How long does AI video take, and how to build longer films from scenes

How long generation takes: our own logs

We logged wall-clock time for each completed run in our own production tool, between 7 September and 21 October 2026. These came from our own copy of MiniMax H3 on Hugging Face ZeroGPU, which has no public queue. Size was 768 × 1344 or 1344 × 768.

A 5-second clip took us about 2 minutes on turbo and 6 minutes on the full model. Five further runs out of 76 did not finish: three were cancelled or failed on a bad input file, and two were lost when our own server restarted. We did not log hosted APIs this way, so we have no equivalent figures for Kling or Wan, and we will not guess them. Other timings from our published tests:

To estimate a job, multiply length by wait by takes. Thirty seconds on H3 turbo at three takes is 30 × 19.9 × 3, about 30 minutes of waiting.

Quota can be the tighter limit: that Hugging Face plan gives 40 GPU minutes a day, and one finished second used about 39 quota seconds, which works out at about 61 seconds of video a day.

Most of the time goes elsewhere

Waiting for renders is the small part. Pictures, judging takes, fixing prompts and editing take far longer. We did not time this consistently, so we will not quote an average. What moves it most is the order you work in: in the dollhouse test, editing an approved picture took 8 requests against 26 messages in a chat that rebuilt the scene each time. That kind of saving dwarfs any difference in render speed.

How long can one generation be?

A long generation is not always better. It can be a good way to get an unbroken take, as in One Prompt Studio, but a single mistake at second 25 means paying for all 30 again.

Five ways to connect scenes

1. Share a frame between two shots

Generate shot A ending on a picture, and shot B starting on the same picture. We did this in the dollhouse scene: 3 seconds, then 6, joined at a shared frame. The composition lined up, but the face had a slightly different skin texture after the join. It works best with a cut or movement to hide the seam.

2. Reuse the same references and cut between angles

Viewers forgive a cut far more than a morph. Generate each shot separately from the same character and location sheets and cut between angles, as in a normal edit. The Lisbon promo was built like this from five shots across three models.

3. Hand off the last frame

Extract the last frame of an approved take and use it as the first frame of the next. It carries the position and light forward, though errors carry too. Tidy the frame before you reuse it.

4. Use a native extend feature

Veo’s extend step on fal is the clearest case. Each extension is another paid generation that can drift, so check the joins.

5. Splice the good seconds

Keep all your takes, failures included, and cut the good seconds together. It only works when sheets keep every element the same across takes.

Rules for a longer film

• Plan the shot list first. Our Nepal explainer was planned as 12 shots, keyframes first, before any video.
• Keep each run small. Six shots and nine pictures made Wan 3.0 skip instructions; two and two came out clean.
• Move the camera on one idea per shot. Short shots only have room for one small move.
• Do the sound last, across the whole cut. Separately generated clips rarely share a room tone.

Written by Oleksiy Onishchenko, who produces AI films for brands and teaches AI video workshops. Founder of easyai.video.