Why do actions still feel slow even after using speed-boosting prompts? Often, it is not the intensity of the descriptive words that is lacking, but a lack of control over the video's temporal pacing. You need to orchestrate when the action starts, how long it takes to complete, and what happens before and after.
Part 01 Why stronger descriptors don't necessarily increase speed
This test uses Seedance 2 Fast to generate samples on a standard single-flight staircase. All videos are at their original speed with no audio. The composition requires a fixed camera angle, keeping both the visible stairs and the flat landing at the bottom in the frame. The first group is five seconds long, sharing the same first frame, model, resolution, and random seed, varying only the intensity word in the action sentence:
A. Normal pace: The woman walks down the stairs at a normal walking pace.
B. Quickly: The woman walks down the stairs quickly.
C. Very quickly: The woman walks down the stairs very quickly.
All three videos completed the staircase, but the three levels of intensity failed to create a distinct speed difference: the progress in the latter segments of B and C is only slightly ahead of A, and C shows no significant acceleration over B. In this set of results, simply intensifying the descriptive words did not produce a clear speed-up effect.
Next, we changed the action from walking to running, and further to a full-speed sprint, while adding specific time requirements. This step changes the action type, footfalls, and duration description simultaneously, rather than being a single-variable control. The action description for the sprint version is as follows:
The woman immediately sprints down the visible flight of stairs at full speed, reaching the tiled floor within the first two seconds. She uses rapid, powerful, alternating footfalls with brief contact on each step.
This five-second sprint version did not meet the "reach in two seconds" requirement: the person is still on the stairs at two seconds, with the first foot hitting the flat floor at approximately 3.5 seconds. Although the arms are positioned in a running stance, the total time to descend did not significantly shorten after using more intense action instructions combined with specific time constraints.
Part 02 Tightening the temporal window for actions
We kept the full prompt, first frame, random seed, and other parameters from the sprint version, but reduced the generation duration from five seconds to four seconds. As a result, the four-second version landed on the flat floor at about 2.5 seconds—earlier than the five-second version, but still failing to meet the "reach in two seconds" requirement.
After shortening the total clip duration, we further adjusted the timing of the character's entry and exit. Keeping the four-second duration, we used a empty shot of the same staircase as the first frame, showing the empty stairs first before the character enters to sprint down and out of the frame. The time arrangement in the actual prompt is:
From 0.0 to 0.7 seconds, hold on the empty staircase. At about 0.7 seconds, an adult European woman with tied-back brown hair, a blue T-shirt, dark jeans and white sneakers enters from the upper-left stair opening already sprinting. She immediately sprints down the entire visible staircase at full speed in one rapid continuous burst, with quick alternating footfalls and brief contact on the steps. Complete the visible descent in about 1.3 seconds: by 2.0 seconds both feet have reached the flat tiled floor at the bottom. She keeps running to the right across the floor and exits the frame, leaving the staircase empty again for the end of the shot.
The model did not strictly execute according to the timestamps in the prompt; both the character's entry and arrival at the floor were later than planned. However, continuing to run out of the frame after descending made the action feel more concentrated. The results of the three time arrangements are as follows:
| Version | Time Arrangement | Observation |
|---|---|---|
| Five-second sprint | Descend from the start, required arrival in 2s | First foot lands at approx. 3.5s |
| Four-second sprint | Same prompt/first frame, only shortened duration | First foot lands at approx. 2.5s |
| Four-second delayed entry | Empty frame first, then concentrated descent and exit | Enters after >1s, first foot lands at approx. 2.5s; visible descent takes 1.3–1.5s |
The time of the first footfall is similar for both four-second videos; the main change in the delayed version is that the action is more concentrated after entering. Because this version changed both the blank-shot first frame and the temporal sequence description, the change cannot be attributed to a single factor. For more on arranging the rhythm of entry, core action, and exit, refer to Shot Duration and Beat Design for AI Video.
Part 03 Accounting for the space the character must traverse
Stairs, handrails, and the flat floor in the frame help us judge whether the person is truly moving downward. When writing prompts, besides writing "downstairs," you can clearly state the starting point, the path along the way, and the destination. The following two prompts are examples that have not yet been separately tested; for actual comparison, keep the speed words, first frame, and clip length consistent, observing only the supplementary effect of spatial descriptions:
Minimal spatial detail: The woman walks quickly downstairs and stops.
Clear spatial detail: The woman walks quickly from the top of the visible staircase, follows the steps downward beside the wall-mounted handrail, and stops with both feet on the tiled floor at the bottom. She follows the staircase throughout.
A single reference image can define the start, stairs, and flat ground, but it cannot directly dictate movement speed. For complex scenes, it is recommended to first confirm that the character can naturally complete the path on a simple, single-flight staircase before adding turns or additional segments. The routes and turns in the prompt must correspond to the reference image; if the character is blocked by a handrail or stair slab midway, mark it as "undeterminable."
Part 04 Separate recording of time and action quality
Each group of videos shown above has only one sample generated under its respective conditions. For actual troubleshooting, we recommend repeating the generation multiple times under the same conditions and keeping both successful and failed samples for comparison.
| Record Item | What to check |
|---|---|
| Entry time | When the person first appears in the frame |
| Start time | When the person begins moving down the stairs |
| First foot on floor | When the first foot touches the flat ground at the bottom |
| Descent complete | When both feet have completed the transition from stairs to floor |
| Action naturalness | Sequential footfalls, continuous descent, absence of sliding, teleporting, or stair distortion |
The arrival times in this text are calculated from the start of the video, using the first foot touching the flat ground as the benchmark; these are frame-sampled estimates and are not equal to both feet finishing the descent. The 1.3–1.5 seconds for the delayed entry version describes only the visible descent after entry, excluding movement off-screen. Samples that do not arrive or are obscured should also be recorded truthfully.
Arriving on time does not equal natural action. If the character arrives early due to foot sliding, teleporting, or stair distortion, it does not count as successful control. Also, watch for drift in the character's appearance or the camera position. When comparing versions, it is recommended to play them side-by-side at original speed and muted before dubbing or editing, stopping immediately after completion to avoid interference from loop playback or frantic sound effects when judging the real speed.



