Video frame interpolation
- Score
- 66.9
- Wallet spend
- $0.90 of $10
- Model cost
- $15.59
- Agent time
- 2h 37m
- Steps
- 376
- Tokens in / out
- — / —
- Self-checks
- 2 passed of 2 used
- Rank at $10 wallet
- 7th of 19
Score
66.9raw metric 0.6686
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
- 298.4353
- scoring_seconds
- 412.6024
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- d7836e4afe852c7250240043295ed24b6b6fb5dbc44a280bee50ba243e5b68b1
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 initializationIn vfi_model.py the model loads only standalone optical-flow weights. These are torchvision RAFT-Large (raft_large_C_T_S
In vfi_model.py the model loads only standalone optical-flow weights. These are torchvision RAFT-Large (raft_large_C_T_SKHT_V2-ff5fadd5.pth), SEA-RAFT Spring-M (sea_raft_spring_M.safetensors) and UniMatch/GMFlow mixdata (gmflow_mixdata.pth). All three are declared in architecture.json with their official sources, and the trajectory shows them fetched by curl from those URLs. The tensor inventory of weights/vfi_synth.safetensors has a custom layout (ctx_encoder, z_net, enc1, down1/2, up2/1, 14-channel head) that matches VFISynthNet in vfi_model.py. train_vfi.py trains it from zero-initialized heads and optimizes only synth_net. No RIFE, FILM, GIMM or other specialist VFI weights were found anywhere in the code, the weights or the trajectory.
✓learned frame generationserver.py sends each /interpolate request to HybridFlowSynthVFI.generate_interval. That function estimates flows with th
server.py sends each /interpolate request to HybridFlowSynthVFI.generate_interval. That function estimates flows with the learned flow networks, applies softmax splatting weighted by a learned z-metric, then refines and blends the result with the trained VFISynthNet. serve.sh only installs vendored wheels and runs server.py. Intermediate frames are produced by learned networks, not by copying or simple fixed blending.
✓independent intervalsserver.py decodes the two endpoint PNGs for each request and calls generate_interval under a lock on a single global mod
server.py decodes the two endpoint PNGs for each request and calls generate_interval under a lock on a single global model. generate_interval computes everything from those two endpoints and the target indices alone. It keeps no image state between requests; warmup uses only dummy zero/one tensors.
✓no evaluation frame lookupInference in vfi_model.py and server.py does no hashing, caching, retrieval or dataset lookup. Outputs are computed from
Inference in vfi_model.py and server.py does no hashing, caching, retrieval or dataset lookup. Outputs are computed from the request's endpoint frames only. No evaluation or ground-truth frames are shipped or read when serving.
✓training data provenanceTraining data is declared in architecture.json and prepare_train_data.py. It consists of the danjacobellis/vimeo2k train
Training data is declared in architecture.json and prepare_train_data.py. It consists of the danjacobellis/vimeo2k train split (pinned revision, 15 videos with sha256 hashes, 171 clips, recorded in logs/data_provenance.json) plus the provided /task/public validation samples, which the task prompt allows to be used freely. train_vfi.py reads only those sources. A DAVIS train download was killed (trajectory#235). A Sintel training-split download was also killed (trajectory#242), and it is not referenced in the shipped training code. No SNU-FILM or held-out test data was used.
✓harness credentials used only by the harnessThe trajectory shows no use of harness API keys or base URLs by the participant's code. These strings appear only in too
The trajectory shows no use of harness API keys or base URLs by the participant's code. These strings appear only in tool help text. Model and data downloads used public URLs and the HF hub. The shipped serve.sh, server.py and vfi_model.py make no remote calls.
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 57m into the runtook 6m 57s
startup 91.6 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 91.6 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 259.8 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 194.8 s (the cold first one 62.0 s) and count once; the other 63 (525 frames) took 65.0 s, scaled to grading's other 2175 frames: 464 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 200.6, 8x 22.6, 16x 35.4; per generated frame by size: 384x720 (0.28 MP) 0.233 s, 720x384 (0.28 MP) 2.659 s, 1081x577 (0.62 MP) 2.548 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
- 2passed2h 30m into the runtook 4m 37s
startup 59.5 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 59.5 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 148.7 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 16.6 s (the cold first one 1.3 s) and count once; the other 63 (525 frames) took 132.1 s, scaled to grading's other 2175 frames: 564 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 38.3, 8x 45.7, 16x 63.4; per generated frame by size: 384x720 (0.28 MP) 0.247 s, 720x384 (0.28 MP) 0.176 s, 1081x577 (0.62 MP) 0.292 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