Video frame interpolation
- Score
- 50.5
- Wallet spend
- $6.18 of $10
- Model cost
- $19.46
- Agent time
- 3h 32m
- Steps
- 581
- Tokens in / out
- 62.5M / 343k
- Self-checks
- 3 passed of 3 used
- Rank at $10 wallet
- 10th of 19
Score
50.5raw metric 0.5046
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
- 349.3242
- scoring_seconds
- 440.6677
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- cbdd550564a401963f9d431e86fdd2114e15db049d9f2a2445194bc01ab66f06
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 initializationTwo models are served. The first is a frozen standalone RAFT Sintel optical-flow checkpoint. It was downloaded from the
Two models are served. The first is a frozen standalone RAFT Sintel optical-flow checkpoint. It was downloaded from the declared ddrfan/RAFT Hugging Face URL at a pinned revision with sha256 (trajectory#24), and its tensor inventory is RAFT-shaped (module.fnet.*). The second is a SynthesisNet trained from scratch. train.py builds it with PyTorch default random initialization and a zero-initialized head; resume only reloads its own earlier checkpoints. The main3s24.pt inventory (enc0.conv.weight [32,24,3,3]) matches SynthesisNet. PWC-Net and LPIPS-Alex are used only inside training losses. PWC-Net is a standalone flow model, which the policy allows. No RIFE, FILM, GIMM-VFI or other specialist VFI checkpoint was found being downloaded, loaded, used as a teacher or used to make pseudo-labels; the trajectory mentions these names only when restating the restrictions.
✓learned frame generationserve.sh runs server.py. The Interpolator loads each member listed in inference_config.json with torch.load(weights_only
serve.sh runs server.py. The Interpolator loads each member listed in inference_config.json with torch.load(weights_only=True) and strict state_dict loading. Each request computes RAFT bidirectional flow, then runs the learned SynthesisNet ensemble (flow correction, fusion logits and an RGB residual) to generate the target frames. linear_fallback runs only when an exception occurs, so the served output comes from the learned models.
✓independent intervalsserver.py's /interpolate handler decodes only the two endpoint PNGs in each request and calls predict(im0, im1, target_i
server.py's /interpolate handler decodes only the two endpoint PNGs in each request and calls predict(im0, im1, target_indices, factor). It keeps no state across requests and does not cache or use other intervals.
✓no evaluation frame lookupThe serving path only computes from the two supplied endpoint frames using RAFT and SynthesisNet. It contains no dataset
The serving path only computes from the two supplied endpoint frames using RAFT and SynthesisNet. It contains no dataset index, hash lookup or stored evaluation frames, and it does not reach the network for ground-truth frames.
✓training data provenancedata_manifest.json and architecture.json declare the training data as the Vimeo6K (danjacobellis/vimeo6k) train-split sh
data_manifest.json and architecture.json declare the training data as the Vimeo6K (danjacobellis/vimeo6k) train-split shards at a pinned revision. training/setup_data.sh and preprocess_videos.py turn these into /workspace/vfi/data/clips17, and train.py, the fine-tuning scripts and every config read only clips17. Model selection used an internal holdout carved from the train split plus the task's public validation examples. The trajectory shows that a Vimeo90K triplet train shard was looked at and an attempted septuplet download was blocked (trajectory#22, #38). Neither feeds the 17-frame clips17 pipeline, and both are training splits in any case. No public val/test split (Vimeo90K test or SNU-FILM) was used.
✓harness credentials used only by the harnessThe only appearance of the harness credentials is incidental, in ps output showing the harness command line (trajectory#
The only appearance of the harness credentials is incidental, in ps output showing the harness command line (trajectory#464). The participant used the Spin compute CLI and its SSH key (runpod.key), which are allowed. training/audit_submission.py only scans files for credential strings and does not use them. No reading, copying or use of ANTHROPIC_API_KEY, OPENROUTER or *_BASE_URL was found.
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.
- 1passed46m 11s into the runtook 2m 28s
startup 41.4 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 41.4 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 40.8 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 8.3 s (the cold first one 0.8 s) and count once; the other 63 (525 frames) took 32.5 s, scaled to grading's other 2175 frames: 143 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 9.9, 8x 11.6, 16x 18.1; per generated frame by size: 384x720 (0.28 MP) 0.062 s, 720x384 (0.28 MP) 0.085 s, 1081x577 (0.62 MP) 0.139 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
- 2failed3h 22m into the runtook 27sfree (stopped before any GPU time)
failed review prerequisites: Initialization metadata must declare a nonempty base_model and base_revision (or none for both when training from scratch)
1 gates
- ✗review prerequisites · Initialization metadata must declare a nonempty base_model and base_revision (or none for both when training from scratch)
- 3passed3h 23m into the runtook 2m 42s
startup 49.2 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 49.2 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 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 11.0 s (the cold first one 0.8 s) and count once; the other 63 (525 frames) took 37.8 s, scaled to grading's other 2175 frames: 168 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 13.3, 8x 13.4, 16x 21.3; per generated frame by size: 384x720 (0.28 MP) 0.073 s, 720x384 (0.28 MP) 0.125 s, 1081x577 (0.62 MP) 0.185 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
- 4passed3h 27m into the runtook 4m 38s
startup 53.7 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 53.7 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 39.6 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 10.1 s (the cold first one 0.8 s) and count once; the other 63 (525 frames) took 29.5 s, scaled to grading's other 2175 frames: 132 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 9.9, 8x 9.9, 16x 18.9; per generated frame by size: 384x720 (0.28 MP) 0.057 s, 720x384 (0.28 MP) 0.108 s, 1081x577 (0.62 MP) 0.179 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.
Final unchanged-package selfcheck B passed:
- Startup: 53.7 s of 300 s.
- Projected throughput: 17.0043 generated frames/s.
- Sample: 600 frames in 39.6 s; projected full workload about 132 s of 2,250 s.
- Digest:
9c3b1941d047b7b717416a433faf676c436a697d88a639b5d1631ba2551ab64c. - 147 files and 4,058,024,677 bytes, matching the final audit exactly.
I’m saving B outside the package and proceeding to the requested two-step submission without further edits.