This is for anyone digitising a family archive. A box of unlabelled tapes becomes a folder of unlabelled files unless somebody writes down what they were, and the person who knows is the person doing the capture — not whoever opens the folder in thirty years. Putting it in the file means it survives copying, emailing, re-muxing and being moved between drives, and it shows up in Windows Explorer, in MediaInfo, in a player's info panel, and in a search across the drive. A sidecar text file does none of that past the first time someone reorganises.

All three fields are optional, and anything left blank writes nothing. The date accepts a bare year or a year and month, so somebody who knows only that it was the summer of 1998 is not forced to invent a day.

Written as title, creation_time plus date, and location — including the Apple-style location key, so that .mov and .mp4 files read it as well as .mkv.

And what was done to it

A restored tape outlives the memory of how it was restored, so the settings go with it three ways, kept separate from the description above — a search for "Christmas 1998" should not have to wade through deinterlacer settings:

• A one-line summary in the file's own metadata — survives copying, emailing and re-muxing, and is readable in MediaInfo or ffprobe without this app.
• A JSON sidecar — the complete configuration, which is what to reach for when you want to treat the next tape from the same deck identically.
• A log — what the tools actually said, for when something looks wrong.

The generated VapourSynth script is saved too, so the render is reproducible by hand.

All three are written from the same configuration the render used, so none of them can describe a setting that was not applied.

Installing

Download the release, unzip it anywhere, and run the exe.

On first run it fetches a 158 MiB support payload containing FFmpeg, a portable VapourSynth with its plugins, and the deinterlacing and denoising libraries. The download is checked against a byte count and a SHA-256 compiled into the app, extracted to a staging directory, and only swapped into place once it is complete and verified — so a failed update cannot leave you without a working install. Nothing is installed into your system Python, and no shell is spawned to do it.

Requires 64-bit Windows 10 or 11.

Building from source

cargo run --release -p vcg-gui

Or run.cmd, which builds the release binary if it is missing and launches it.

cargo test                  # the full suite
cargo clippy --all-targets  # clean, and kept clean in CI

The payload can be managed on its own:

cargo run --release -p vcg-deps --bin vcg-deps -- status --deps <DIR>
cargo run --release -p vcg-deps --bin vcg-deps -- install --deps <DIR>

status exits non-zero when the payload is not ready, so it can gate a script.

There is also a command-line prober that runs the same analysis and prints what it found:

cargo run --release -p vcg-tools --bin vcg-probe -- "path/to/tape.avi"
cargo run --release -p vcg-tools --bin vcg-probe -- --quick "path/to/tape.avi"

Built on

QTGMC / havsfunc · vs-jetpack · VapourSynth · FFmpeg · BM3DCUDA · egui

---

By VideoCaptureGuide.