ADR 0005: SeedVR2 Video Quality Proof Requires Five-Second Reader-First Clips¶
Status: Accepted.
Context¶
Recent SeedVR2 video work produced several misleading proof shapes:
- sub-second clips could show that a command ran, but they did not reliably expose cadence errors, time-mixing, or motion distortion;
- sampled contact sheets and heuristic scores were useful review aids, but they were not enough to let a reader judge whether movement stayed coherent;
- terse labels such as
wavelet,lab, andoffwere implementation names, not reader-first explanations.
For video restoration quality, the reader has to see motion. If the proof is shorter than a few seconds, it is too easy to miss the exact failure modes that matter most: skipped-motion feel, time dilation, repeated cadence artifacts, or frame-to-frame distortion.
Decision¶
For SeedVR2 public video quality claims, MLX-Gen uses a reader-first minimum proof duration of five contiguous seconds.
The rule is:
- any public quality claim about a SeedVR2 video profile, checkpoint, or comparison must use at least five seconds of contiguous source video at the original FPS;
- shorter clips are allowed only for narrow internal diagnostics and must not be presented as the primary quality proof;
- full-length video runs may still be kept as host-safety or long-run stability evidence, but they are secondary to the five-second reader-first proof when judging visible quality.
Required public proof artifacts for SeedVR2 video quality claims are:
- restored MP4 files for the compared candidates;
- one comparison MP4 when more than one candidate is being compared;
- one contiguous motion-strip sheet;
- at least one detail crop sheet focused on moving subjects or other distortion-prone regions;
- a readable validation report that explains what the labels mean, what profiles were rejected, and what conclusion the reader should draw.
Reader-facing labels must be intelligible. Raw implementation tokens may appear in parentheses, but the visible description must explain the behavior in plain language.
Consequences¶
Positive¶
- Readers can judge motion quality directly instead of inferring from a sub-second clip or a single score.
- Public docs distinguish between route proof, long-run stress proof, and visible quality proof.
- SeedVR2 validation reports become easier to audit and harder to overclaim.
Negative¶
- Public validation takes longer to produce than a tiny smoke clip.
- Some earlier proof artifacts become internal audit evidence instead of public-quality evidence.
Neutral¶
- Tiny clips and sampled metrics still remain useful for fast local debugging.
- Long-run full-video runs remain valuable for memory, host-safety, and output-integrity checks.
Enforcement¶
- Docs, backlog completion notes, and release evidence must not present sub-second SeedVR2 clips as the primary proof of video quality.
- New SeedVR2 comparison claims must link the exact five-second proof bundle and readable report.
- Review should reject terse labels when a reader cannot understand what a mode or profile does without reading code.
- Comparative family claims must be grounded in direct visual proof, not heuristic ranking alone.
Validation¶
Compliance is validated by:
- preserved five-second proof MP4s and comparison MP4s;
- preserved motion-strip and detail-crop review sheets;
- a checked-in validation report with human-readable labels and conclusions;
- direct visual review of the preserved proof artifacts.
Backlog links¶
- Route implementation: 0032 SeedVR2 video restoration and upscaling
- Follow-up audio work: 0046 SeedVR2 video audio copy-through
Related¶
- ADR 0004: SeedVR2 Video Host Safety And Proof Boundaries
docs/upscaling.mddocs/assets/validation/seedvr2-video-2026-06-21/