Carry the time reference
An observation needs its connection to the source shot. Time references make the evidence actionable during editing.
The compact view stays connected to the original material.
W.02 / AI × MOVING IMAGE
Morphena
Let an editing agent inspect a film without swallowing every frame.
SOURCE / GITHUB · 2026-09-15Open source ↗
WHY / DESIGN / BUILD
What shaped the design.
Compile adaptive frame selections and motion evidence into contact sheets and structured records, retaining time references back to the original material.
Carry the time reference. An observation needs its connection to the source shot. Time references make the evidence actionable during editing.
Spend attention unevenly. Stillness and motion do not need identical treatment. Adaptive selection is the central design direction.
Why this
was worth building.
A smaller representation should help an agent make decisions while preserving a route back to the source. That is the question behind Morphena.
Video is dense. An editing agent has a limited context window, but it still needs evidence of what happened, when it happened and how the shot moved.
The underlying
logic.
Compile adaptive frame selections and motion evidence into contact sheets and structured records, retaining time references back to the original material.
01Source video
02Evidence selection
03Timed records
04Edit decisions
The decisions
inside it.
Spend attention unevenly
Stillness and motion do not need identical treatment. Adaptive selection is the central design direction.
Compression becomes an evidence-design problem whose coverage must be evaluated.
What you
can inspect.
A Python project and agent workflow for compiling, auditing and working with time-referenced video evidence.
Inspect the repository ↗
The repository documents experiments and optional models. This portfolio does not claim an independently verified compression or coverage benchmark.
Have a related
problem?
Tell me what you want to make possible, where the difficulty is, and what a useful result would look like.