Video frame interpolation
- Score
- 25.7
- Wallet spend
- $83.76 of $200
- Model cost
- $1.89
- Agent time
- 4h 49m
- Steps
- 1,044
- Tokens in / out
- 143.5M / 536k
- Self-checks
- 2 passed of 2 used
- Rank at $200 wallet
- 11th of 14
Score
25.7raw metric 0.2566
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
- 353.8069
- scoring_seconds
- 789.5969
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- 35d15914d7b9843ff51d1470a95d5ee13eca847a95bd5372b3cc0bedbd5f0524
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 initializationThe only pretrained runtime weights are torchvision RAFT-Large C_T_V2, which is a standalone optical-flow model and is a
The only pretrained runtime weights are torchvision RAFT-Large C_T_V2, which is a standalone optical-flow model and is allowed. The session downloaded it from download.pytorch.org, and its sha256 (1bb1363a…) matches the shipped file. The tensor inventory has standard RAFT keys. FusionNet starts from random weights in training/train.py: run_train.sh trains 40k steps from scratch, and the continue/extend scripts only resume its own checkpoints. The fusion_*.pt inventories hold only FusionNet keys (enc1 [32,21,3,3], head [4,32,3,3]). No VFI checkpoint download or teacher was found. AlexNet (LPIPS) and PWC-Net were used only as diagnostic metrics.
✓learned frame generationserve.sh installs the vendored wheels offline and runs serve.py. serve.py loads RAFT from the local .pth and FusionNet f
serve.sh installs the vendored wheels offline and runs serve.py. serve.py loads RAFT from the local .pth and FusionNet from weights/fusion_{75k,10k,15k}.pt. For each target it runs bidirectional RAFT flow, warps both endpoints, builds the 21-channel fusion input, runs FusionNet and composes the frame as a sigmoid-gated blend plus a bounded tanh RGB correction (model.py). It can average this with a temporal-reversal prediction. Every output frame is produced by the learned network, and architecture.json lists no fallback.
✓independent intervalsThe /interpolate handler in serve.py uses only the two endpoint PNGs and target_indices from the current request. Nothin
The /interpolate handler in serve.py uses only the two endpoint PNGs and target_indices from the current request. Nothing persists between requests except the fixed loaded weights and a GPU lock. The checkpoint and setting chosen for each frame count are fixed in INFERENCE_SETTINGS and do not depend on other intervals.
✓no evaluation frame lookupserve.py has no hashing or fingerprinting of inputs, no stored frame database, no network calls and no retrieval logic.
serve.py has no hashing or fingerprinting of inputs, no stored frame database, no network calls and no retrieval logic. Outputs are computed only by the RAFT and FusionNet forward passes. Settings were selected using the task-shipped examples, not evaluation frames.
✓training data provenancetraining/data_provenance.json, download_data.sh, download.log, prepare_data.py and prepare.log show that training used o
training/data_provenance.json, download_data.sh, download.log, prepare_data.py and prepare.log show that training used only the 25 train shards of the Vimeo90K Triplet split from huggingface.co/datasets/bijinc/vimeo-90k-medium at a pinned revision (25656 rows). Shard hashes are retained in vimeo_train_sha256.txt. Supervision is the real Vimeo middle frame. The trajectory shows no use of test splits or evaluation data, and checkpoint/setting selection used only the task-shipped examples.
✓harness credentials used only by the harnessNo trajectory step reads, copies or uses the harness API key or base URL, and no extra agent CLI sessions were launched
No trajectory step reads, copies or uses the harness API key or base URL, and no extra agent CLI sessions were launched with them. training/backup_to_r2.py uploads checkpoints with Spin transfer-session credentials, which are not harness credentials.
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 40m into the runtook 4m 11s
startup 62.0 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 62.0 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 118.4 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 13.7 s (the cold first one 2.8 s) and count once; the other 63 (525 frames) took 104.7 s, scaled to grading's other 2175 frames: 447 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 32.7, 8x 34.9, 16x 49.5; per generated frame by size: 384x720 (0.28 MP) 0.198 s, 720x384 (0.28 MP) 0.124 s, 1081x577 (0.62 MP) 0.211 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
- 2passed4h 45m into the runtook 3m 32s
startup 51.0 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 51.0 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 71.9 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 12.4 s (the cold first one 2.5 s) and count once; the other 63 (525 frames) took 59.5 s, scaled to grading's other 2175 frames: 259 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 17.5, 8x 18.8, 16x 30.7; per generated frame by size: 384x720 (0.28 MP) 0.109 s, 720x384 (0.28 MP) 0.107 s, 1081x577 (0.62 MP) 0.186 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
The agent's last message before its session ended.
Submitted at 2026-10-09 11:58 UTC.
frontier selfcheck passed on an H100: startup was 51 seconds, and the 2,250-frame runtime projection was 259 seconds. The final routes use the 75k, 10k, and 15k FusionNet checkpoints, with factor-specific RAFT settings.
Training data, pretrained component origins, and the permitted sample-selection process are documented in architecture.json. The selfcheck does not report a quality score.