Code Video Model Official Repo
JavaScript
CRITERIA: composition, camera_motion, prompt_alignment.
SOURCE is the Three.js layout/camera reference. CANDIDATE is the generated video.
The prompt is the intended content, action and camera specification.
For every evidence entry ADD: source_seconds, source_observation, prompt_clause.
source_observation describes what is ACTUALLY VISIBLE in the SOURCE at that time;
observation describes the CANDIDATE and explicitly names matches/discrepancies.
prompt_clause is a verbatim relevant substring of the quoted prompt, or null when
that comparison has no corresponding prompt clause. Do not paraphrase a quote.
Compare corresponding normalized progress while keeping both real timestamps.
Paired still views are provided at the same normalized progress. Use these for
direct layout comparison, and use the complete videos for motion/event progression.
composition: compare subject center and relative size in the frame, horizon/vanishing
point, camera elevation and visible object faces, foreground/background ordering,
occlusion, relative left/right/front/back layout, and progression of these properties.
'Both contain the same scene/objects' is not enough for a high score.
Inspect a scene-specific layout signature: relative order and screen-space size
of salient objects, distinctive openings/obstacles, depth/occlusion relationships,
and how these change. Describe deviations, not merely shared object categories.
Do not deduct for textures, shading, exposure or realistic-vs-rendered appearance
when geometry and framing agree. Only actual layout/visibility differences count.
camera_motion: identify source and candidate direction/path (pan, truck, dolly, orbit,
crane, roll, zoom), viewpoint/elevation changes, trajectory extent, parallax and
timing of turns or stops. A forward shot vs an orbit, opposite orbit directions,
or a static shot vs requested motion is a central mismatch even if appearance is good.
Score composition and camera_motion against SOURCE independently of prompt_alignment.
Matching the prompt must NEVER excuse or reduce a source-alignment discrepancy in
those first two subscores. If prompt/source conflict, a candidate may legitimately
score high on prompt_alignment and low on camera_motion; report both honestly.
prompt_alignment: explicitly evaluate the requested composition/camera clauses,
subjects, actions, event order and intentional edits. A bag-of-objects match is not
enough. Identify prompt/source conflicts instead of silently picking one reference.
Do not impose an action not stated in the prompt or visible in the source.
Never repeat a requested orbit angle, action or time-freeze as an observed fact
unless the visible sequence supports it. If the main subjects become unobservable,
do not claim their posture/action remains correct. Required behavior that is
unverifiable must be identified as unverified, not praised as completed.
Report visible agreement AND disagreement. Generic scene descriptions without
specific paired source/candidate evidence cannot justify 90+ camera alignment.
A major/critical failure in any central structural criterion caps the final
Structure score at 39/19, even if other criteria score well.