shotdrift

You asked for a slow push in. Measure whether you got one.

Generative video is given a camera move and is under no obligation to deliver it. The usual check is a person watching sixty clips and forming an impression. shotdrift measures the camera path out of the pixels — pan, zoom, roll, frame by frame — and reports whether one physical camera could have produced it.

$ pip install shotdrift
$ shotdrift pan_demo.mp4 --expect push-in

pan_demo.mp4
  1280x720 -> measured at 512x288, 97 frames @ 24 fps

  camera path
    dominant move      pan
    pan                0.7159 of frame width (x -0.7159, y -0.0000)
    zoom               1.000x
    roll               -0.00 deg
    reversals          0  (measured, not judged)
    jerk               0.35
    incoherence        0.0000
    morph              0.041  (used to find cuts, not judged)
    closure            0.0000
    confidence         0.98

  asked for: push-in
    NOT HELD - asked for push in / dolly in; zoom moved +0.0001, under the 0.02
               floor - that move did not happen

  no findings: the motion is consistent with one physical camera.

That clip is a perfectly good shot — smooth, coherent, one physical camera. It is also not the shot that was ordered: it pans, and the push-in never happened. Both facts matter and they are reported separately, because "is this a real camera move?" and "is it the move I asked for?" are different questions.

Exit codes make it usable in a loop: 0 clean, 1 findings at or above --min-severity, 2 could not run, 3 the clip could not be measured.

Why pixels and not a solve

Nothing else is available. You are handed an mp4 by a model that will not tell you what it did. No camera metadata, no depth, no solve — so the measurement has to come from the frames themselves, which means it also works on footage from a camera, a render, or a competitor's demo reel.

Everything is reported in fractions of the frame width, never pixels, so a bound means the same thing on a 720p proxy and a 4K master.

What it measures

A grid of tiles is tracked by phase correlation, then one similarity transform is fitted to all of them:

dx = tx + s*x - r*y
dy = ty + s*y + r*x

That yields pan, scale and roll — and the part the fit cannot explain is the other half of the product. If tiles refuse to agree with any single camera, something in the frame is moving independently of one.

	what it means
incoherence	tile disagreement the camera model cannot explain
jerk	third derivative of the path: a carried camera changes speed smoothly
wander	travelled a long way, arrived nowhere
breathing	scale oscillates without going anywhere
closure	adding up every step vs. measuring first-to-last directly

closure is worth a sentence. The per-frame chain says where frame k ended up by summing every step; measuring frame 0 against frame k directly asks the same question using none of those steps. Where both are trustworthy they must agree. It is an independent route to the same number rather than a second opinion from the same code.

Severities are clean, soft, broken. soft exists so that a real camera with a person walking through it can be reported without being blocked — the first thing anyone does with a noisy alarm is stop reading it. broken is the default exit gate; --min-severity soft makes it strict.

Declared moves

$ shotdrift *.mp4 --expect push-in --min-severity broken
$ shotdrift --list-moves

static locked push-in dolly-in zoom-in pull-out dolly-out zoom-out pan-left pan-right tilt-up tilt-down roll-cw roll-ccw

Each is checked three ways, because the three failures need different fixes: did it happen at all, was it the right way round, and was it held or did a quarter of the travel run backwards.

The signs are camera-relative. A camera panning right makes the picture move left, so pan-right expects a negative picture-x. Writing those the intuitive way round made a correctly measured clean pan report "x went the other way" — the measurement was right and the vocabulary was wrong.

Edited sequences

A clip containing cuts is segmented and each shot measured on its own. Run as one take, an edited reel reported 13 reversals and a broken closure: all true, and useless, because there were six shots and no single camera move was ever there to hold.

Cuts come free from the morph residual already being computed. There are two signatures, and the second nearly got missed: where structure survives the join — a whip transition, a dissolve — morph spikes to 1.7–8.6 against the 0.36 real within-shot footage reaches. Across a clean cut between unrelated shots, nothing correlates at all, so no camera can be fitted and morph is NaN. The first version only tested for the spike, so it caught smeared transitions and walked straight past an ordinary hard cut.

Use --no-segment to force one measurement over everything.

What it cannot do

This section is the useful one.

It does not detect invented geometry. A tool like this ought to catch the melting-background tell, and I could not make it work. Measured over a fixed warp budget, real footage reached a structural residual of 1.06 while a literal cross-dissolve between two entirely different worlds read 0.358. Real scenes contain people, LED walls and vision mixes, so they change structure more than melting geometry does. No threshold separates them in either direction, so that check is not shipped. A dissolve under a locked-off camera comes back clean here, and that is pinned as a test so it cannot quietly change.

Direction changes are not a fault signal. Counting reversals seemed obvious. Real operated footage showed 27 direction changes on a slow zoom — an operator riding a rocker — against 3 in a deliberately broken control. The real material scores worse than the fault, so the count carries no signal alone. The question it was trying to answer is answered properly by --expect, which measures backtracked travel against the move actually requested.

Subject motion is a confound, by construction. A person crossing a locked-off frame raises incoherence. That is reported as soft, with advice saying so, rather than pretended away. Trimmed fitting rejects the worst-fitting quarter of tiles, which handles a subject occupying part of the frame and will not save you from one that fills it.

Cut detection can miss a cut between two visually similar shots, and closure degrades to NaN once the first and last frames no longer overlap. Both report "not measurable" rather than a number — an anchor that cannot be trusted is skipped, because cannot measure must never be reported as measured bad.

Calibration

Thresholds are not taste. Every bound was set by measuring real single-camera footage and synthetic controls with known faults; calibration.md records the distributions and validate_real.py re-derives them.

The requirement that shaped the tool is silence on real footage: across 24 shots from two events, nothing broken fires, and the three soft findings are all genuine operator behaviour — two stage cameras that reframed and came back, one of which also zoomed in and back out.

measured over 24 real shots	real max	gate	soft at
incoherence	0.00068	—	0.004
closure	0.0092	—	0.02
jerk	2.35	speed ≥ 0.005	1.9
breathing	110	scale_amp ≥ 0.04	4.0
wander	52	travel ≥ 0.05	4.0

Three of those five are gated rather than thresholded, and that is the most useful thing in this README. They are ratios, and on a locked-off camera each denominator is legitimately near zero, so noise divided by nothing produces an enormous number on the most ordinary footage there is. Real maxima of 2.3, 110 and 52 against bounds of 1.9 and 4.0 are not near-misses — the ratios have come apart, and no bound can be raised to cover that without covering every real fault too. What works is refusing to judge the quality of a move until there demonstrably was one. Every gate above has a measured gap behind it: real footage reaches a speed of 0.003 where the slowest control with a genuine move sits at 0.014.

The harness passed while the tool was wrong

Worth the paragraph because it is the failure mode of every calibrated tool.

shotdrift <clip> reported BROKEN on four of seven real clips while validate_real.py printed VALIDATION PASSED — same code, same files, same afternoon. The harness measured the first 8 seconds; the tool measures 600 frames. And breathing was gated on accumulated scale change against a fixed bound, which on a locked-off camera walks 0.0025 at 100 frames to 0.0263 at 1200 while the per-frame rate stays flat at 2e-5. The gate was crossed somewhere past 500 frames by nothing but clip length, and 8 seconds at 50p is 400 frames — just underneath it.

So: a fixed bound on a cumulative quantity is a time bomb, and a control that does not run the shipped configuration validates a configuration nobody uses. The gate is now peak-to-peak amplitude, which does not grow with length and is what a viewer can actually see, and the harness defaults to the whole clip.

breathing also lost its broken bound in the same round, for the third instance of the pattern this tool keeps running into: real operated footage reached a breathing ratio of 169 against the deliberate fault's 46. An operator zooming in and back out is scale oscillating without going anywhere.

Python

from shotdrift import measure

r = measure("take_07.mp4", expect="push-in")
print(r.verdict, r.ok)
for s in r.shots:                      # cuts are detected; one entry per shot
    print(s.start, s.end, s.verdict, [f.code for f in s.findings])

Frames you never wrote to disk

from shotdrift import measure_frames, report

r = measure_frames(batch, expect="push-in")   # (n, h, w, c), numpy or torch
print(report(r))

No ffmpeg, no temp file. Takes uint8 0..255 or float 0..1, greyscale or RGB(A), and a channels-first batch is refused by name rather than measured sideways. report() renders exactly what the command line prints — the CLI calls it, so the two surfaces cannot drift apart.

In ComfyUI

shotdrift-comfyui measures the batch inside the graph that produced it, and can stop the queue when a take does not hold the move it was given. That is the point of measuring here rather than afterwards: an unattended run of sixty takes is only worth doing if the bad ones announce themselves.

Requirements

Python ≥ 3.9, numpy, pillow, and ffmpeg on PATH. No OpenCV, no torch, no network, no GPU. A tool that needs a 2 GB wheel to measure a camera move does not get run.

Development

pip install -e ".[test]"
pytest -q                                  # 83 tests
PYTHONPATH=src python validate_real.py     # real footage + controls, as shipped

MIT. Built by Syamjith NK — cinematographer and AI creative technologist, Abu Dhabi.