Three Video Frames Do Not Define Three Equal Time Steps

Separate frame position, presentation time and duration

A delivered video input can be described as three images, an ordered array, or images associated with times. Those descriptions preserve different information. An index such as 0, 1 or 2 locates an element in the supplied sequence. Advancing one index does not, by itself, specify an elapsed interval.

A presentation timestamp, or PTS, places a frame on the media presentation timeline. With an established rational unit a/b seconds per tick, its time coordinate is PTS x a/b seconds. Two valid timestamps expressed in the same effective unit give a presentation gap of (PTS2 - PTS1) x a/b seconds. Frame duration separately describes how long a frame lasts.

This article begins with delivered, presentation-ordered frames and established timestamp meanings. It explains how to compare their timing information. Presentation time does not automatically identify camera capture time, UTC, or the wall-clock moment when processing finished. The related clock guide addresses the separate relationship between event timestamps and clocks.

References: [VT1][VT2]

Read the FFmpeg fields within their stated scope

The named reference is FFmpeg 8.0. AVStream distinguishes time_base, start_time, duration, nb_frames, avg_frame_rate and r_frame_rate. Its start_time describes the first presentation PTS in stream units; stream duration is another field, and nb_frames is zero when unknown. Neither a count nor a stream-wide average supplies every interval. The r_frame_rate base-rate estimate is expressly a guess, so it cannot establish uniform spacing.

AVFrame defines pts and duration separately, with duration zero when unknown. Its best_effort_timestamp is a heuristic estimate in stream units. Preserve that estimated status when describing supplied timing evidence.

AVFrame.time_base also has an explicit qualification: the reference discusses future population by decoders or filters and says encoders or filters ignore it by default on input. A field's existence therefore establishes no universal propagation guarantee. The comparison needs valid delivered timestamps and their actual effective units established for the named path; this article supplies no decoding or export procedure.

References: [VT1][VT2]

VPTS-111: the images survive, but the second interval changes

Fictional example VPTS-111 stipulates three presentation-ordered frames A, B and C with valid PTS values 1000, 1040 and 1120. Their known effective unit is 1/1000 second per tick. These are invented inputs and arithmetic; no video was processed or measured.

A has presentation time 1000/1000 = 1.000 s. B is at 1.040 s and C at 1.120 s. The A-to-B gap is (1040 - 1000)/1000 = 0.040 s, or 40 ms. The B-to-C gap is 0.080 s, or 80 ms. Consecutive positions have unequal time spacing.

Now stipulate a second representation containing exactly the same images in the same order, but reconstructing times from index at uniform 25 frames per second, anchored at A's 1.000 s. One index step becomes 1/25 = 0.040 s. Indices 0, 1 and 2 are assigned 1.000, 1.040 and 1.080 s.

C is now placed 1.120 - 1.080 = 0.040 s, or 40 ms, earlier than its supplied presentation time. Both representations contain three images. Their second interval differs, and their first-to-last spans differ: 120 ms in the supplied timeline versus 80 ms in the uniform reconstruction. An image-count check would miss the change.

The 120 ms span does not establish total clip duration

VPTS-111 supplies the presentation starts of A, B and C. Subtracting A's time from C's gives 120 ms; it leaves the interval after C's start unresolved. The example specifies neither C's duration nor an ending boundary, so total clip duration cannot be concluded from that subtraction.

The 80 ms B-to-C gap also establishes no particular acquisition loss or decoder fault. It is a property of the stipulated delivered presentation timeline. Diagnosing why that spacing exists would require separate evidence about the actual source and transformations.

Every nth frame selects positions, not fixed media-time steps

A consumer selecting every nth frame follows sequence positions. In an irregularly spaced sequence, equal index strides need not cover equal presentation intervals. If the requirement is expressed in media time, the consumer also needs an explicit selection policy: for example, what happens when no frame exists at a requested time, or when several candidates qualify. This article does not choose that policy for a model or task.

The consequence can arise even without estimating motion. Suppose a fictional review requirement includes supplied frames whose presentation starts fall within the first 100 ms after A. On VPTS-111's supplied timeline, C starts 120 ms after A and falls outside that window. Under the uniform reconstruction, C appears only 80 ms after A and falls inside it. The same image has acquired a different temporal meaning.

Preserving timestamps makes that difference visible. It does not establish that a model consumes those timestamps, interprets them correctly, or meets the application's quality requirement.

Compare image-array offers at the consumer's input boundary

Consider two fictional offers for the same A, B and C images, with matching dimensions. Offer A retains the supplied timeline and its frame associations. Offer B supplies an ordered image array plus a 25 fps label. If B's consumer reconstructs the uniform times above, the offers deliver different temporal inputs despite matching image counts and pixels. Neither offer has a stipulated task-quality or performance result.

Ask the provider to show which effective presentation time belongs to each delivered image and how that association reaches the consuming stage. The evidence needs the applicable units and origin mapping, together with the status of missing or estimated timestamps. A timing record detached from the array does not establish which image a time describes. A record retained only before a later transformation leaves the consumer's input unresolved.

Different raw integers need not mean different spacing. In VPTS-111, subtracting the common 1.000 s origin yields 0, 40 and 120 ms and preserves both gaps. A documented exact change of tick units can likewise preserve the timeline. Compare the resulting intervals and frame associations after the stated mapping, rather than requiring identical integer labels.

An exporter may deliberately drop, duplicate or retime frames. Obtain evidence of the transformation actually performed and compare it with the required deliverable; a tool name alone establishes none of those actions. If arbitrary original timestamps have been discarded and no recoverable timing record remains, image count plus a headline rate cannot reconstruct their original intervals.

An image-only array may be sufficient when timing is irrelevant to the agreed task. If the task depends on temporal spacing, an array whose timing has not been established leaves that part of the input requirement unanswered. Matching dimensions, count or even presentation intervals still proves neither model temporal correctness nor overall software quality. The related Wan guide addresses generated video representations and execution work; those output-side quantities do not settle this incoming timeline.

Discuss a separate physical equipment requirement

If a separately selected system also needs equipment receiving, rack integration, onsite installation or IT cabling, Leviathan's Complete Physical Deployment service and related project inquiry provide a route to discuss those physical stages. Share a nonconfidential equipment list, site, required connections and proposed handoff to confirm scope, requirements and availability.

Video preparation, timestamp handling, model integration and application evaluation remain with the appointed software team. The inquiry concerns the agreed physical equipment work.

References: [LS1]