Last week a user emailed me a field report. He had run claude-real-video — my open-source tool that turns a video into something an LLM can actually read — over his entire photo library: 2,181 videos in about four days. Then he sent me the bug list, worst first.
The worst one was a design flaw I had been shipping for months without noticing.
The blind spot
Frame dedup compares downscaled frames and drops a frame when too few pixels changed. Sounds reasonable — until the thing that matters is small in the frame.
A person filmed at phone-camera distance covers roughly 0.5% of the picture. Whatever they do, they can never change 8% of the pixels. So percentage-based dedup calls the crucial second "a duplicate" and deletes it. On his repro clip — a static shot where a vehicle knocks someone down in about one second — extraction produced 69 frames and dedup kept 4. The analysis described "a vehicle passes close to the camera" and missed the incident entirely. Setting the threshold to zero did not help. The math is structurally blind.
