Video frame interpolation
- Score
- 57.2
- Wallet spend
- $30.84 of $50
- Model cost
- $1.76
- Agent time
- 9h 20m
- Steps
- 689
- Tokens in / out
- 68.6M / 532k
- Self-checks
- 4 passed of 4 used
- Rank at $50 wallet
- 7th of 19
Score
57.2raw metric 0.5724
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
- 361.6826
- scoring_seconds
- 448.8025
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- 37aac99de0bcab57311812a26c76a597d2239c548c509d8da2f93f062e47ff03
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 initializationStage 1 of code/vfi/train.py trains the interpolator (a FlowRefiner plus a SynthNet U-Net with FiLM time conditioning, f
Stage 1 of code/vfi/train.py trains the interpolator (a FlowRefiner plus a SynthNet U-Net with FiLM time conditioning, from code/vfi/model.py) from scratch. Stage 2 is initialized only from the stage-1 EMA. The tensor inventory of weights/interpolator.pt has only the custom module names (ema.flow.stem, ema.flow.enc1.film and so on), which fits a custom architecture rather than a pretrained VFI checkpoint. The only third-party learned component is torchvision's generic RAFT optical-flow model. code/vfi/flow.py loads the shipped C_T_SKHT_V2 weights into it and keeps it frozen; it is not a frame-interpolation specialist. I found no download or use of RIFE, FILM, GIMM or other VFI checkpoints in the trajectory. Those names appear only in the task text and in the declarations.
✓learned frame generationserve.sh launches code/server/server.py, which loads weights/interpolator.pt and the RAFT weights. Each target frame is
serve.sh launches code/server/server.py, which loads weights/interpolator.pt and the RAFT weights. Each target frame is produced by Interpolator.forward: RAFT bidirectional flow, then learned flow refinement, backward warping, occlusion masks, and a learned synthesis residual on the blended base, with t = target_index/(frame_count-1). The optional ensemble branch runs only if weights/interpolator_b.pt exists, and that file is not shipped. The output is generated by the learned model, not by copying or lookup.
✓independent intervalsThe /interpolate handler in code/server/server.py decodes exactly the two endpoint frames of the request and computes ev
The /interpolate handler in code/server/server.py decodes exactly the two endpoint frames of the request and computes every target from them alone. The only state kept across requests is throughput (fps) statistics used for TTA gating, which defaults to off in serve.sh. No image content or results are cached or shared between intervals.
✓no evaluation frame lookupNeither the server nor the model code contains frame fingerprinting, hashing, stored reference frames or dataset lookup.
Neither the server nor the model code contains frame fingerprinting, hashing, stored reference frames or dataset lookup. The shipped weights hold only model tensors. According to architecture.json and logs/eval_log.json, the three public task example intervals (media/sequence-0027) were used only for motion statistics and smoke tests, and none of their frames is stored or served.
✓training data provenancearchitecture.json and logs/data_manifest.json declare the sources as REDS train_sharp, the DAVIS 2017 train ImageSet, Go
architecture.json and logs/data_manifest.json declare the sources as REDS train_sharp, the DAVIS 2017 train ImageSet, GoPro train/sharp, Vimeo-90k medium train shards (downloaded, declared unused) and Vimeo-1080p train shards. Retained download commands are in code/download_data.sh and code/download_parallel.sh, with logs such as logs/download_vimeo90k.log. code/vfi/prepare_data.py filters each source to its train split only. Validation used a holdout taken from the training manifest (code/vfi/eval_holdout.py). In the trajectory, the HF repo file listings (step 23) only enumerated val/test files; nothing shows them being downloaded for or used in training or selection. I found no SNU-FILM use.
✓harness credentials used only by the harnessTrajectory searches for ANTHROPIC, OPENROUTER, BASE_URL, API_KEY, printenv and 'claude -p' found no participant use or c
Trajectory searches for ANTHROPIC, OPENROUTER, BASE_URL, API_KEY, printenv and 'claude -p' found no participant use or copying of harness credentials and no extra harness CLI launches. The only token used was HF_TOKEN, for Hugging Face dataset API reads and downloads (trajectory#23, trajectory#24), which is permitted.
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 4 of 4.
- 1passed1h 34m into the runtook 2m 52s
startup 46.4 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 46.4 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 48.8 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 7.7 s (the cold first one 0.6 s) and count once; the other 63 (525 frames) took 41.1 s, scaled to grading's other 2175 frames: 178 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 11.4, 8x 14.0, 16x 22.2; per generated frame by size: 384x720 (0.28 MP) 0.077 s, 720x384 (0.28 MP) 0.079 s, 1081x577 (0.62 MP) 0.135 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
- 2error3h 11m into the runtook 1m 42sfree (stopped before any GPU time)
The self-check could not finish (NotFoundError: Modal Sandbox with container ID ta-01M5DR2NRWBB58RHKBA9J5FP1S not found. This means this Sandbox has already shut down. (Error code: PF2J9NT2)).
1 gates
- ✓review prerequisites
- 3passed3h 14m into the runtook 2m 44s
startup 46.3 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 46.3 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 55.4 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 8.8 s (the cold first one 0.6 s) and count once; the other 63 (525 frames) took 46.7 s, scaled to grading's other 2175 frames: 202 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 12.6, 8x 15.8, 16x 25.6; per generated frame by size: 384x720 (0.28 MP) 0.087 s, 720x384 (0.28 MP) 0.088 s, 1081x577 (0.62 MP) 0.160 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
- 4passed8h 32m into the runtook 2m 37s
startup 53.8 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 53.8 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 37.9 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 6.1 s (the cold first one 0.4 s) and count once; the other 63 (525 frames) took 31.8 s, scaled to grading's other 2175 frames: 138 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 8.1, 8x 10.6, 16x 18.1; per generated frame by size: 384x720 (0.28 MP) 0.059 s, 720x384 (0.28 MP) 0.060 s, 1081x577 (0.62 MP) 0.114 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
- 5passed9h 13m into the runtook 2m 52s
startup 52.2 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 52.2 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 54.7 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 7.2 s (the cold first one 0.5 s) and count once; the other 63 (525 frames) took 47.5 s, scaled to grading's other 2175 frames: 204 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 11.2, 8x 13.9, 16x 28.3; per generated frame by size: 384x720 (0.28 MP) 0.088 s, 720x384 (0.28 MP) 0.072 s, 1081x577 (0.62 MP) 0.129 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.
The submission is final.
What was shipped (/workspace/output, 68 files, 4.01 GB):
- Model: a single interpolator trained from scratch (
code/vfi/model.py, 10.4 M params) — a time-conditionedFlowRefinerthat turns the frozen RAFT endpoint flow into per-timestep motion fields, plus a FiLM-conditioned U-NetSynthNetthat fuses endpoints, warped frames and an occlusion mask. The only pretrained components are RAFT (raft_large, frozen, generic optical flow) and LPIPS-AlexNet v0.1 (training loss only). No specialist VFI weights anywhere. - Weights: the EMA of a two-stage training run (76 000 steps, then 12 218 steps of an LPIPS-aligned fine-tune with the pixel term down-weighted, since PSNR is diagnostic only). sha256
a1a9ea06…, 41 657 091 B. - Serving:
serve.shinstalls the vendored cp312/cu128 wheels into/tmp, then starts the server; one H100, single model, no TTA, no ensemble.
Validation: the final frontier selfcheck on the shipped folder passed — startup 52.2 s / 300 s, POST /interpolate smoke test OK, generation rate 11.0 frames/s against the 1.0 frame/s gate, server healthy afterwards. A separate interface test suite on a GPU machine covered the raw-RGB8 byte count, the JSON fallback, empty and single-target requests, and non-multiple-of-16 and 1×5 frames — the last two exposed two real bugs (RAFT's correlation pyramid needs ≥128×128 inputs, and an empty target list crashed the batcher), both fixed and re-tested.
Measured on the public example (the only eval-distribution-matched sample): mean LPIPS 0.03736 (4×/8×/16×: 0.02205/0.03269/0.05734) versus 0.03760 for the end of the main run.
Rejected after measurement (all recorded in logs/eval_log.json): a third LPIPS-aligned training stage, a 50/50 weight soup of the two stages, mirror TTA, model ensembles, and float32 inference.