First and Last Frame AI Video: A Keyframe Workflow Guide

Most AI video work fails in the middle. You describe a scene well enough that the opening frame looks right, and then the model wanders: the subject drifts, the camera invents a move you never asked for, and by the last second you are looking at a different picture than the one you started with. The first-frame and last-frame method exists to end that drift. You hand the model two fixed points and let it draw the line between them.

The idea is borrowed from traditional animation, where keyframes have always been the backbone. You set the pose at the start, you set the pose at the end, and the in-between frames are generated to connect them. In an AI video tool the two keyframes are literal images: one uploaded as the opening, one as the closing, paired with a prompt that describes the change between them. The model is no longer improvising a whole scene. It is solving a much smaller, much more tractable problem: how do I get from here to there.

The One Rule That Makes It Work

Everything hinges on a single discipline, and it is worth stating before any example: keep the subject identical and change only the state.

Same character, same object, same camera framing, same art style, same resolution and aspect ratio. What you are allowed to change is atmosphere, lighting, weather, the position of particles, and a modest shift in posture. What you are not allowed to change is the thing the viewer is tracking: a face that becomes a different face, a building that gains a floor, a camera that swings to a new angle. Break those and the model has to warp one image into an incompatible one, and the middle of your clip collapses into distortion.

The cleanest transitions are the ones where a stranger could not tell your two keyframes apart at a glance. A valley hidden by cloud and the same valley revealed is a great pair. A hiker and a different hiker is not a pair at all.

Five Transitions Worth Building

Clouds parting over a valley

Start with a figure standing on a rocky summit, a thick sea of cloud hiding everything beyond. End with the same figure in the same stance, the cloud sunk and pulled apart to reveal the valley below. The motion is slow and downward, the jacket hem lifting in a light wind.

A hiker in a jacket standing on a granite peak, unchanged throughout. Thick clouds slowly sink and disperse downward, unveiling the wide valley beyond. The hem of the jacket flutters gently. Fixed camera, soft afternoon side light, realistic outdoor photography, atmospheric haze, slow natural movement, smooth transition, no shape distortion.

A hiker on a summit as a sea of clouds parts to reveal a valley

This is the ideal teaching example because the only thing moving is weather. The subject is a statue. When you are learning the method, start with a scene where the figure does nothing and the world changes around it.

Light falling across a reading table

Begin with a person seated at a wooden table, head down, a pool of sunlight on one side, dust hanging still. End with the same seated figure lifting her gaze toward the window, the sunlight slid to the other side of the boards, dust now drifting.

A young woman seated at a wooden table in a vintage library, posture stable, slowly lifting her eyes toward the window. Sunlight shifts gradually across the tabletop, fine dust particles float gently, curtains sway slightly. Soft diffused natural light, film portrait texture, minimal movement, fixed camera, stable face without distortion.

Here you are changing light and a single small gesture. The trick with any face is to keep the change tiny. A gaze lifting a few degrees survives; a head turning to a new angle does not.

A sunbeam finding the turtle

Open in dim blue water with an ancient turtle holding still and no light from above. Close with the same turtle in the same spot, one shaft of sun now reaching down to light its shell, plankton drawn into the beam.

An ancient giant sea turtle hovering in the same position in deep blue water. A single sunbeam slowly pierces down from the surface and illuminates the shell. Tiny plankton drift toward the light, subtle water flow, soft underwater light-ray effect, cinematic marine photography, calm and slow motion, subject unchanged.

A sea turtle in deep water as a sunbeam slowly illuminates it

Underwater scenes are forgiving because the medium itself is in constant soft motion. A beam arriving is a lighting change, not a structural one, so the turtle never has to be redrawn.

A balloon filling with air

Start with a deflated vintage balloon slumped on grass, fabric wrinkled, no flame. End with the same balloon at the same spot swelling into shape, a soft orange flame flickering inside, a thin thread of smoke rising.

A vintage hot air balloon resting on a green meadow, position fixed. The envelope slowly inflates and rounds out, a gentle orange burner flame flickers inside, a thin wisp of smoke drifts upward. Fixed camera, warm golden-hour light, realistic landscape photography, slow gentle change, no shape distortion, basket stays steady.

Inflation is a form change, which is usually the risky kind. It works here because the outline grows predictably and the basket anchors the composition. Give any shape-shifting scene a fixed object to hold the frame still.

Tide rising under a beached boat

Open with an old wooden boat sitting fully on dry sand, the sea low and flat in the distance. Close with the same boat, water crept up under the hull, small waves lapping it, ripples spreading.

An old wooden boat resting on a coastal beach in the same position. The tide rises slowly, seawater creeping under the hull little by little, small waves softly lapping the boat, the sea surface undulating gently. Warm dusk light, realistic coastal photography, slow natural movement, unchanged composition, boat stays fixed.

An old wooden boat on a beach as the tide slowly rises around it

Water rising is one of the safest transitions there is. The boat is a still object, the change is a level, and the motion is continuous and slow. If you want a first success with this method, make it a tide.

Writing the Prompt for a Pair

With two keyframes uploaded, the prompt has one job: name the change. Do not spend it describing the scene, because the images already did that. Spend it on the verb.

Describe the motion as slow and directional. Say what moves and, just as importantly, say what stays put: fixed camera, subject unchanged, composition stable. Those reassurance phrases are not filler; they actively suppress the model's urge to add a camera move or reshape the subject. A good keyframe prompt reads like a stage direction with a list of props that must not move.

Keep the aspect ratio and resolution identical between the two images. A mismatch there is one of the most common causes of flicker, because the model has to reconcile two frames that are literally different sizes.

When the Middle Still Breaks

If the clip warps halfway through, the two keyframes were probably too far apart. A person seated and a person standing is a jump, not a transition; split it into two generations, seated to leaning, leaning to standing, and join them. If the camera drifts when you wanted it fixed, say fixed camera twice and remove any word that implies movement you did not ask for. If the face changes, shrink the pose change until the head barely moves at all. And if the whole thing jitters, slow the described motion down and lengthen the duration, because fast interpolation between two stills is where the artifacts live.

The Point

The first-frame and last-frame method is really a request for control. A single prompt leaves the model to invent everything between the start and the end, and invention is where consistency dies. Two keyframes replace that invention with a constraint: get from this exact image to that exact image, and do not change the thing we are both looking at. Give it a subject that stays, a change that is slow and clearly named, and a fixed frame to travel inside, and the middle stops wandering. The transition does the storytelling for you.

Haomings AI Video
Create with Haomings AI Video
Turn photos and text into lifelike videos with Haomings AI Video.