A five-second output setting and a prompt that says “turn toward the camera at two seconds” describe different requirements. The first sets the clip duration. The second asks for an event inside it.
Before comparing video generators, define what counts as that event and how you will record it. This is a test-design workflow adapted from FablessBridge's published note. The example task has not been run, and the numbers below are arithmetic examples—not measured model results.
1. Define the event before writing the score
“Turns toward the camera” can mean the head starts moving, the face reaches a side view, or the gaze reaches the camera. Those moments produce different timestamps.
Pick the observable condition first. If completion is what matters, describe the completed state in the prompt and use that same condition during review. Record the requested output duration separately from the event timestamp. A dedicated timing control, where available, is a separate input from an ordinary sentence in a prompt.
2. Change one condition
An unexecuted example from the source uses a single action: a person turning their face toward the camera from the same starting image. Compare a prompt without a time instruction against one with it. Keep the starting image, framing and requested clip length consistent.
Do not change the camera move and the action while also adding timing instructions. Otherwise, the comparison no longer isolates the timing instruction. Even if a seed can be fixed, retain the actual settings and outputs rather than assuming identical results.
Before submitting any jobs, choose the number of trials, cost cap and stopping conditions. This article does not ask you to generate a batch or claim that one model will follow the timing reliably.
3. Keep a record that allows “unfinished”
The source proposes these fields; a CSV header is a convenient way to preserve the same structure:
trial_id,input_conditions,requested_time_s,observed_start_s,observed_completion_s,reason_not_judgeable
Keep the model, provider route, date and source version with the record. Use the same observation rule across clips. If the action never finishes, mark it unfinished instead of choosing a convenient intermediate pose as completion. An unjudgeable result needs a reason; it should not silently disappear from the comparison.
4. Inspect the unedited output
Observe the original clip before trimming or speed changes. Watch the motion at normal speed, then step through the relevant frames to locate its start and completion. If only an edited timeline survives, do not use its timestamps as direct evidence about the original prompt.
For a completed action, the difference is:
timing difference = observed completion time - requested completion time
A requested time of 2.0 seconds and an observed completion at 2.4 seconds would give +0.4 seconds. That is an illustrative calculation, not a measured clip.
5. Separate timing accuracy from whether the shot is usable
A timestamp alone cannot tell you whether the action is coherent or visually natural. Record whether the action completed and whether its appearance is usable, alongside its timing difference. One close result does not establish reliable behavior across a model, provider route or later version.
For a shot that must match a musical beat or a caption, editing can also be part of the solution. Keep “the generator met the requested time” separate from “the editor aligned the shot afterward.” Both can produce a usable sequence, but they describe different evidence.
Put the acceptance condition in the brief
When commissioning a product or service film, identify which action must coincide with a caption or soundtrack cue. That gives the production team something concrete to review and helps decide whether timing should be addressed during generation or editing.
Original Japanese procedure · FablessBridge production examples
To discuss coordinated motion, captions and sound for a video, send a production brief or email fabless-bridge@shian.space.
Disclosure: This article was prepared with AI assistance from a published FablessBridge procedure. No new generation experiment or performance benchmark was performed for this post.
Top comments (1)
Interesting approach—measuring the actual event timing rather than just clip boundaries gives a much clearer picture of model consistency.