How to extend an AI video: from the last frame to the next clip

Video models deliver clips of 4 to 5 seconds (in the GenBrasil test, none went beyond that). A 15 or 20 second scene has to be assembled. There are two ways: different shots cut in an editor, and continuation, where the end of one clip becomes the start of the next. This article is about the second.

The technique

1. Generate clip 1 (from text or from an image).
2. Extract clip 1's last frame.
3. Use that frame as the first frame of clip 2, with a prompt describing what happens next.
4. Repeat and join the clips.

Extracting the last frame with ffmpeg

# grab a frame near the end of the file
ffmpeg -sseof -0.1 -i scene1.mp4 -frames:v 1 -q:v 1 last_frame.png

Open the image and check: if it is blurred by motion, step back a bit (-sseof -0.3) and choose a sharper frame. Then upload that image as the first frame on GenBrasil and describe only the continuation: “the same woman opens the door and walks into the hallway; camera follows from behind”.

Joining the clips

A hard cut works when there is a shot change. For continuity, a short transition hides the seam:

# 0.5 s dissolve between 5 s clips (offset = duration of the 1st - 0.5)
ffmpeg -i scene1.mp4 -i scene2.mp4 \
  -filter_complex "xfade=transition=fade:duration=0.5:offset=4.5" \
  -c:v libx264 -crf 18 scene1_2.mp4

Both clips need the same resolution, frame rate and pixel format. If they come from different models, standardize first with -vf scale=1280:720,fps=24.

The problem: drift

Each continuation starts from a generated frame, not the original. Small errors accumulate: the face changes a little, the light color drifts, the style becomes something else. After three or four links the difference is visible. How to reduce it:

• Re-anchor every 2 or 3 clips. Go back to the character's original image (or the one generated with a LoRA) instead of the last frame.
• Repeat the fixed description. Clothes, hair, light and lens described identically in every prompt.
• Keep the same model through the chain. Switching models midway changes the style.
• Prefer simple actions. One movement per clip, so the last frame is clean.

When cutting beats continuing

Not every scene needs continuity. A set of short shots (wide, medium, detail) cut in the editor usually looks more professional and is much cheaper, because each shot can be redone on its own. Use continuation when the action must be continuous: a walk, a following camera, a transformation.

Costs

The chain multiplies cost: 4 links are 4 approved generations plus the discarded ones. Sketch the chain on a light model and only then generate the links on the final model, following the model ladder. Some models advertise longer clips (up to 30 seconds in a single shot); if your scene fits that, it may be simpler than chaining.

Frequently asked questions

How do I get a video's last frame?

With ffmpeg: ffmpeg -sseof -0.1 -i video.mp4 -frames:v 1 last_frame.png. If the frame comes out blurred, step back a bit in time.

Why does the face change after a few continuations?

Because each link starts from a generated frame, and small errors accumulate. Re-anchor every 2 or 3 clips with the character's original image.

Is it better to continue the scene or cut to another shot?

Cutting to another shot is usually cheaper and more professional. Continue only when the action must be continuous.

Rather build your own workflow? See the hourly GPU templates.

Read next: post-production on a GPU · how an AI video model works