Video frame interpolation
- Score
- 2.3
- Wallet spend
- $6.98 of $10
- Model cost
- $0.78
- Agent time
- 1h 20m
- Steps
- 159
- Tokens in / out
- 9.4M / 56k
- Self-checks
- 1 passed of 1 used
- Rank at $10 wallet
- 15th of 19
Score
2.3raw metric 0.02256
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
- 324.1637
- scoring_seconds
- 447.3165
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- 89278d823bd03dd0fd25d9e765518855a3f6d420612f7af2d8690d7b749a67a5
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 initializationserver.py loads torchvision raft_large(weights=None) and fills it from weights/raft_large_C_T_SKHT_V2-ff5fadd5.pth. That
server.py loads torchvision raft_large(weights=None) and fills it from weights/raft_large_C_T_SKHT_V2-ff5fadd5.pth. That file was fetched from official download.pytorch.org (trajectory#27) and its inventory shows the standard torchvision RAFT tensors (feature_encoder.*), so it is a standalone flow model, which is allowed. The synthesis UNet (weights/interp_unet_ft.pt) has a 19-channel input conv (model.b0.net.0.weight [32,19,3,3]). It was trained from random init in train.py (trajectory#83) and fine-tuned from that checkpoint in finetune.py (trajectory#93). The only frozen network used in the loss is LPIPS-alex. No RIFE, FILM, GIMM-VFI or other VFI weights were used as init, teacher or pseudo-labeler; architecture.json lists teachers: [] and fallback_models: [].
✓learned frame generationserver.py builds each intermediate frame with the trained UNet, a residual synthesis over the warped frames, RAFT flows,
server.py builds each intermediate frame with the trained UNet, a residual synthesis over the warped frames, RAFT flows, occlusion maps and the time map. The UNet weights come from the training and fine-tuning runs recorded in logs/train.log and logs/ft.log. Output frames are not copied or blended from fixed templates; they are produced by the learned model.
✓independent intervalsserver.py's /interpolate handler computes flows only from the request's own endpoint frames and generates every intermed
server.py's /interpolate handler computes flows only from the request's own endpoint frames and generates every intermediate t=idx/(T-1) from them. Nothing is cached or stored between requests, and no other interval is used.
✓no evaluation frame lookupserver.py and serve.sh contain no hash or fingerprint lookup, no stored evaluation frames and no network calls at serve
server.py and serve.sh contain no hash or fingerprint lookup, no stored evaluation frames and no network calls at serve time. serve.sh installs only the local wheels with --no-index, and the outputs are computed by RAFT plus the UNet.
✓training data provenancedata_manifest.json declares UCF101 frames from huggingface mavi88/ucf101_frames. The download of shard_00 is retained at
data_manifest.json declares UCF101 frames from huggingface mavi88/ucf101_frames. The download of shard_00 is retained at trajectory#80-82, and train/fine-tune use triplets from that shard (trajectory#83, #93). The declared 'train' split label is unverified because the HF repo has no split structure, but this is an action-recognition corpus, not a VFI validation or test split such as SNU-FILM or Vimeo-test. Model selection used only the task's public example (trajectory#96, #108), which is allowed. Vimeo test repos were only listed (trajectory#99) and never downloaded or used.
✓harness credentials used only by the harnessThe trajectory shows only HF_TOKEN used for HF downloads (allowed) and the harness's own selfcheck/submit tools. There w
The trajectory shows only HF_TOKEN used for HF downloads (allowed) and the harness's own selfcheck/submit tools. There was no use of ANTHROPIC, OPENAI or OPENROUTER keys or base URLs, and no claude or codex CLI was launched. server.py reads no such 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 1 of 4.
- 1passed1h 8m into the runtook 7m 50s
startup 52.5 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 52.5 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 120.0 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 19.6 s (the cold first one 2.2 s) and count once; the other 63 (525 frames) took 100.4 s, scaled to grading's other 2175 frames: 435 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 31.6, 8x 36.0, 16x 51.1; per generated frame by size: 384x720 (0.28 MP) 0.191 s, 720x384 (0.28 MP) 0.190 s, 1081x577 (0.62 MP) 0.344 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. The system is ready for grading.