Video frame interpolation
- Score
- 86.5
- Wallet spend
- $149.67 of $200
- Model cost
- $21.19
- Agent time
- 9h 56m
- Steps
- 600
- Tokens in / out
- 67.8M / 817k
- Self-checks
- 2 passed of 3 used
- Rank at $200 wallet
- 3rd of 14
Score
86.5raw metric 0.8654
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
- 417.9803
- scoring_seconds
- 669.7947
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- fb5206516fcfd485e3d9d315abc4b4900b2f59c20c4281d18b46b7649314942e
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 initializationNo specialist frame-interpolation weights are used. The pretrained parts are SEA-RAFT, a standalone optical-flow model (
No specialist frame-interpolation weights are used. The pretrained parts are SEA-RAFT, a standalone optical-flow model (MemorySlices Tartan-C-T-TSKH-spring540x960-M, fetched at a pinned sha), and the frozen general-image encoders Depth Anything V2 Small and Base. Their sources, revisions and sha256 hashes are recorded in pretrained_manifest.json. The own FrameRefiner models (refiner.pt and semantic_base.pt) were trained from random initialization during the session. The 'teacher' caches are SEA-RAFT flow computed on observed frame pairs; they are not outputs of a VFI model. PWC-Net, AlexNet/LPIPS and DPFlow appear only in development or training-loss code and are not served.
✓learned frame generationserve.sh runs server.py. server.py calls Interpolator.generate. interpolator.py strict-loads weights/refiner.pt and weig
serve.sh runs server.py. server.py calls Interpolator.generate. interpolator.py strict-loads weights/refiner.pt and weights/semantic_base.pt and runs EndpointFlow (SEA-RAFT from weights/sea_model.safetensors). hierarchy.predict then encodes semantics, splats and interpolates, and refines at each scale. So every output frame is produced by the learned models. The only non-model shortcut is that identical inputs are returned as copies, which is a trivial edge case.
✓independent intervalsEach POST request is decoded, sorted by index and handled with interpolator.generate(first, last, times, factor). The se
Each POST request is decoded, sorted by index and handled with interpolator.generate(first, last, times, factor). The server keeps no state across requests and has no cache. The pyramid recursion in hierarchy.py stays within a single request.
✓no evaluation frame lookupThe serving code uses only the two endpoint frames from the request. Neither interpolator.py, hierarchy.py nor server.py
The serving code uses only the two endpoint frames from the request. Neither interpolator.py, hierarchy.py nor server.py contains any lookup table, hash matching or stored evaluation frames. The weights hold only model tensors.
✓training data provenancearchitecture.json declares every training corpus: the task examples; the Vimeo90K train intersection from bijinc/vimeo-9
architecture.json declares every training corpus: the task examples; the Vimeo90K train intersection from bijinc/vimeo-90k-medium at a pinned revision, which vimeo_split_audit.json shows is entirely in the official train lists with no test IDs; DAVIS2017 train only, with validation excluded and the archive sha256 recorded in davis_native_manifest.json; and REDS train_sharp from snah/REDS at a pinned revision. The trajectory shows matching downloads. The GoPro and REDS validation/test files were only listed, never downloaded. The task prompt allows unrestricted use of the provided validation samples.
✓harness credentials used only by the harnessNo trajectory step reads, copies or uses harness or model-API credentials. Searches for ANTHROPIC, OPENAI, BASE_URL and
No trajectory step reads, copies or uses harness or model-API credentials. Searches for ANTHROPIC, OPENAI, BASE_URL and 'codex exec' found nothing. The compute-lease commands are ordinary sanctioned tooling.
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 3 of 4.
- 1failed8m 12s into the runtook 10m 15s
failed startup: Server did not become healthy in 300s (last /health probe: RuntimeError: Submission HTTP request failed: ConnectionRefusedError: [Errno 111] Connection refused); end of its log: Traceback (most recent call last): File "/submission/server.py", line 68, in <module> main()
2 gates
- ✓review prerequisites
- ✗startup · Server did not become healthy in 300s (last /health probe: RuntimeError: Submission HTTP request failed: ConnectionRefusedError: [Errno 111] Connection refused); end of its log: Traceback (most recent call last): File "/submission/server.py", line 68, in <module> main() File "/submission/server.py", line 15, in main interpolator = Interpolator() ^^^^^^^^^^^^^^ File "/submission/interpolator.py", line 13, in __init__ self.flow = EndpointFlow() ^^^^^^^^^^^^^^ File "/submission/flow.py", line 19, in __init__ self.model.load_state_dict(load_file(str(ROOT / "weights/sea_model.safetensors")), strict=True) File "/tmp/vfi-python/torch/nn/modules/module.py", line 2593, in load_state_dict raise RuntimeError( RuntimeError: Error(s) in loading state_dict for RAFT: Missing key(s) in state_dict: "cnet.layer2.0.downsample.1.weight", "cnet.layer2.0.downsample.1.bias", "cnet.layer2.0.downsample.1.running_mean", "cnet.layer2.0.downsample.1.running_var", "cnet.layer3.0.downsample.1.weight", "cnet.layer3.0.downsample.1.bias", "cnet.layer3.0.downsample.1.running_mean", "cnet.layer3.0.downsample.1.running_var", "fnet.layer2.0.downsample.1.weight", "fnet.layer2.0.downsample.1.bias", "fnet.layer2.0.downsample.1.running_mean", "fnet.layer2.0.downsample.1.running_var", "fnet.layer3.0.downsample.1.weight", "fnet.layer3.0.downsample.1.bias", "fnet.layer3.0.downsample.1.running_mean", "fnet.layer3.0.downsample.1.running_var".
- 2error42m 8s into the runtook 4m 59sfree (stopped before any GPU time)
The self-check could not finish (NotFoundError: Modal Sandbox with container ID ta-01M4751MV71YBMSP9SA2TA8HDS not found. This means this Sandbox has already shut down. (Error code: 7BWHE9MA)).
- 3passed1h 8m into the runtook 4m 1s
startup 70.3 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 70.3 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 102.3 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 18.1 s (the cold first one 3.3 s) and count once; the other 63 (525 frames) took 84.2 s, scaled to grading's other 2175 frames: 367 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 22.5, 8x 38.7, 16x 40.0; per generated frame by size: 384x720 (0.28 MP) 0.164 s, 720x384 (0.28 MP) 0.171 s, 1081x577 (0.62 MP) 0.261 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
- 4passed9h 45m into the runtook 6m 24s
startup 72.4 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 72.4 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 235.5 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 41.8 s (the cold first one 5.4 s) and count once; the other 63 (525 frames) took 193.7 s, scaled to grading's other 2175 frames: 844 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 54.4, 8x 80.3, 16x 99.6; per generated frame by size: 384x720 (0.28 MP) 0.379 s, 720x384 (0.28 MP) 0.428 s, 1081x577 (0.62 MP) 0.607 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 /workspace/output/ with frontier submit --confirm.
- Learned endpoint-only interpolation; all pretrained origins declared.
- No prohibited specialist VFI checkpoints or teachers used.
- Final offline self-check passed: 72.4s startup and 844s projected generation/export for 2,250 frames.
- Passed odd-sized, extreme-aspect-ratio, and 4096² interface tests.
Self-check reports runtime compliance, not an official quality score.