Video frame interpolation
- Score
- 64.2
- Wallet spend
- $41.88 of $200
- Model cost
- $1.32
- Agent time
- 9h 51m
- Steps
- 522
- Tokens in / out
- 50.4M / 477k
- Self-checks
- 3 passed of 3 used
- Rank at $200 wallet
- 8th of 19
Score
64.2raw metric 0.6423
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
- 368.2663
- scoring_seconds
- 493.4332
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- 35565fe1539a8d4fd55e301a08969738fd79969149d3433f603dc7f09a43453e
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 initializationServing loads two sets of weights. The first is a frozen torchvision RAFT-large general optical-flow model (C_T_SKHT_V2)
Serving loads two sets of weights. The first is a frozen torchvision RAFT-large general optical-flow model (C_T_SKHT_V2). Its sha256 ff5fadd5... matches across weights/CHECKSUMS.txt, data_manifest.json and the torch-hub file the trajectory shows was symlinked in, and its tensor inventory is the standard RAFT layout. The second is the task-trained Refiner U-Net, stored as ema.* tensors with a 24-channel input and an 8-channel output. architecture.json declares that the Refiner was randomly initialized. The trajectory shows the v1 run starting from scratch, with later runs resuming only from the participant's own checkpoints (/workspace/ckpt3/last.pt, /workspace/init.pt). The trajectory shows no RIFE, FILM or other VFI-specialist checkpoint being downloaded or loaded. PWC-Net and AlexNet/LPIPS appear only as training-loss components.
✓learned frame generationserve.sh launches serve.py. serve.py builds VFIModel from vfi_model.py using weights/refiner.pt and weights/raft_large.p
serve.sh launches serve.py. serve.py builds VFIModel from vfi_model.py using weights/refiner.pt and weights/raft_large.pth. Each output frame is produced by learned synthesis: RAFT estimates the flows, those are scaled by t, the inputs are warped with grid_sample, and the Refiner predicts the mask, an RGB residual and flow residuals. None of the generation paths is a fallback such as copying or blending a nearby frame. architecture.json lists no fallback models.
✓independent intervalsserve.py interpolate() uses only the two endpoint frames of the request, frames[0] and frames[count-1], with timesteps i
serve.py interpolate() uses only the two endpoint frames of the request, frames[0] and frames[count-1], with timesteps i/(count-1). The module-level STATE holds only the loaded model, and nothing is cached or carried between requests. The recursive mode recurses only within the current request.
✓no evaluation frame lookupserve.py makes no network calls and reads no stored frames, hashes or ground-truth files. Outputs are computed purely fr
serve.py makes no network calls and reads no stored frames, hashes or ground-truth files. Outputs are computed purely from the request's endpoint frames. The task's public examples appear only in train_code/train7.py validate(), where they are used for development model selection, as data_manifest.json declares. They are not shipped as data the server reads.
✓training data provenancedata_manifest.json declares the training corpora as Vimeo-90K train (HF bijinc/vimeo-90k-medium) and Inter4K (HF Shilin-
data_manifest.json declares the training corpora as Vimeo-90K train (HF bijinc/vimeo-90k-medium) and Inter4K (HF Shilin-LU/Inter4k), and train_code/train7.py reads those prepared directories. UCF-101 was downloaded during the session, but every training log shows ucf=0, so it was never used. Public task examples were used only for validation/selection, and the manifest declares this. The trajectory shows no use of benchmark or evaluation test sets such as SNU-FILM for training.
✓harness credentials used only by the harnessSearching the trajectory found no access to or use of harness or Anthropic credentials. The Spin compute calls the parti
Searching the trajectory found no access to or use of harness or Anthropic credentials. The Spin compute calls the participant made were for running their own jobs through the provided tooling. serve.py makes no outbound authenticated 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 3 of 4.
- 1passed28m 25s into the runtook 7m 44s
startup 57.3 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 57.3 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 199.2 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 37.0 s (the cold first one 4.4 s) and count once; the other 63 (525 frames) took 162.2 s, scaled to grading's other 2175 frames: 709 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 53.9, 8x 59.4, 16x 84.6; per generated frame by size: 384x720 (0.28 MP) 0.318 s, 720x384 (0.28 MP) 0.364 s, 1081x577 (0.62 MP) 0.557 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
- 2passed1h 44m into the runtook 3m 15s
startup 45.7 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 45.7 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 95.8 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 15.2 s (the cold first one 1.6 s) and count once; the other 63 (525 frames) took 80.7 s, scaled to grading's other 2175 frames: 349 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 20.2, 8x 25.5, 16x 48.9; per generated frame by size: 384x720 (0.28 MP) 0.153 s, 720x384 (0.28 MP) 0.150 s, 1081x577 (0.62 MP) 0.268 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
- 3passed9h 47m into the runtook 4m 4s
startup 49.0 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 49.0 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 96.5 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 15.9 s (the cold first one 1.8 s) and count once; the other 63 (525 frames) took 80.6 s, scaled to grading's other 2175 frames: 350 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 19.9, 8x 26.9, 16x 48.8; per generated frame by size: 384x720 (0.28 MP) 0.154 s, 720x384 (0.28 MP) 0.157 s, 1081x577 (0.62 MP) 0.277 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.
Submission complete.
Final state
Submitted: /workspace/output/ — 52 files, 3.09 GB, submitted 05:28:55 UTC after a passing selfcheck (startup 49.0 s of 300 s, 1/1 speed gates, 72/270 sampled intervals projected well inside the generation budget).
Model: frozen torchvision RAFT-large (C_T_SKHT_V2, standalone optical-flow weights) estimates bidirectional flows; a task-trained 7.8M-param U-Net refiner (24→8 ch: blend mask, RGB residual, two flow residuals) warps and blends the endpoints. Trained only on public corpora (Vimeo-90K triplet train split, Inter4K), no specialist VFI checkpoints anywhere, no cross-request state.
What happened this session:
- The armed post-training waiter never fired — its own launching shell contained the pgrep pattern, so it matched itself forever. Killed it and relaunched the fine-tune manually.
- The first rented H100's 8 h prepaid lease expired at 03:39 UTC (taking the v6 100k checkpoint and its snapshots with it); extension of an already-expired lease was refused.
- Rebuilt the training environment on a fresh H100: re-downloaded Inter4K (15.5 GB), re-prepared 84k frames, re-uploaded the only surviving checkpoint (the shipped v6 iter-6000 EMA, verified by md5) and fine-tuned it for 20k iterations with a gap-16-heavy sampling distribution.
- Scored candidates with the exact LPIPS-AlexNet metric on the three public intervals: v7 final EMA 0.4792 (fp32) / 0.4808 (val proxy) vs 0.4777 / 0.4738 for the previously shipped checkpoint. Weight soups (α = 0.5, 0.75) and output ensembling measured 0.4783–0.4794 — within noise, so no ensembling was shipped.
Measured but not shipped: recursive midpoint prediction was clearly worse (0.4715); flip-TTA was within noise (+0.0007) so the simpler, faster direct path was kept. FloLPIPS components stayed at the 0-clip in my replication for every candidate, so checkpoint selection rested on LPIPS; the 4x and 8x components are at the target cap (normalized 1.000), and 16x is ~0.885.
The pod was stopped (final charge $13.96) before submission.