Start and end frames in AI video: how frame control actually works

Most AI video is generated from a prompt alone, which means the model decides where the shot begins. Frame control takes that decision back. You supply the opening image, and optionally the closing one, and the model generates the motion that connects them.

It is the difference between asking for a shot and directing one. Here is how the two modes behave, when each is worth using, and where the limits are.

Three modes, not two

A video request lands in one of three shapes:

• Text to video. No images. The model invents the whole scene. Most freedom, least control, most attempts.
• Image to video (start frame). One image, which becomes the first frame. You describe the motion outward from it.
• Start and end frame. Two images. The first frame and the last frame are both fixed, and the model generates everything in between.

The third mode only engages when both images are present. Supplying an end frame alone does nothing useful — there is no defined starting point to travel from — so the Video Studio treats a lone image as a start frame.

When a start frame is enough

Most of the time. If you have a still you like and you want it to move, a start frame plus a motion prompt is the whole job, and it is the cheapest path to a usable clip because the composition is already settled.

Use a start frame alone for: animating your own generated art, product shots, landscapes, atmospheric loops, anything where the destination does not matter as much as the movement. Our walkthrough on animating a photo covers this case end to end.

When the end frame earns its keep

Fixing both ends is worth the extra work in three situations.

A transformation. Something becomes something else — a sketch becomes a painting, a bare room becomes a furnished one, a closed box becomes an open one. Both states exist as images; the interesting part is the middle.

A cut you have already planned. If clip two has to begin where clip one ends, generating clip one with the correct end frame means the cut is invisible. This is the practical way to build a sequence longer than a single generation, since clips run 4, 6, 8 or 10 seconds and no longer.

A brand-critical closing frame. When the last thing on screen has to be exactly right — a product in a specific position, a composition that a title will sit over — fixing it removes the guesswork.

The rule that makes it work: the two frames must agree

This is where most end-frame attempts fail. The model interpolates between your images; it does not reconcile them. If the two frames disagree about something fundamental, the clip spends its duration morphing awkwardly rather than moving.

Agreement means:

• Same subject. A different person or product between frames produces a melt, not a cut.
• Same lighting direction and colour temperature. Warm to cold across six seconds reads as a fault.
• Same aspect ratio and framing logic. A wide start and a tight end is a legitimate push in; a wide start and a completely different composition is not a move, it is two shots.
• Plausible physical continuity. The model needs a believable path from A to B. Generate both frames from the same seed image or the same character setup where you can — the character consistency guide covers how to hold a subject steady across images.

A good habit: produce the end frame by editing the start frame rather than generating it fresh. Magic Edit in the Image Studio will move, remove or alter one element while leaving everything else identical, which gives you two frames that agree by construction. The image editing guide walks through it.

Limits worth knowing before you plan

Reference images are not available alongside frames. Frame modes are pure-frame: the images you supply define the first and last frame, and you cannot additionally pass a separate style or character reference in the same request. If you need a consistent style, bake it into the frames.

Raw photographs of identifiable real people are refused, in every frame mode. This is a consent boundary. This is a policy boundary to prevent unauthorized generation of real people.

Durations sit on a grid. 4, 6, 8 or 10 seconds. A request for 5 snaps down to 4, which will cut the last beat off a timed transformation. Plan transformations at 6 or 8 seconds; a two-state change rarely reads at 4.

The middle is invented. Frame control fixes the ends, not the path. If a specific intermediate pose matters, generate it as its own clip with its own frames rather than hoping.

A workflow for a two-shot sequence

1. Generate or choose frame A.
2. Edit frame A into frame B, changing only what should change.
3. Generate clip one with A as start and B as end, 6 seconds, at 360p (25 credits) to check the interpolation.
4. If the middle behaves, re-render at 1080p (95).
5. For clip two, use frame B as the start frame, so the cut is invisible.
6. Assemble on the timeline, where the two clips become one continuous move.

That sequence gives you twelve coherent seconds out of two generations, which is how longer AI video is actually built — our guide to how long AI videos can be covers why assembling beats asking for length.

Frequently Asked Questions

Do I need both a start and an end frame?

No. A start frame alone is the common case and usually the right one: it fixes the composition and lets the motion prompt do the rest (especially when keeping product labels sharp in AI video). Add an end frame when the destination matters — a transformation, a planned cut, or a closing composition that has to be exact.

Why does my end-frame clip morph strangely in the middle?

Because the two frames disagree about something the model cannot reconcile — a different subject, a lighting flip, or an implausible physical path between them. The fix is to derive the end frame by editing the start frame, so the two share everything except the thing that is meant to change.

Can I use a photo of a real person as a start frame?

No. Raw photographs of identifiable real people are refused in every frame mode, deliberately. Generated portraits, illustrated characters, products, animals and places all work.

Can I supply a style reference as well as frames?

Not in the same request — frame modes take frames only. Build the style into the images before you generate the video. In practice this is a better workflow anyway, because you can see and fix the style in a still for the price of an image rather than a video.

How many seconds should a transformation clip be?

Six or eight. A transformation needs time to read as one continuous change; at four seconds the viewer sees a jump rather than a transition. Remember the duration grid — asking for five gets you four.

Try it with one edit

The cheapest experiment: generate one image, edit a single element to make the second frame, and render the pair at 360p for 25 credits. If the middle behaves, you have learned the whole mode. A free account starts with 150 one-time credits and no card, and the image-to-video tool page is the shortest way in.