One character across shots.
Preserve approved character attributes and object counts. Our sequences expose changes in appearance, added objects and inconsistent shapes.
VISION RESEARCH · GENERATION AND REPAIR
Keep the character.
Repair the shot.
A generated sequence can change a character’s appearance, add an object or lose the intended action between shots. Within our vision research, the “Making AI Work” video programme investigates those failures and ways to correct them.
The work has produced prototype sequences, documented failure cases, local repair implementations and verification procedures. Comparative performance and generalization across scenes remain research objectives.
ONGOING MODEL RESEARCH
Understand a sequence.
Assess a correction.
Nemotron and Gemma multimodal models are candidates for inspecting frames and dialogue against scene requirements. Apple’s AFM 3 family adds a separate framework-based research path.
Explore the models and access options ↗The intended order of events.
Study contact, motion and timing: an action finishes, a character looks, then reaches. A plausible frame can still belong to the wrong sequence.
Correct a selected region.
Investigate local repairs that retain approved parts of a shot, its timing and audio. Broader comparisons with regeneration are proposed next.
Dialogue at the right moment.
Examine speech timing, expression and action alignment. Lip-sync pilots and wider validation remain part of further development.
THE EXPERIMENTAL ENVIRONMENT
“Making AI Work”
as a set of test cases.
Short narrative sequences give the programme specific constraints: an object should retain its shape, a character should remain recognizable, and a correction should preserve already approved content.
AI-assisted workflows support reference preparation, version tracking, editing, clip assembly and quality checks. Human review governs creative acceptance. A file that renders successfully can still need visual or dialogue corrections.
We study how to control generated sequences and use agents to prepare, edit and review video.
PROPOSED COMPARISONS
What would make
the method useful?
Compare structured scene requirements with ordinary prompting. Compare local repair with reference-guided and whole-shot regeneration. Compare matched manual and agent-assisted production tasks.
Measure constraint violations, accepted shots, unintended changes, correction effort and cost. Include unsuccessful attempts, not only the final accepted sequence.
Read the methods and evaluation proposal ↗