Generated video has no stillness
A real fixed camera holds perfectly still for 28.8 minutes. No generated footage we measured stays still longer than 4 seconds.
Joe Cox, Resoa
The measurement
| source | quiescent fraction | longest continuous stillness |
|---|---|---|
| real fixed camera, 50 min | 0.992 | 1726 s (28.8 min) |
| real, second window | 0.988 | 1043 s |
| real, third window | 0.991 | 849 s |
| clinical prior, hospital inpatient room | 0.674 | minutes |
| generated, frontier model clip | 0.320 | 4 s |
| generated, ours | 0.175 | 1 s |
| generated, ours (longer run) | 0.104 | 3 s |
Three orders of magnitude.
The likely cause is mundane and structural. Video generators train on trimmed, action-centric clips. Nothing in that data teaches a model that a monitored room is mostly nothing happening. Stillness is not a thing the training distribution contains.
This matters beyond realism. If synthetic monitoring footage is continuously eventful, a detector trained on it never sees the base rate it will face in deployment.
We threw the first instrument away
Version one was void, and it took a specific check to prove it. Its threshold was the median of the quietest decile plus three times that decile's absolute deviation, which is scale-invariant: it adapts to whatever distribution you hand it, so it pinned every source to roughly 86% "active" regardless of content. The check that killed it: exposure-normalising the real footage changed the mean frame difference by 64x, and moved the reported quiescent fraction by 0.002. A measurement invariant to a 64x change in its own subject is not a measurement.
Version two is absolute. A pixel counts as changed if it moves at least 12 grey levels. A frame counts as active only if those changed pixels form a contiguous blob covering at least 0.15% of the frame. Both constants are fixed across every source and never fitted to the data. The area gate is what provides noise immunity without per-source calibration, because sensor noise is spatially uncorrelated and does not form large connected regions.
It also passes a known-answer test before reporting anything, and aborts if it fails. Run on a sequence where every frame is provably the same image under a known lighting transform, it reports a duty cycle of 0.0000 and a maximum blob of 0 pixels.
That test produced the best evidence the instrument works. The first version of the test reported 3.85% activity on the supposedly static control and looked like a failure. Every flagged sample fell within one sampling interval of a single frame: the scheduled lights-on step, which we had left inside our own control window. The instrument was correctly detecting a real event. The test was wrong; the instrument was not.
Scoring against what we sealed
The prediction was registered before any footage was measured. Two of four predictions were wrong.
Real quiescent fraction above 0.60: pass, at 0.99. We underestimated badly.
Real longest still run above 60 s: pass, at 849 to 1726 s.
Generated quiescent fraction below 0.10: fail. Actual 0.10 to 0.32; the frontier clip at 0.32 is three times our stated bound.
Primary discriminator, real-to-generated ratio above 5x: mixed. It reaches 9.5x and 5.7x against our own footage, but only 3.1x against the frontier clip, which is our own pre-registered weak-support band rather than confirmation. We are not rounding that up.
The run-length result is the one that survives every framing.
The comparison is not apples to apples, and we said so first
The real capture is an ordinary, barely occupied room under a fixed camera, at 99% quiescent. It is not a hospital room, whose clinical prior is 67% quiescent. It bounds what static-camera reality looks like, not what a ward looks like.
This matters less than it might, because the two independent references sit at 0.99 and 0.674, and generated footage at 0.10 to 0.32 is far below both. The conclusion does not depend on which reference you choose.
What this licenses
It licenses one sentence: generated footage lacks the quiescent structure of real fixed-camera footage, by a measured factor, on an instrument that passes a known-answer test.
It does not license any claim that this harms downstream models. That needs a train-on-synthetic, test-on-real study, which is unrun. The general principle that temporal realism matters is already established elsewhere; the contribution here is a measurement in video, not the principle.
On privacy
The real-footage rows are aggregate statistics only. No frames were displayed, copied, exported or published, and none will be. The source is described only as an ordinary, barely occupied room under a fixed camera. Published with the capture owner's explicit consent.
Code and data
Full write-up, pre-registration, code and raw numbers: huggingface.co/resoajoe/loglens-longvideo-drift
Reproduce: code/eventrate2.py in that repo