AI Video Quality
Ten lessons on making an AI-animated series without AI slop. Small shots, locked references, automatic checks on one Mac, a judge that asks narrow questions, and one human watch.
Written from a real production run in November 2026. The checks were tested on episodes with known faults before they were trusted. Model and tool behaviour changes, so re-test on your own footage.
- 01 What AI slop is, and why a whole season failed A kids’ animated series spent a month of credits and produced no usable episode. The failures were not the model’s fault. They were choices about scope, checking and who decides. 6 min read
- 02 Design shots the model can actually make The cheapest slop to fix is the slop you never ask for. Write the script around what the model does well, and keep every shot small enough to judge. 5 min read
- 03 One maker, one reviewer, one approver Agents that can all give orders will undo each other. Give each job to exactly one place, write the rules down, and let only the owner say go. 5 min read
- 04 Lock pictures, not adjectives Words drift; pictures hold. Build every new frame from an approved reference pack, keep a ledger of what has happened in the story, and never build from the last clip. 5 min read
- 05 Assemble without creating new slop Some of the worst slop in this project was made in the edit, not by the model: ghost images from dissolves, dead silence between cuts, and lip sync that drifted almost three seconds. 5 min read
- 06 A scan that reads every frame A small Python script reads every frame of an episode in about a minute and marks the moments most likely to be slop. It finds where to look. It does not decide. 6 min read
- 07 Read every word on screen Writing is where generated video fails most visibly, and it is one of the few things a machine can check almost perfectly: read it, and compare it with the script. 5 min read
- 08 Check every spoken line against the script Speech-to-text on the Mac writes down every word with its time. Lining that up with the script finds missing, wrong and mumbled lines in under a minute. 6 min read
- 09 A judge that asks narrow questions Vision AI judges miss far more than they invent, and they lean toward “looks fine”. Ask them small, specific questions built from the script, and make two signals agree. 6 min read
- 10 The fix loop, and the one human watch Findings become a timestamped fix list, fixes come back as a new version with a change log, and only what changed is checked again. A person watches once, with sound. 6 min read