Skip to content

Code Video Model Official Repo

JavaScript
4.6 KB
Raw Blame
CRITERIA: motion_continuity, artifact_freedom.
Judge the candidate's visible sequence against intended continuous motion.
The raw prompt supplies intent, not evidence that its requested motion actually occurs.
A wrong but smooth camera path or different composition is not a technical-quality
defect; those are Structure issues. Likewise, low-poly/shader-like appearance alone
does not imply low Quality if edges, geometry and motion are stable.
A smooth path must still be physically and perceptually natural: unintended
camera intersection with opaque solid furniture/walls, implausible acceleration, near-plane
clipping that obscures the scene, or a disorienting approach into a surface are Quality
defects, not excused by continuity alone. A deliberate close-up is not the same as
passing through a solid object. Judge this distinction from visible evidence.
Do not penalize a coherent static shot solely because the source or prompt calls
for motion; missing requested movement is a Structure failure. Actual accidental
freezes or jerky stop/restart events are Quality defects when visibly supported.

motion_continuity: assess continuity of viewpoint and framing, natural velocity
and acceleration, plausible subject trajectories, and stable scene progression.
Distinguish (a) ordinary regular stepping at low native FPS,
(b) actual abrupt viewpoint/scene jumps, cuts, teleportation or unstable trajectories,
and (c) hosted sampling gaps. FPS metadata alone is NEVER proof of a jump/artifact;
identify concrete discontinuity in visible scene geometry or motion. Do not claim a
smooth full sequence merely because a few sampled stills look coherent.
Differences between frames roughly one second apart cannot by themselves establish
a sudden jump: a natural turn or head-bob can happen between them. For such gaps,
describe the motion as uncertain rather than declaring a cut without additional
visible discontinuity evidence. Confidence must reflect this uncertainty.
artifact_freedom: explicitly inspect geometry/topology or anatomy deformation,
object identity/count persistence, unexplained appearance/disappearance, flicker,
texture crawling, melting/smearing, duplication/ghosting, edge tearing, abnormal
blur and compression. Distinguish plausible occlusion, intentional blur and intended
effects from failures. For a significant defect give its visual symptom, location,
phase/timestamp and whether local/transient or persistent.
Additional consecutive-frame pairs and endpoint stills are provided. Use them to
check object continuity where a sparsely sampled video impression is ambiguous.
Uniformly sampled intermediate frames are also provided to bridge the longer gaps.
A subject turning from side-on to front-on is not itself a camera cut. Track the
background landmarks and the subject separately across intermediate timestamps.
Only call a transition a jump/cut when visible discontinuity cannot be explained
by the intervening camera/subject movement; do not ignore supplied intermediate frames.
Compare the SAME physical part over close timestamps. A changing visible face,
foreshortened rigid component, perspective enlargement, an object leaving the field
of view, natural folds or a curved soft surface are NOT automatically deformation.
A catastrophic geometry claim requires visible loss of coherent shape/connectivity
that perspective, material flexibility and occlusion cannot plausibly explain.
Do not use "perspective" as a blanket excuse either. Compare persistent landmarks:
rigid straight edges, connected corners, proportions within the same surface,
object identity, and texture attachment. Slow melting can be smooth between adjacent
frames yet still fail across the full shot. Newly visible content must be consistent
with plausible occlusion/reveal rather than simply assumed correct because the camera moved.
Minor warping or imperfect fine details in otherwise coherent objects is not
catastrophic. Give the concrete impossible change, not just words like "melting".
If the supplied close-time evidence contradicts a broad impression from video
sampling, use the actual visible evidence and reduce confidence in that impression.
If the candidate changes abruptly into incoherent debris, noise or a different
scene, do not praise those as realistic effects merely because the opening is good.
But do not invent such a failure if it is not actually visible.

Evidence must cover start/middle/end, not only the opening. Each check's reason
must explicitly distinguish visible problems from native-FPS and sampling limitations.
The final Quality mean is capped at39/19 when either criterion has a major/critical
failure, so one good aspect cannot hide an unusable sequence.