AI Video Editing Flows

Tests we ran ourselves

Tools we ran ourselves on the same short clips, with the timings, the measured result and what failed. Every number comes from our own runs; nothing is copied from a vendor page.

Tests so far

How we test

  1. Same inputs every time: three 60-second clips from the open film Tears of Steel (CC-BY 3.0, Blender Foundation), with a checksum for each file so anyone can use the same ones.
  2. One cold run, then three timed runs per clip and setting. We publish the median, the spread, peak memory and the CPU load on the laptop just before each run.
  3. A written reference for quality: the film's official English subtitles for words and speech timing; the source frames for anything drawn into the picture.
  4. Every page has a What failed box. Failures and surprises are part of the result.
  5. We never publish a time saved, because that would need a manual baseline we have not measured.

What we test next

  • WhisperX word timings against faster-whisper on the same clips. On 5 Nov it would not install here: its current releases require Python 3.13 or older (this machine runs 3.14), and the older releases pin a ctranslate2 version with no build for 3.14. Next run: a separate Python 3.12 environment.
  • Long-video-to-Shorts tools against highlights we mark by hand first. We looked at both on 6 Nov and could not run either here yet. OpenShorts runs through Docker and its README lists a Google Gemini key as required for its AI features (a local language model can replace it for picking moments only); we run no accounts or keys in these tests. FunClip needs Python 3.12 with PyTorch and its models, and this machine runs Python 3.14. Next run: a Python 3.12 environment for FunClip, and OpenShorts only if it can run with a local model and no key.
  • A clean talking-head recording for silence cutting, where auto-editor should do much better than on a film score.

Page checked on Nov 4, 2026.