Video frame interpolation
- Score
- 56.5
- Wallet spend
- $18.08 of $200
- Model cost
- $20.79
- Agent time
- 3h 17m
- Steps
- 411
- Tokens in / out
- — / —
- Self-checks
- 2 passed of 2 used
- Rank at $200 wallet
- 10th of 19
Score
56.5raw metric 0.5652
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
- 354.5353
- scoring_seconds
- 521.0543
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- d2549a30d6a6b7d3f033b0964b08bc4bd2dee85a161458a50723f7620fb827e2
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 are only three shipped weight files: - torchvision RAFT-Large C_T_SKHT_V2. The trajectory shows it downloaded from
There are only three shipped weight files: - torchvision RAFT-Large C_T_SKHT_V2. The trajectory shows it downloaded from download.pytorch.org and copied into the output. - SEA-RAFT spring-M, a standalone optical-flow model fetched from the MemorySlices HF repo. - vfi_refiner.pt, whose keys match the custom FlowSynthRefiner. Round 1 deleted vfi_refiner.pt before training from scratch (trajectory#229-249). Rounds 2 and 3 fine-tuned that same checkpoint. train_vfi.py uses only the flow nets plus LPIPS/PWC-Net from the evaluator as losses. No RIFE, FILM or GIMM-VFI weights, teacher, distillation or pseudo-labels appear in the code or the session. Hashes match architecture.json.
✓learned frame generationserver.py's /interpolate handler calls MODEL.interpolate_interval in vfi_model.py. That path loads the shipped RAFT/SEA-
server.py's /interpolate handler calls MODEL.interpolate_interval in vfi_model.py. That path loads the shipped RAFT/SEA-RAFT weights and the trained FlowSynthRefiner. It computes learned bidirectional flow, splats the flow to time t, then runs the learned refiner (occlusion mask plus RGB residual, with flip TTA) to generate the frames. Generation is therefore done by learned models, not by copying frames.
✓independent intervalsserver.py keeps no state across requests apart from the global model and a dummy warmup. Each request uses only its own
server.py keeps no state across requests apart from the global model and a dummy warmup. Each request uses only its own two endpoint PNGs and the frame index or time. The recursive bisection blend inside interpolate_interval uses frames predicted within that same request. There is no cross-request cache or shared buffer.
✓no evaluation frame lookupThe serving code in vfi_model.py and server.py makes no network or file lookups of target frames and contains no hash ta
The serving code in vfi_model.py and server.py makes no network or file lookups of target frames and contains no hash tables or stored evaluation outputs. Output is computed from the endpoints alone. Evaluation and selection used the provided /task/public validation samples, which the task explicitly lets participants use freely, plus held-out train-split sequences.
✓training data provenanceAll corpora are declared in architecture.json and the manifests: - **snah/GOPRO_Large train/ only.** prepare_gopro_train
All corpora are declared in architecture.json and the manifests: - **snah/GOPRO_Large train/ only.** prepare_gopro_train.py filters train/, and the manifest lists 22 official train sequences. The trajectory confirms test/ was never extracted (trajectory#186-213). - **Fanqi-Lin/GoPro-Raw-Videos and flying101/cinematic-slowmotion.** prepare_extra_train.py, the extra manifest and the log list the exact source files. - **The provided public sequence-0027.** SNU_FILM and vimeo appeared only in an HF dataset search listing (trajectory#180) and were never downloaded or used.
✓harness credentials used only by the harnessThe only credential the trajectory shows in use is the read-only HF_TOKEN, used to download declared public HF datasets
The only credential the trajectory shows in use is the read-only HF_TOKEN, used to download declared public HF datasets and weights. That is its stated purpose. There are no uses of ANTHROPIC, OPENAI or BASE_URL credentials, no model-API calls, and no harness credential in shipped code. serve.sh and server.py contain none.
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.
- 1passed1h 51m into the runtook 2m 50s
startup 52.1 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 52.1 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 58.7 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 15.8 s (the cold first one 1.5 s) and count once; the other 63 (525 frames) took 42.9 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 14.9, 8x 16.3, 16x 26.3; per generated frame by size: 384x720 (0.28 MP) 0.084 s, 720x384 (0.28 MP) 0.157 s, 1081x577 (0.62 MP) 0.293 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 12m into the runtook 2m 58s
startup 47.4 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 47.4 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 68.3 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 17.9 s (the cold first one 1.3 s) and count once; the other 63 (525 frames) took 50.4 s, scaled to grading's other 2175 frames: 227 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 13.5, 8x 20.6, 16x 33.1; per generated frame by size: 384x720 (0.28 MP) 0.099 s, 720x384 (0.28 MP) 0.178 s, 1081x577 (0.62 MP) 0.337 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