What is interlacing?
Interlacing sends half a picture at a time: one field carries the odd-numbered lines, the next carries the even ones, captured a fraction of a second apart. That is what the "i" in 1080i and 480i means, and 1080p does not do it. A modern display shows both fields at once, so anything that moved between them appears twice, one line apart — the combing on moving edges. Deinterlacing rebuilds whole frames from those field pairs, and it has to happen before anything sharpens the picture.
If moving objects in a clip have fine horizontal stripes along their edges while everything standing still looks perfectly normal, you're looking at interlacing. It isn't damage, and it isn't a bad transfer. It's the footage working exactly as designed, on a kind of display that no longer exists.
A bandwidth trick from the 1930s
Broadcast television couldn't carry enough bandwidth to send complete frames at a smooth rate. The workaround was to send half a picture at a time: all the odd-numbered lines, then all the even-numbered lines, each one a fraction of a second apart. Each half is called a field. A cathode ray tube drew them in sequence fast enough that your eye combined them, and you got the motion smoothness of fifty or sixty pictures per second for the bandwidth of twenty-five or thirty.
It was a genuinely clever piece of engineering, and it survived into DV camcorders, DVDs and early digital broadcast. Which is why so much family footage from roughly 1980 to 2010 is interlaced.
Why it looks wrong now
Modern displays don't work that way. They show whole frames. So the two fields, captured at genuinely different moments, get displayed at the same instant. Anything that moved in the gap between them appears twice, offset, one stripe apart. That's the comb.
Stationary parts of the frame look fine because nothing moved between the two fields, which is why the effect appears to cling to moving objects specifically.
Why it has to be fixed first
This is the part people get wrong, and it's expensive. An AI upscaler has no concept of fields. Hand it a combed frame and it sees a regular, high-contrast horizontal pattern, exactly the sort of fine structure it was trained to preserve and sharpen. So it preserves and sharpens the comb. You end up with crisp, high-resolution interlacing artefacts, which are far harder to remove than the originals and often impossible to fully undo.
The order that works is deinterlace, then denoise, then upscale, then colour. Every step makes the following ones harder if you take it out of turn, and interlacing is the least forgiving of the lot.
Telling interlacing apart from things that look similar
Comb teeth only appear on movement, and they're perfectly horizontal and evenly spaced. Rolling shutter skew leans vertical edges over but doesn't stripe them. Compression blocking shows up as square patches, usually in dark or busy areas, and it doesn't care whether anything moved. Ghosting from a bad tape transfer smears rather than combs.
If you're unsure, step through a few frames one at a time on a shot with fast motion. Interlacing is unmistakable frame by frame, and almost invisible at normal speed on some displays, which is how it survives into finished edits without anyone noticing.
The mechanical check, and the two ways a file lies about it
The eyeball test needs a shot with real motion in it. The mechanical version is ffmpeg's idet filter, which classifies the frames themselves: ffmpeg -i in.mp4 -vf idet -f null -. The last "Multi frame detection" line is the verdict — TFF and BFF counts against Progressive. That is the same line Crisp reads: probe.interlacing() parses it and answers "unknown" rather than guessing when it cannot classify at least eight frames.
The container flag is not the answer. Measured on the 1936 newsreel restored for this site: the original MPEG-2 honestly reported field_order=tt, but once the clip had been cut and re-encoded — the state most footage reaches you in — the container claimed progressive while every frame was still combed. Looking at the pictures instead, idet read 96 of 96 frames interlaced on the file that lied, and 82 of 82 progressive on the same clip after inverse telecine. Re-encoding falsifies the flag. It does not hide the combing from a detector that reads frames.
A resize does hide it. Vertical scaling interpolates the two woven fields into each other, so the teeth soften into a ghost. On that same clip the 640×480 frames read interlaced and the identical frames scaled to 1280×720 read progressive — the damage still plainly there, still visible as smearing, the detector simply blind to it. That is also why combed footage tends to look like motion blur rather than comb teeth by the time it has become a web-sized video. Run the check on the original, before anything has been scaled.
Doing it on a Mac
Crisp straightens this out for you. When you enhance or upscale, it checks the source for combing and fixes it before anything sharpens it — and it picks the right fix rather than one fix for everything, because footage that started on film needs the pulldown reversed instead. The deinterlacing walkthrough covers what it is choosing between, and how to do it by hand if you would rather. The exception is the Max quality lane, which restores the frames it is handed: there Crisp warns you before spending the render, because an hour of generative restoration faithfully reconstructing comb teeth is the most expensive version of this mistake.
It does not guess between the two fixes, because guessing wrong is destructive in one direction: reversing a pulldown on genuinely interlaced video drops one real frame in five that nobody ever duplicated. So the frame rate narrows the choice first — a 3:2 pulldown can only exist near 30fps, and 25fps PAL material is interlaced but never telecined — and then the inverse-telecine answer is rendered on a short sample and handed back to the detector before the real render commits to it. If that sample does not come back clean, the job falls back to the deinterlace, which is always safe. Whichever route ran is named on the job's route line, so you are told which of your frames changed rather than finding out later.
Two other lanes are exceptions, and Crisp says so instead of skipping the step silently. HDR sources do not get the automatic fix at all; there the app points you at HandBrake's Decomb, or ffmpeg's fieldmatch,decimate for film, and you straighten it before you bring it back. A timeline export does straighten each clip as it assembles the edit, but with the deinterlace rather than the pulldown reversal — a concatenated edit forces one common frame rate, and dropping every fifth frame of one clip would fight it. So film that went through a telecine is blended on that lane rather than restored, and the single-clip enhance is the one that gives the photographed frames back.
Once you bring it a progressive file, everything after that runs on your Mac. Nothing is uploaded, which tends to matter for exactly the sort of irreplaceable family footage that is most likely to be interlaced in the first place.
Interlacing, in detail
Can I just upscale interlaced footage and fix it afterwards?
No, and this is the single most expensive mistake with old video. The upscaler treats comb teeth as real detail and sharpens them. Deinterlacing afterwards then has to work against artefacts that have been reinforced and enlarged. Deinterlace first, always.
Does deinterlacing lose quality?
It costs something, unavoidably, because you're reconstructing whole frames from half-pictures. A good deinterlacer blends or interpolates intelligently rather than simply discarding one field. Throwing away every other line is the fast method and it halves your vertical resolution.
How do I know whether my footage is interlaced?
Find a shot with fast horizontal movement and step through it a frame at a time. If moving edges break into fine horizontal stripes while static areas stay clean, it's interlaced. Footage from DV camcorders, DVDs and pre-2010 broadcast very often is.
Is 1080i actually 1080p?
Not in terms of motion. 1080i carries the same line count but delivers it as two half-height fields per frame, so a moving subject never has a full 1920x1080 sample at a single instant. Deinterlaced 1080i and true 1080p are not equivalent.
Related
Download Crisp for Mac Free to try, one-time $129 to remove the watermark. Runs entirely on your Mac.