d48 tier, stage-D2 checkpoint at iteration 50,000 — architecture rtmosr_ea_film, 9,722,409 parameters, 2x.
Selected best-by-DISTS at 0.141940 across 11 candidates.
Acceptance — four legs, on identical frames
A checkpoint is accepted by passing every leg, not because one number improved.
| Leg | Result | Verdict |
|---|---|---|
| Sweep (DISTS) | iter050000 at 0.141940, best of eleven | selected |
| Face gate | 0 violations in 60 images | PASS (blocking) |
| Temporal | 5.9147 vs bicubic 6.0240 (-1.8%) | PASS |
| Invention probe | 6.14% vs teacher 8.50%, bicubic floor 2.14% | PASS |
Measured throughput
NVIDIA Jetson Thor — TensorRT 10.13.3, FP16, builderOptimizationLevel=5, CUDA graph enabled. Median enqueueV3 GPU time, compute-only.
| Input | Output | Latency (median) | Throughput |
|---|---|---|---|
| 540p | 1080p | 6.335 ms | 158 fps |
| 720p | 1440p | 12.350 ms | 81 fps |
| 1080p | 4K | 28.742 ms | 35 fps |
p95 and p99 within 0.2 ms of the median at every shape. These are compute-only figures on one named device; they are not a promise about your pipeline or end-to-end playback.
TensorRT parity for this checkpoint: 75.51 dB against the torch fp32 reference (project threshold >40 dB).
Weights
The weights are not attached to this release. They are published on Hugging Face at ltomes/userdevice-rtus, revision v0.1.0, with SHA256SUMS and the sweep, temporal and invention evidence beside them.
Keeping one home for the artifacts is deliberate: two copies of a weights file is two things to verify and one of them will drift.
Full model card
Licence, attribution, training data, degradation recipe and limitations are in docs/MODEL-CARD.md at this tag. They are deliberately not copied here — the card is edited over time and a release note is not.