The first half-second of your clip lurches
Opening frames often resolve the still before motion starts, which reads as a lurch. Generate extra length, start already in motion, and trim.
The take is good from about frame fifteen. The problem is frames one through twelve. The subject pops into the pose, the camera hiccups toward the composition, or the whole frame breathes once as if the still were being pulled onto a moving shot. Then the clip settles. The opening is where many image-conditioned generators reconcile the input still with the motion they are about to run.
At the default 25 fps, half a second is roughly a dozen frames. That is enough for a viewer to feel a lurch and not enough for them to enjoy it. The working habit is not a cleverer adjective for smoothness. It is to start the shot already in motion on the page, generate more duration than you will ship, and trim the settle off the front.
What the lurch actually is
Image-to-video is conditioned on a still that is, by definition, a moment of zero velocity. The sampler has to invent motion from a photograph. The first frames are where that invention is least constrained by its own earlier output and most constrained by a static picture, so you get a characteristic family of artefacts:
- Pose snap. The body in the still is not quite the body the motion prompt implies. The model corrects toward the prompt in a handful of frames: a chin that tucks, a shoulder that drops, a hand that finds the mug.
- Camera settle. You asked for a slow push-in. Frame one is the still's exact crop. Frames two to ten hunt for the push. That hunt reads as a hitch, then the move begins.
- Texture lock-in. Fine detail (hair edge, background, a print) crawls once as the clip decides what the still "meant," then holds. Related to boil, but concentrated at the head rather than throughout.
- Energy delay. The prompt said "she walks." She stands, then walks. The still was a stand, so the walk starts late, and the start is a lurch rather than a stride.
This is the inverse of tail drift. The tail is where accumulated error takes over. The head is where the still still has too much say. Both get worse on long, ambitious clips; only the head is fixable by throwing the first half-second away.
A useful check: export frame one next to a frame from one second in. If they differ in pose or crop more than the stated action accounts for, the opening is a settle, not part of the shot. Do not grade it, do not prompt "no lurch," and do not ship it as the hook. Cut it.
Write the shot as already underway
Prompts that begin the action produce a wind-up. Prompts that continue an action produce a clip you can open on.
| Avoid as an opening | Write instead |
|---|---|
| "She starts to walk toward the counter" | "She is mid-stride toward the counter, weight on the front foot" |
| "The camera begins a slow push-in" | "Slow push-in already in motion toward her face" |
| "He looks up and then smiles" | "He is looking at camera, mid-smile, as he finishes a nod" |
| "Steam starts to rise from the cup" | "Steam is already rising from the cup in a thin, steady line" |
| "She begins to turn her head" | "Her head is already turning toward camera, halfway through the glance" |
Present continuous, already in progress. The still should support that lie. A planted, square-on hero still of someone standing still will fight "mid-stride" and you will get the lurch plus a warped leg. Either pick a still that already has the weight shift (a frame grab from a previous take is ideal) or write an in-progress action a planted body can do: a blink, a breath, a smile that is already there, hair already moving.
Put one action and one camera move in the brief, same as any other image-to-video shot. Then add a clause that denies the wind-up: "The motion is already underway on the first frame; there is no settle, no pause, no start." That clause is weaker than a still that implies motion, and stronger than nothing. Use both.
Generate long, ship the middle
The production rule: the duration you request is not the duration you publish. Request enough length that a half-second (sometimes a second) can come off the head and still leave the beat intact.
Practically:
- Time the beat you actually need, then add a buffer on the front. A three-second gesture is a four-second generation. A five-second line is a six-second generation if the model offers that rung.
- Prefer a duration ladder with leftover seconds over a model that only offers the exact length you want to ship. Whole-second rungs make "generate 6, keep 5" a setting. Clips longer than five seconds cost more because video is billed by what it actually renders. The buffer is still cheaper than a reroll of a clip you cannot open on.
- Do not steal the buffer from the tail if the tail is the part that drifts. Coherence often degrades as the generation runs. The usable window is a middle slice: after the settle, before the morph. Where motion coherence breaks as clips get longer is the map for the tail. This post is the map for the head. You may be cutting both ends. That is a successful take with handles, not a failed one.
- Trim on a preview, not on a guess. Drop the clip on the timeline, mark the first frame that already feels like the shot, and cut there. The editor's 480p preview pass runs at no credit cost and carries a short per-user cooldown; the final export is charged once. Scrubbing at preview resolution is enough to find a lurch. It is not enough to judge pores. You are answering a timing question.
If the model only offers 5 or 10 seconds and your beat is 5, generate 10 and keep the middle. A long generation you trim is still one job. Two 5-second generations you hope will open cleanly is two jobs and two lurches.
First-last-frame is the other structural fix when you can author the start. Feed a first frame that is already a moment of motion (extracted from a previous take, not a posed still) and a last frame that is the landing. The sampler no longer has to invent velocity from a freeze. The opening still can lurch if that extracted frame is a freeze in disguise, so pick one where the body is between poses.
When not to trim, and what to do instead
The lurch is the whole clip. If the hitch never settles, you do not have handles. You have a still fighting the brief. Change the still or shrink the action. Trimming a four-second lurch leaves nothing.
The lurch is a teleport. A body that changes rooms in the first half-second is a position snap, not a settle. Pin both ends; do not just cut the pop and hope the new frame one matches the previous shot.
You need that first frame on screen. Thumbnails, hold-on-first-frame platforms, and any edit that freeze-frames the opener will display the still, not the settled shot. In that case the still is the first frame of the movie and it has to be a frame you will show. Restage the still until freeze-framing it is acceptable, then animate. Trimming does not help a product tile that shows frame one.
Audio starts on frame one. If you generated native sound, trimming the head also trims the line. Either write a half-second of room tone into the performance ("a breath, then the line") so the cut lands on speech, or attach the voice in the edit after picture is timed.
Tail handles are a different policy: they give you a cut point before drift. Generate with both when the duration ladder allows it, and treat the published clip as the slice between them.
FAQ
Is the lurch worse on image-to-video than on text-to-video?
It shows up more clearly on image-to-video because the first frame is a photograph the rest of the clip has to become. Text-to-video can invent a first frame that already has velocity, which hides the settle and leaves you a first frame you cannot match later. If you need a specific opening image, accept the settle and trim. If you need opening energy, start from a frame grab that already has it.
How much should I take off the front?
Start with half a second and watch. If the hitch is gone, stop; cutting more just shortens a good take. If it is still there at one second, the still and the brief disagree and a longer trim is covering a generation problem. Fix the input rather than throwing away two seconds of every clip.
Will a higher frame rate reduce the lurch?
No. More frames in the same half-second is a smoother picture of the same settle. You will still cut it. Default generations here run at 25 fps; interpolating the head before you trim just gives you more frames to discard.
Can I prompt "smooth start, no sudden movement" instead of trimming?
Write the already-in-motion clause; it helps. Do not rely on it as the whole fix. The still is static, the sampler knows that, and a negative about sudden movement is the same family of instruction that leaks the concept it names. Handles are mechanical. They work on takes the prompt only partly saved.