Video frame interpolation
- Score
- 47.1
- Wallet spend
- $13.96 of $50
- Model cost
- $23.13
- Agent time
- 3h 58m
- Steps
- 363
- Tokens in / out
- — / —
- Self-checks
- 2 passed of 2 used
- Rank at $50 wallet
- 10th of 19
Score
47.1raw metric 0.471
The score is the raw metric (balanced LPIPS and FloLPIPS gap closed against a frozen baseline), which already runs from 0 to 1, times 100.
Verifier details
- scoring_protocol
- vfi-lpips-flolpips-two-anchor-v2
- metric_protocol
- vfi-lpips-alex01-flolpips-sequence-v1
- psnr_cap_db
- 100
- metric_seconds
- 336.6789
- scoring_seconds
- 563.0259
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- 2353fd1bb3cea476fccd980316bbc30df48568f60961b79f964b6e3106491a8f
After grading, a reviewer model reads the submission's files and the agent's trajectory and checks each of the task's rules. A failed check makes the score 0. Step numbers link into the trajectory.
✓non specialist initializationThere is one pretrained component: SEA-RAFT MemorySlices/Tartan-C-T-TSKH-spring540x960-M, pinned at rev eb97ef34 with sh
There is one pretrained component: SEA-RAFT MemorySlices/Tartan-C-T-TSKH-spring540x960-M, pinned at rev eb97ef34 with sha256 cb8cfbf1…. It is a standalone optical-flow model, which the policy allows. The trajectory shows it fetched through hf_hub_download and copied with shutil.copy2 (trajectory#140-155 chunk, trajectory#91), and the final sha matches the declared one (trajectory#352). VFIRefinementNet was trained from random initialization in this session: V1 came from train.py with no --resume-ckpt (trajectory#130 chunk). V2 warm-started from V1 and V3 from V2 (trajectory#256-271, trajectory#316-330 chunks). The shipped refiner is a 0.75/0.25 weight average of V2 and V3, saved in trajectory#347. Its sha 6578ceda… matches architecture.json. Its tensor inventory (adapt_f*, bottleneck.*) matches the custom refiner and no RIFE/FILM/GIMM key structure. No specialist VFI checkpoint was downloaded or used as a teacher, and teachers_and_fallbacks is empty.
✓learned frame generationserver.py calls HybridFlowVFI.interpolate_sequence. vfi_model.py computes SEA-RAFT flow between the two endpoint frames,
server.py calls HybridFlowVFI.interpolate_sequence. vfi_model.py computes SEA-RAFT flow between the two endpoint frames, does multi-scale softsplat/backwarp, and runs the learned VFIRefinementNet, which predicts residual intermediate flows and softmax blending of candidates. The output frames are synthesized by the network, not copied or looked up.
✓independent intervalsThe server accepts exactly two input frames (index 0 and frame_count-1) per request. prepare_interval is computed once p
The server accepts exactly two input frames (index 0 and frame_count-1) per request. prepare_interval is computed once per request from those frames only, then interpolate_time runs for each target time. No state is kept across requests; warmup uses dummy tensors only.
✓no evaluation frame lookupNeither server.py nor vfi_model.py contains a lookup table, hashing of inputs, network calls or stored evaluation frames
Neither server.py nor vfi_model.py contains a lookup table, hashing of inputs, network calls or stored evaluation frames. The weights are the SEA-RAFT flow net plus the trained refiner. Public validation examples were used only for checkpoint selection and averaging choices, which the policy allows.
✓training data provenanceEvery corpus is declared in architecture.json and data_manifest.json, and prepare_data.py retains the download code: vim
Every corpus is declared in architecture.json and data_manifest.json, and prepare_data.py retains the download code: vimeo1080p train shards (train/ files only), Fanqi-Lin/GoPro-Raw-Videos (robot-manipulation footage such as fold_towel and pour_water, not SNU-FILM GoPro test footage), vimeo-90k-medium train shards, and the shipped public validation example. data_download.log records the counts. Searching the trajectory for SNU-FILM or test-split downloads found nothing.
✓harness credentials used only by the harnessThe shipped code reads only HF_TOKEN (prepare_data.py, for downloads) and MODEL_NAME (server.py). Searching the trajecto
The shipped code reads only HF_TOKEN (prepare_data.py, for downloads) and MODEL_NAME (server.py). Searching the trajectory for ANTHROPIC, OPENAI and BASE_URL found only the Spin CLI help text (trajectory#1-9 chunk). Compute and S3 access went through the Spin CLI, which is allowed.
LLM adversarial review is not proof of arbitrary program behavior. Host storage and task-specific deterministic checks remain required.
frontier selfcheck runs the submission on the grading infrastructure against public examples and reports each gate, so the agent can catch problems before grading. Each run gets four. This agent used 2 of 4.
- 1passed2h 14m into the runtook 3m 19s
startup 56.5 s of 300 s; 1/1 gates passed; The generation rate was projected from a sample: 72 of the evaluation's 270 intervals (600 of its 2250 frames), the first 9 counted once and the rest scaled to the evaluation, not timed over a whole pass.
5 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 56.5 s
- ✓POST /interpolate · the first public interval (4x, 384x720): 3 raw RGB frames of the requested size, accepted by the grader's parser
- ✓generation rate · 600 frames in 53.2 s, projected from 72 of 270 requests at concurrency 1: the first 9 intervals, which bring each of the sample's 9 (size, factor) shapes for the first time, took 8.8 s (the cold first one 0.8 s) and count once; the other 63 (525 frames) took 44.4 s, scaled to grading's other 2175 frames: 193 s of the 2250 s allowed for 2250 frames, including transport and export. 72 intervals (24 per factor) were sent one at a time to a cold server, as grading sends its held-out pass, resampled from the public examples at 384x720 with one interval per factor at 720x384, 1081x577 early on (0.29 megapixels per generated frame on average). Request seconds by factor: 4x 11.5, 8x 15.2, 16x 25.4; per generated frame by size: 384x720 (0.28 MP) 0.083 s, 720x384 (0.28 MP) 0.087 s, 1081x577 (0.62 MP) 0.156 s. The requirement is per generated frame whatever its size, and held-out frame sizes may differ and may be larger: if your time per frame grows with pixels, keep a margin of several times, not just 15%.
- ✓healthy after requests
- 2passed3h 53m into the runtook 2m 59s
startup 55.6 s of 300 s; 1/1 gates passed; The generation rate was projected from a sample: 72 of the evaluation's 270 intervals (600 of its 2250 frames), the first 9 counted once and the rest scaled to the evaluation, not timed over a whole pass.
5 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 55.6 s
- ✓POST /interpolate · the first public interval (4x, 384x720): 3 raw RGB frames of the requested size, accepted by the grader's parser
- ✓generation rate · 600 frames in 46.4 s, projected from 72 of 270 requests at concurrency 1: the first 9 intervals, which bring each of the sample's 9 (size, factor) shapes for the first time, took 7.6 s (the cold first one 0.7 s) and count once; the other 63 (525 frames) took 38.7 s, scaled to grading's other 2175 frames: 168 s of the 2250 s allowed for 2250 frames, including transport and export. 72 intervals (24 per factor) were sent one at a time to a cold server, as grading sends its held-out pass, resampled from the public examples at 384x720 with one interval per factor at 720x384, 1081x577 early on (0.29 megapixels per generated frame on average). Request seconds by factor: 4x 9.5, 8x 12.8, 16x 23.0; per generated frame by size: 384x720 (0.28 MP) 0.073 s, 720x384 (0.28 MP) 0.075 s, 1081x577 (0.62 MP) 0.137 s. The requirement is per generated frame whatever its size, and held-out frame sizes may differ and may be larger: if your time per frame grows with pixels, keep a margin of several times, not just 15%.
- ✓healthy after requests