An Archive Clip Has Repeated Fields. Recover Its Film Cadence—or Keep the Video Timing?
An Archive Clip Has Repeated Fields. Recover Its Film Cadence—or Keep the Video Timing?
A combed frame tells you that two fields inside one video frame disagree. It does not tell you why, and the why decides everything.
Three different things produce that disagreement. A film frame can be carried by two video frames, so one video frame holds the tail of one film frame and the head of the next; the original progressive frames are still in there, at least partly, and field matching can reassemble them. A video camera can record two genuinely different moments in the two fields of one frame; there is no progressive original underneath, and any filter that pairs those fields is inventing one. A conversion can average fields into a blend; the detail is already gone and there is nothing exact left to match.
So a repeated field pattern in your clip is a hypothesis about the past, not a property of the file. The file will only help you test it where something moves.
The short answer, before the reasoning: where the provenance notes and the observed motion both support a regular film-to-video transfer, match the fields and remove the repeats — for the sections that support it. Where the material is native interlaced, blended, or mixed, build a viewing copy that carries the recording's own temporal progression and label it exactly that way. Many archive reels need both treatments in different places, and an honest deliverable is a section list, not one setting applied to the whole file.
Test the cadence hypothesis across motion and cuts
Start with the paperwork, because the picture is weaker evidence than it looks. What element was transferred, on what machine, at what rate, for what purpose, and what happened to the tape afterward? A note that says an original was transferred to video at 30 fps gives the cadence hypothesis real support. A note that says a tape is a dub of a dub with an unknown intermediate gives you almost nothing, and no amount of stepping through fields will recover what the note omitted.
Read the rate precisely. A 24-to-30 transfer and a 23.976-to-29.97 transfer share the same ratio — four film frames become five video frames in both — but they do not produce identical durations, and the difference will matter when you check totals later.
Then go to motion. Pick a passage where something identifiable travels steadily across the frame or turns through a known angle: a hand, a car, a title crawl, a head turning. Static shots will lie to you, because pairing errors are invisible when nothing moves. Step through the material field by field, not frame by frame.
What you are looking for is whether the same instant appears more than once. In a transfer built from whole repeated frames, you will see one video frame that duplicates its neighbour, once every group of five. In a transfer where the repeat was placed at field level, you will not see a duplicated whole frame at all; you will see a field that repeats while its partner already belongs to the next instant. Both are recoverable in principle. They are recovered differently, and a recipe tuned for one will misbehave on the other.
Native interlaced motion looks different again: every field is a new moment, and the two fields inside a video frame are separated by an interval — roughly a sixtieth of a second in a 60-field system. If you see that steady field-by-field advance with no repetition, there is no film frame underneath and the question changes.
One combed frame cannot decide this. It tells you the two fields disagree; it cannot tell you whether the disagreement is a straddled film frame, a blend, or a real interval of time. And one apparently regular interval cannot decide it either — a short stretch can look perfectly periodic while the rest of the reel is something else.
So check the cuts. Transfers break at splices, at reel changes, and wherever video material was inserted into a film record. Step across several edits and write down whether the repeat pattern persists through them, restarts, shifts by a frame, or stops entirely. A pattern that holds for two minutes and then shifts once is a different problem from a pattern that holds throughout, and a pattern that disappears at a title card tells you the reel is a compilation.
By the end of this pass you should be able to label sections, not the file. Regular transfer. Native interlaced. Blended. Mixed, with boundaries. Blended material deserves its own category because there is no correct pairing to find: the conversion averaged moments together, and a matcher has nothing exact to lock onto.
One boundary case is worth naming so you don't chase it. Film shot at 24 and transferred at 25 for a 50-field system has a one-to-one frame relationship and no repeated fields at all — everything simply runs about four percent fast. That is a speed question, not a cadence question, and it is a different decision with different checks.
Separate field matching from removing repeated frames
Inverse telecine is usually described as one operation. It is two, and keeping them apart is most of what makes the work checkable.
The first task is matching: decide which fields describe a single instant and assemble them into a candidate progressive frame. The second is decimation: drop the outputs that repeat a neighbour. The FFmpeg filter documentation draws exactly this distinction — its fieldmatch section treats the matching of fields and the later removal of duplicate frames as separate stages, and flags limitations for mixed interlaced material (the filter documentation draws this distinction; what the filter does to your particular file is a separate question).
Matching has to answer two things at once, and it answers both by assumption. It needs the field order — which field is top and which is bottom — and it needs the pairing pattern, which fields belong together in one frame. Set the field order the wrong way round and the frames still assemble. They assemble the wrong pairs. Nothing about the output announces the mistake, and if the shot is locked off you will see almost nothing at all; the error shows up in motion as combing or judder that looks like a source problem rather than a decision you made.
Decimation has its own way of going quietly wrong. Remove one output in every five and you will hit the correct frame count and the correct duration. Remove the wrong one in every five and you will also hit the correct frame count and the correct duration, while shifting the sequence or duplicating the wrong moment at a pattern break.
A chain that runs to completion has produced a file, not a verdict.
Here is a report a project can inherit and then misread. Call the fixture B022. Its sequence was built so the numbers would come out clean: 96 frames at 24 fps became 120 frames at 30 fps and returned as 96 frames at 24 fps. All three container durations were reported as four seconds.
The arithmetic is worth doing anyway. 96 ÷ 24 = 4. 120 ÷ 30 = 4. 96 ÷ 24 = 4. The 120-frame intermediate carries 24 more frames than it has instants, which is one repeat in every five — the same 4:5 relationship you would expect from a 24-to-30 transfer, or equivalently from a 23.976-to-29.97 transfer. Nothing in the report is wrong.
What it recorded is the problem. Only counts and durations were checked: not the recovered frames against the original, not the motion across the sequence, not the sound. The run artifact is not in hand here, so this report cannot be presented as independently recovered evidence. It describes an ideal synthetic cadence, and an ideal synthetic cadence is not an archive.
Two details show how narrow the ideal case is. The fixture uses integer 24 and 30, so four seconds is exact; the same 96 and 120 frames at 23.976 and 29.97 would each run about 4.004 seconds — still agreeing with each other, which is the point. Duration agreement is that cheap. And the report never says whether its repeat was a whole duplicated frame or a field interleave, which is precisely the fact that decides whether matching is routine or delicate.
That is the trap in miniature: a clean count can coexist with wrong pairings, and a matching four-second duration can coexist with damaged motion at a cut. Both outcomes look like success in a file listing.
There is also a limit worth stating plainly. Field matching does not infer where the material came from, does not recover detail that a blend averaged away, and does not produce a moment the camera never recorded. If the source is genuinely mixed, the documentation's own caveat applies and no amount of tuning removes it.
Compare recovery with a timing-preserving viewing copy
When the cadence evidence holds for a section, you can build a recovery derivative: progressive, at the film rate, repeats removed, cuts placed where the pattern actually broke. That derivative makes a claim — these frames correspond to the transfer's film frames in these sections — and the claim should be written down with the section boundaries that qualify it.
When the material is native interlaced, blended, or unresolved, the alternative is not a failed version of the same thing. It is a different product with a different claim: a viewing copy that preserves the recording's own temporal progression without asserting a film rate.
For native interlaced motion this is the more faithful object, not the consolation prize. There is no film frame to recover, and the two fields are two moments. A treatment that emits one output per input field carries both moments forward in order, at twice the frame rate. A treatment that emits one output per input frame merges the two moments into one picture, which is usually what a viewing context can play but is also a real loss of temporal precision. Neither is dishonest. Choosing between them is choosing which moments you keep, and the choice belongs in the record rather than in a preference.
Segment the work when the reel is mixed. Use recovery where the pattern is supported, the timing-preserving treatment where it isn't, and put the boundaries in the notes. The failure mode to avoid is one global chain that produces a smooth-looking result and quietly mangles the sections it doesn't fit. A montage hides that failure very well — motion damage reads as grain, as a soft frame, as an edit that "always looked like that." If a single process damages mixed passages, stop and segment rather than conceal it.
Keep the source transfer untouched. Every one of these products is a derivative, and the only copy you cannot rebuild is the one you overwrote.
The practical consequence of deferring reconstruction is mild. A pitch can use a qualified viewing copy — a progressive file at the recording's own rate, with a note saying what it is and which sections remain unresolved — without anyone calling it a recovered original. The reconstruction can proceed later, with better evidence, and nothing about the pitch has to be retracted.
Require several checks before naming the result
Check the frame sequence and the motion first. Does the output advance monotonically in real time, or does it stutter, hold, or step backward somewhere? Where you have a known original, compare frame to frame; where you don't, watch the motion through a movement you understand and look for a hitch that shouldn't be there.
Check the edits. Do the cuts land on the same source frames in the derivative as they do in the source? A one-frame shift that appears after a pattern restart is the signature of a mis-set pattern, and it is invisible in a duration readout.
Check the total duration, knowing what it can and cannot tell you. The fixture's three four-second figures rule out a gross rate error. They cannot see a swap, a shift, or a dropped instant.
Check the sound. Sync at the head proves very little; sync at the middle and the tail proves more. If a speed question is entangled with the cadence question, pitch will drift too, and that is a different investigation with different evidence.
Keep three kinds of agreement separate in your own mind, because they are not interchangeable. Frame-count agreement is arithmetic. Pixel identity requires a known original to compare against. Perceived fidelity requires eyes and a playback context, and it will sometimes disagree with both of the others in both directions. Confusing any two of these is how a derivative gets described as a restoration.
Record the work while you do it: the tool and its version, the assumed field order, the pattern you inferred, the section boundaries, and every passage you could not resolve, with timecodes. An unresolved passage with a timecode is a finding. An unresolved passage folded into a montage is a liability.
The fixture report would need three more things before it could support any conclusion. Labeled moving objects, so a recovered frame can be compared against a known original frame and a wrong pairing becomes visible. An edit placed inside the sequence, so a pattern break can be observed rather than assumed. Synchronized sound, so drift shows up on its own. Then the same recipe run against native-video and mixed-cadence material, to see whether it damages what it doesn't fit. That is a plan, not a result, and none of it has been run here; the fixture's artifact is missing and the checks above remain to be executed on authorized material. A specialist review may also change where the example has to stop and use a local treatment instead.
What a finished pass produces is a short document with a route in it. It names the sections where cadence was recovered, the evidence supporting each one — observed pattern, cut behavior, duration, sound — the assumption made about field order and pattern, the tool and version, and the sections left unresolved with their timecodes. It keeps the source transfer. The timing-preserving copy sits in that document as a route with its own justification, not as an apology for work that stalled.
And the discipline that holds all of it together is a single refusal: an ideal synthetic cadence, or one clean frame, or a smooth playback, is never proof about an unknown historical recording. A recovered sequence that looks right is still an interpretation, and the difference between an interpretation and a recovery is whether you can say what you checked.
Frequently asked questions
What does a combed frame actually tell you?
It tells you that two fields inside one video frame disagree. It does not tell you why. The disagreement can come from a film frame carried by two video frames, from a native interlaced camera recording two genuinely different moments, or from a conversion that averaged fields into a blend. The why decides whether matching can recover anything exact.
When should you recover film cadence, and when should you keep the video timing?
Where provenance notes and observed motion support a regular film-to-video transfer, match fields and remove repeats for the sections that support it. Where the material is native interlaced, blended, or mixed, build a viewing copy that preserves the recording's own temporal progression and label it that way. Many archive reels need both treatments in different places, so an honest deliverable can be a section list rather than one setting for the whole file.
Why can a clean frame count and matching duration still hide a failed inverse telecine?
The fixture with 96 frames at 24 fps, 120 at 30 fps, and 96 at 24 fps shows that counts and durations can agree while wrong pairings, pattern shifts, or dropped instants remain invisible. Duration agreement rules out only a gross rate error. It cannot see a swap, a shift, a mis-set pattern, or damaged motion at a cut. Frame sequence, motion, edits, and sound need separate checks.
What is the difference between field matching and decimation?
Matching decides which fields describe a single instant and assembles candidate progressive frames. Decimation drops outputs that repeat a neighbour. Matching needs the correct field order and pairing pattern; setting field order wrong still assembles frames but pairs the wrong fields, which may show only as combing or judder in motion. Decimation can hit the right frame count and duration while removing the wrong one in every five, shifting the sequence or duplicating the wrong moment at a pattern break.
What should be recorded before calling a result a recovery rather than an interpretation?
Record the tool and its version, the assumed field order, the inferred pattern, the section boundaries, and every unresolved passage with timecodes. Also record the evidence supporting each recovered section, such as observed pattern, cut behavior, duration, and sound. Keep the source transfer untouched. A recovered sequence that looks right is still an interpretation; the difference between an interpretation and a recovery is whether you can say what you checked.