Video frame interpolation
- Score
- 44.1
- Wallet spend
- $5.03 of $10
- Model cost
- $2.52
- Agent time
- 10h 7m
- Steps
- 842
- Tokens in / out
- 87.1M / 844k
- Self-checks
- 3 passed of 3 used
- Rank at $10 wallet
- 11th of 19
Score
44.1raw metric 0.4411
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
- 354.9414
- scoring_seconds
- 505.1474
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- 31fe00c318642c20f3489a3c46f8df33e55f4cf5cf720413fa64301ec979793e
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.
The score comes from a regrade that re-scored this same submission without reviewing it again; this is the review from its original grading (2026-10-21).
✓non specialist initializationThe only pretrained component is the frozen generic optical-flow model, torchvision RAFT-large C_T_SKHT_V2. architecture
The only pretrained component is the frozen generic optical-flow model, torchvision RAFT-large C_T_SKHT_V2. architecture.json declares its URL and sha, and the weight inventory shows RAFT tensor names at the declared 21,106,607 bytes. vfi_model.build_raft builds raft_large(weights=None) and loads that file. The synthesis U-Net (unet.safetensors, unet.* tensors) starts from random initialization with a zero-initialized output layer and is trained from scratch. architecture.json lists teachers=[] and fallback_models=[]. No RIFE, IFRNet or other VFI-specialist checkpoint appears in the code, the weights or the trajectory. RIFE appears only in the task text and in a 'RIFE-style' comment.
✓learned frame generationserve.sh runs server.py. server.py loads unet.safetensors with a strict load into Interpolator(base=48). Each output fra
serve.sh runs server.py. server.py loads unet.safetensors with a strict load into Interpolator(base=48). Each output frame is produced by the learned network: RAFT flow plus learned flow residuals, warping, a learned softmax blend over candidates and a learned RGB residual. There is no hand-coded fallback such as a plain linear blend or frame duplication standing in for the model.
✓independent intervalsserver.py interpolate takes exactly the two frames in the current request. It computes endpoint_flow and occlusion masks
server.py interpolate takes exactly the two frames in the current request. It computes endpoint_flow and occlusion masks from those two frames only and evaluates the model at each t=idx/factor. Nothing is cached or kept as state across requests, and the handler rejects requests that do not have exactly two input frames.
✓no evaluation frame lookupThe serving path makes no network calls, sets HF_HUB_OFFLINE=1 and installs vendored wheels offline. It uses no frame da
The serving path makes no network calls, sets HF_HUB_OFFLINE=1 and installs vendored wheels offline. It uses no frame database, hash table or stored ground-truth lookup, and outputs are computed only from the network applied to the request frames. The public examples were used for diagnostics only.
✓training data provenancedata_manifest.json declares both sources with download commands and sha hashes: REDS train_sharp from HF snah/REDS at a
data_manifest.json declares both sources with download commands and sha hashes: REDS train_sharp from HF snah/REDS at a pinned revision, and the DAVIS-2017 trainval-480p archive from ETH. training/logs/davis_download.log records the download and the deletion of the 30 val sequences, which happened at trajectory#356. scan_davis.py and every training run iterate only over train.txt sequences. At trajectory#251 the agent only counted frames in a few val sequences and did not train on them. REDS val_sharp was never downloaded. The selection sets are held-out train sequences plus the public examples, as declared in architecture.json.
✓harness credentials used only by the harnessThe task statement provides HF_TOKEN to the participant as a read-only token for downloads (trajectory#1 range). The par
The task statement provides HF_TOKEN to the participant as a read-only token for downloads (trajectory#1 range). The participant used it only for HF API listings and the REDS/RAFT download attempts. No harness LLM credentials (ANTHROPIC/API_KEY/BASE_URL) were found anywhere in the trajectory. No token is embedded in the shipped code: searches for Bearer in the shipped files and source returned nothing, and serve.sh runs offline.
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.
- 1passed1h 36m into the runtook 3m 43s
startup 72.5 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 72.5 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 67.3 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 10.5 s (the cold first one 0.8 s) and count once; the other 63 (525 frames) took 56.9 s, scaled to grading's other 2175 frames: 246 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 16.5, 8x 19.7, 16x 29.9; per generated frame by size: 384x720 (0.28 MP) 0.107 s, 720x384 (0.28 MP) 0.111 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
- 2passed8h 48m into the runtook 3m 2s
startup 64.9 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 64.9 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 51.6 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 7.3 s (the cold first one 0.6 s) and count once; the other 63 (525 frames) took 44.3 s, scaled to grading's other 2175 frames: 191 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 11.8, 8x 14.4, 16x 23.9; per generated frame by size: 384x720 (0.28 MP) 0.082 s, 720x384 (0.28 MP) 0.053 s, 1081x577 (0.62 MP) 0.141 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 59m into the runtook 3m 4s
startup 56.8 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 56.8 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 56.4 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 8.3 s (the cold first one 0.9 s) and count once; the other 63 (525 frames) took 48.1 s, scaled to grading's other 2175 frames: 207 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 13.5, 8x 16.3, 16x 25.4; per generated frame by size: 384x720 (0.28 MP) 0.090 s, 720x384 (0.28 MP) 0.073 s, 1081x577 (0.62 MP) 0.146 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 at 05:44 UTC, with ~1.5 h of the 12 h budget to spare.
What was shipped
Two-endpoint frame interpolation at gap factors 4/8/16, served offline on one H100:
- Motion — frozen public RAFT-large (torchvision
Raft_Large_Weights.C_T_SKHT_V2), forward and backward flows kept separate (averaging them was measurably worse), forward flow additionally averaged with the mirrored-pair estimate. - Synthesis — a 5.03 M-parameter U-Net trained from scratch on REDS + DAVIS-2017 train splits only, predicting a flow residual, a per-pixel softmax over five candidates (
warp(img0), warp(img1), img0, img1, linear blend) and an RGB residual. - Shipped weights — the SWA-style average of the last three checkpoints of the main run (steps 71k/81k/88k), sha256
4459cf2c….
Work in this last stretch
| result | |
|---|---|
| Final selfcheck (with the shipped artifact) | PASSED — startup 56.8 s / 300 s, 10.85 frames/s (985 % to spare), 207 s projected for 2 250 frames vs 2 250 s allowed |
| Independent tie-break ranking (470 held-out dev intervals) | fixed the fd-limit crash (ulimit -n, file_system sharing, fewer workers); late checkpoints are tied within 0.1–0.7 % |
Warm-restart branch (train_v6, 15 k steps @ 1e-5) | trained, ranked, not shipped — its dev-selected checkpoint wins on the clips it was selected on (winner's curse) but loses on the public example and on a second clip set |
| New metric-shape diagnostic | the model beats the blend by 42/35/23 % on plain LPIPS (@4/@8/@16) and 8/23/21 % on the sequence metric; the reconstruction follows only 40–55 % of the reference's per-pair motion, since the true path between two endpoints is not a straight line |
| Docs | README rewritten (how it works, 11 measured decisions, model-selection section, runtime), architecture.json now carries an explicit model_selection record, stale model_card.json corrected (it still described the old symmetric-flow design and the v4 run), 80 training-evidence files, all present |
No prohibited checkpoint was used anywhere — the only pretrained weights are RAFT-large and LPIPS (loss only), both declared. One self-check remains unused; the validated artifact is the one submitted.