A short moment, grown patiently one piece at a time.
You have a lovely five-second clip: a cyclist on a coastal road, steam curling off a café counter, a slow push-in on a lantern. It looks right, and it is far too short. Asking for a 25-second video in one go usually gives you something looser and less controlled than the clip you already like. A steadier route is to grow the video in short pieces, each one picked up exactly where the last one stopped.
This guide walks through that method: planning, writing each step, budgeting, and knowing when to stop.
How last-frame continuation works
Most extension tools use the same trick. They take the final frame of your clip and treat it as a still photo. That photo becomes the starting image for a new image-to-video generation, guided by a prompt you write. The new piece is then joined to the original, and the duplicated frame at the join is trimmed so you do not see a stutter.
Two consequences follow:
The new piece only "knows" what is in that last frame, plus your words. It does not remember the motion before it.
Whatever is wrong with the last frame (blur, a half-closed eye, a hand mid-gesture) gets inherited by everything that follows.
Some systems instead use the tail of the clip as a short reference video, which carries more sense of motion. Either way, the ending of your clip is the starting line for the next piece.
Step 1: Plan the beats before you extend
Write the whole sequence as short beats, one per piece, before you generate anything. For a 25-second clip built from a 5-second original, you need four more beats.
For the cyclist:
Rides along the cliff road, sea on the left (your existing clip).
Road curves inland, a stone wall appears on the right.
She slows near a small harbour village.
She stops and looks out at the boats.
The camera rises slowly to a wide view of the bay.
Each beat should change one thing: the location, the action, or the camera. If a beat changes all three, the model has too much room to wander and the join will show it.
Step 2: Choose a clean ending frame
Look at the last frame of your source clip before you extend. Scrub back a few frames if you need to. You want:
No motion blur on the main subject.
Eyes open, hands in a natural position.
The subject fully in frame, not cut by the edge.
Light that matches the rest of the shot.
If the final frame is poor, trim the source by half a second and extend from an earlier, cleaner frame.
Step 3: Write the prompt for the next beat only
This is where most people go wrong. They paste the original scene description again, and the model tries to restart the scene. The picture in the last frame already carries the setting. Your prompt should describe what happens next.
Compare:
Too much: "A woman in a yellow jacket cycles on a coastal road at golden hour, sea on the left, cliffs, seagulls, cinematic."
Better: "She keeps cycling as the road bends right, away from the sea. A low stone wall slides into frame on the right. Same camera, tracking from the side at a steady pace. Warm low sun, long shadows."
Three example prompts for different kinds of beats:
Continuing motion: "The cyclist eases off the pedals and coasts down a gentle slope toward a harbour village. Camera keeps tracking from the left at wheel height. Same warm evening light."
Quiet detail shot: "Steam rises from the espresso machine and drifts left across the frame. The barista's hand sets a white cup on the counter. Camera holds still, slight shallow focus. Soft window light from the right."
Camera reveal: "The camera slowly cranes upward and backward, revealing the whole harbour behind the lantern. The lantern flame stays steady. Dusk light, cool blue sky with a warm glow below."
Notice the pattern: what moves, what the camera does, what the light does. Nothing about the character's clothes or the whole room.
Keep camera direction and light consistent
If your last frame shows a camera tracking left to right, say so again, or the new piece may drift into a different move and the cut will feel like two shots glued together. The same goes for light. Name the direction and mood in every prompt ("low sun from the left", "overcast, soft") instead of assuming it will carry over.
Changing the camera on purpose is fine, but do it at a beat where the story changes, and change only that.
Choose length per hop and resolution
Shorter hops are easier to steer. Many models allow takes of 15 or even 30 seconds, but a five-second hop gives you a natural place to check the result and redo just that piece. Longer hops make sense for calm, slow scenes.
For resolution, extend at the resolution you plan to publish. A 480p extension joined to a 1080p original will look soft at the seam.
Budget with a worked example
A tool such as the AI video extender on SpicyAPI bills only for the seconds you add, at the model's normal per-second price. Your source clip is not billed. The numbers below are public prices at the time of writing (October 2026), for five new seconds per hop.
Say you want to go from 5 seconds to 25 seconds, which is four hops of five seconds:
Wan 3.0 at 480p: $0.225 per hop, so about $0.90 for four hops.
Wan 3.0 at 720p: $0.45 per hop, so about $1.80.
Wan 3.0 at 1080p: $0.90 per hop, so about $3.60.
MiniMax H3 at 768p: $0.185 per hop, so about $0.74.
Then add a retry allowance. Plan on redoing one hop in four. For the 720p route, that is one extra $0.45, bringing the realistic total to about $2.25. A sensible habit: test your first beat at a lower resolution to confirm the direction works, then run the final sequence at your delivery resolution.
Know when to stop and cut
Every hop starts from a slightly softened, slightly altered frame, and small errors compound. Faces shift, colours warm up, textures turn waxy. Around the fourth or fifth hop in a row, you will often see it. Signs to watch for:
Skin or fabric looks smoother than in the original.
The light colour has crept toward orange or grey.
Background details change between pieces.
When that happens, do not push on. End the sequence there, then start a fresh clip from a new still of the scene and cut between the two in an editor. A hard cut to a different angle reads as a deliberate edit, and it resets the quality.
Short checklist
Write all beats on one page before generating.
Each beat changes one thing only.
Check the last frame: sharp, subject fully in frame, eyes open.
Prompt describes what happens next, not the whole scene.
Repeat the camera direction and light in every prompt.
Use five-second hops for anything that needs control.
Extend at your delivery resolution.
Budget one retry for every four hops.
Stop and cut when faces, colour or texture start to drift.
Extending is closer to editing than to one big request. Plan the beats, keep each prompt small, and one good moment can grow into a full scene.