Video frame interpolation
- Score
- 0.0
- Wallet spend
- $27.05 of $50
- Model cost
- $1.33
- Agent time
- 7h 43m
- Steps
- 599
- Tokens in / out
- 59.5M / 486k
- Self-checks
- 4 passed of 4 used
- Rank at $50 wallet
- 15th of 19
TimeoutError: Sequence request exceeded its deadline
Score
0.0
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.
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 initializationThe served network is the model_v2.VFIModel class (a bidirectional correlation-flow model with base=48). It is built fro
The served network is the model_v2.VFIModel class (a bidirectional correlation-flow model with base=48). It is built from plain nn modules and loads no pretrained weights. The training scripts (train2/3/4) create cls(base=args.base) with fresh weights. --resume only reloads the participant's own earlier checkpoints, and the run chain B→G→H is recorded in logs/commands.txt. The model.pt inventory has only model.* and raw.* keys that match model_v2 shapes at base 48. Two pretrained networks are used: LPIPS-AlexNet and a standalone PWC-Net. Both are declared in architecture.json and used only in the loss and metric, not in the served path. Trajectory searches found no download of RIFE, FILM or other VFI checkpoints (for example, no flownet.pkl).
✓learned frame generationserve.sh sets VFI_CKPT to model.pt and runs vfi/server.py. The server builds Interpolator(CKPT, mode='oneshot'), which l
serve.sh sets VFI_CKPT to model.pt and runs vfi/server.py. The server builds Interpolator(CKPT, mode='oneshot'), which loads state['model'] into VFIModel. The model estimates bidirectional flows and occlusion and runs a synthesis U-Net for each timestep t=k/n. The linear blend runs only if inference throws an exception and is not the normal path.
✓independent intervalsEach /interpolate request uses only its own inputs[0] and inputs[-1] endpoints. Apart from the cached engine, the server
Each /interpolate request uses only its own inputs[0] and inputs[-1] endpoints. Apart from the cached engine, the server keeps no state across requests, and warmup uses random tensors. Oneshot inference uses only that request's endpoint pair for every t.
✓no evaluation frame lookupThe serving code has no frame hashing, fingerprinting, stored-frame table or nearest-neighbor lookup. Outputs come only
The serving code has no frame hashing, fingerprinting, stored-frame table or nearest-neighbor lookup. Outputs come only from the network's forward pass on the request endpoints. The public examples and held-out REDS clips were used only for dev evaluation, according to data_manifest.json.
✓training data provenancedata_manifest.json declares two training sources. The first is the Vimeo-90K train split from HF bijinc/vimeo-90k-medium
data_manifest.json declares two training sources. The first is the Vimeo-90K train split from HF bijinc/vimeo-90k-medium at a pinned revision, limited to data/train-* files. The second is REDS train_sharp from HF snah/REDS, downloaded with a retained curl command. The trajectory shows only train-split downloads: the test parquet files appear in the step-23 repo listing but were never downloaded. No GoPro, SNU-FILM or other evaluation sets were used.
✓harness credentials used only by the harnessThe trajectory and shipped code show HF_TOKEN used only for HF dataset access and Spin used for compute. No ANTHROPIC, A
The trajectory and shipped code show HF_TOKEN used only for HF dataset access and Spin used for compute. No ANTHROPIC, API_KEY or BASE_URL credentials are read or used by the participant code or the serving path.
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.
- 1passed57m 14s into the runtook 6m 15s
startup 42.8 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 42.8 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 41.3 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 6.8 s (the cold first one 0.4 s) and count once; the other 63 (525 frames) took 34.4 s, scaled to grading's other 2175 frames: 150 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 7.0, 8x 11.0, 16x 22.4; per generated frame by size: 384x720 (0.28 MP) 0.064 s, 720x384 (0.28 MP) 0.066 s, 1081x577 (0.62 MP) 0.133 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 5m into the runtook 2m 20s
startup 39.8 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 39.8 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 46.4 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 7.4 s (the cold first one 0.4 s) and count once; the other 63 (525 frames) took 39.0 s, scaled to grading's other 2175 frames: 169 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 8.3, 8x 12.6, 16x 24.2; per generated frame by size: 384x720 (0.28 MP) 0.072 s, 720x384 (0.28 MP) 0.071 s, 1081x577 (0.62 MP) 0.142 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
- 3passed5h 25m into the runtook 3m 38s
startup 43.7 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 43.7 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 9.1 s (the cold first one 0.7 s) and count once; the other 63 (525 frames) took 86.7 s, scaled to grading's other 2175 frames: 368 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 24.6, 8x 30.2, 16x 39.9; per generated frame by size: 384x720 (0.28 MP) 0.161 s, 720x384 (0.28 MP) 0.097 s, 1081x577 (0.62 MP) 0.156 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
- 4passed7h 38m into the runtook 4m 30s
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 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 6.6 s (the cold first one 0.4 s) and count once; the other 63 (525 frames) took 31.3 s, scaled to grading's other 2175 frames: 136 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 7.1, 8x 10.5, 16x 19.3; per generated frame by size: 384x720 (0.28 MP) 0.059 s, 720x384 (0.28 MP) 0.060 s, 1081x577 (0.62 MP) 0.115 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.
Shipped system: a from-scratch learned VFI network (3.77M params, base-48) — coarse-to-fine bidirectional correlation-flow pyramid (1/8 → 1/4 → 1/2) → per-timestep flow refinement → learned occlusion mask over warped endpoints → full-resolution residual synthesis U-Net. One-shot timestep inference (every target frame synthesized directly from the two endpoints at t = k/(T−1), flows estimated once per interval), no TTA, no ensemble, single checkpoint.
Weights: run B's step-110k EMA → 7k-step fine-tune with the metric-aligned flow-difference-weighted LPIPS loss (run G) → 3.5k-step low-LR continuation (run H, step 120,500). All weights trace to random init; only LPIPS-AlexNet v0.1 (training loss) and PWC-Net (dev metric + loss weighting) are pretrained, neither in the served path.
Final validation: the last self-check passed on the grading image with wide margins — POST /interpolate 0.372 s, 16.5 frames/s projected (16× the required 1 fps; 136 s of the 2250 s budget for 2250 frames), healthy after requests, model loaded in 0.9 s on the H100. Local interface validation also passed.
Improvements tested and rejected on dev metrics (all documented in logs/commands.txt): longer/REDS-only fine-tunes (better on hard clips, worse on public-like content), a temporal flow-matching loss that measurably lowered FloLPIPS but slightly raised LPIPS, recursive inference, flip/scale TTA, unsharp-mask sharpening, 50/50 blending with linear interpolation, PWC flow substitution, output-space checkpoint ensembles and weight-space soups — none beat the shipped checkpoint. Diagnostic work also showed the FloLPIPS components sit ~3× above the task's baseline even for the best checkpoint (the model's residual error concentrates in moving regions), so those three components are unreachable with this model; the reachable score sits in the LPIPS components, where the shipped model is roughly 50–65% of the way from baseline to target on the public examples.