Video frame interpolation
- Score
- 46.3
- Wallet spend
- $38.39 of $50
- Model cost
- $32.11
- Agent time
- 10h 34m
- Steps
- 715
- Tokens in / out
- 72.2M / 378k
- Self-checks
- 3 passed of 3 used
- Rank at $50 wallet
- 11th of 19
Score
46.3raw metric 0.4626
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
- 327.4139
- scoring_seconds
- 521.4674
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- 062beb54c67765fa1e1865b526549bdb16b4537e165f654599a11e3454bfe8f5
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 initializationarchitecture.json declares random initialization with base_weights null. In training/train_ema.py, init_model builds the
architecture.json declares random initialization with base_weights null. In training/train_ema.py, init_model builds the EMA-VFI-style network from random weights, with optional zero-init of the final residual conv. Resume and EMA-init read only from the participant's own /workspace/runs. The trajectory shows the EMA-VFI, RAFT and IFRNet repos were cloned as source code only (trajectory#16). A GIMM-VFI raft-things.pth link showed up only in a WebSearch result (trajectory#73) and was never downloaded. The served weights.pth (sha a0c49e50) was copied from the participant's own run, /workspace/runs/ema_final_anneal/ema_030000.pth (trajectory#704). Its tensor inventory (feature_bone.*, unet.*) matches the shipped model code. The pretrained components it declares are AlexNet/LPIPS, RAFT and PWC-Net, which are general or standalone flow/perceptual models. They are used only for losses or metrics and are not served.
✓learned frame generationserver.py loads weights.pth strictly into build_model (modeling.py → model/*), then generates each frame with a model fo
server.py loads weights.pth strictly into build_model (modeling.py → model/*), then generates each frame with a model forward pass: feature_bone, calculate_flow at time t, then coraseWarp_and_Refine. There is no non-learned fallback path and no external model.
✓independent intervalsserver.py handles each request on its own. It decodes only that request's two endpoint PNGs and computes t from the targ
server.py handles each request on its own. It decodes only that request's two endpoint PNGs and computes t from the target index. No global state persists between requests; the only caches are image-independent coordinate, mask and kernel caches. A search of the model code found no request caching. Warm-up runs on zero tensors only.
✓no evaluation frame lookupThe serving path reads only the request's two input frames and the fixed weights. It makes no network calls, holds no st
The serving path reads only the request's two input frames and the fixed weights. It makes no network calls, holds no stored frame database and does not read target paths. Training data came only from Vimeo train shards plus 85 general training videos. Checkpoint selection used the public task examples and reserved training videos, not evaluation frames.
✓training data provenancetraining/download_data.py and the manifests in training/data/*_download.json declare the corpora as Hugging Face snapsho
training/download_data.py and the manifests in training/data/*_download.json declare the corpora as Hugging Face snapshots pinned to fixed revisions: bijinc/vimeo-90k-medium at 17ba080a (train parquet shards) and danjacobellis/vimeo_fm85 at 4aad54ce (train split). The trajectory shows these downloads being run (e.g. trajectory#29, #36). Both are declared public sources with a retained download script and logs. LPIPS is the only loss model in the main lineage. The PWC-Net flow loss branch was rejected. No VFI teacher or pseudo-labels were used.
✓harness credentials used only by the harnessThe trajectory has no ANTHROPIC/OPENAI credential use, no claude or codex CLI launches, and no use of `spin openrouter k
The trajectory has no ANTHROPIC/OPENAI credential use, no claude or codex CLI launches, and no use of `spin openrouter key create`. The one environment probe only printed whether the HF_/SPIN/FRONTIER variables were set (trajectory#4). Compute ran through the sanctioned spin CLI. The SSH key came from `spin compute ssh` for the participant's own lease (trajectory#160, #161), so it is a compute-access key, not a harness model-API credential. The single Agent call (trajectory#8) is the participant harness's own built-in subagent tool.
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 16m into the runtook 3m 55s
startup 67.2 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 67.2 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 90.4 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 12.4 s (the cold first one 1.2 s) and count once; the other 63 (525 frames) took 78.0 s, scaled to grading's other 2175 frames: 336 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 28.1, 8x 29.0, 16x 32.1; per generated frame by size: 384x720 (0.28 MP) 0.147 s, 720x384 (0.28 MP) 0.143 s, 1081x577 (0.62 MP) 0.186 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
- 2passed6h 17m into the runtook 3m 25s
startup 57.8 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 57.8 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 80.9 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 11.8 s (the cold first one 1.1 s) and count once; the other 63 (525 frames) took 69.2 s, scaled to grading's other 2175 frames: 298 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 22.0, 8x 24.7, 16x 32.9; per generated frame by size: 384x720 (0.28 MP) 0.130 s, 720x384 (0.28 MP) 0.133 s, 1081x577 (0.62 MP) 0.184 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
- 3passed10h 31m into the runtook 2m 54s
startup 54.5 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 54.5 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 54.6 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 0.7 s) and count once; the other 63 (525 frames) took 47.3 s, scaled to grading's other 2175 frames: 203 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 13.7, 8x 16.2, 16x 23.7; per generated frame by size: 384x720 (0.28 MP) 0.089 s, 720x384 (0.28 MP) 0.090 s, 1081x577 (0.62 MP) 0.099 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 completed successfully at 2026-10-24 11:32:03 UTC.
Final package:
- EMA step 30,000, selected over later checkpoints using matched quality evaluations
- FP16 inference
- RGB8 and legacy PNG interfaces validated
- Determinism, request independence, and irregular-shape tests passed
- Official offline H100 selfcheck passed
- Startup: 54.5 s
- Throughput: 11.07 frames/s
- Checked digest:
cf105457c04d0adce78601d35881d651521f4b816ec10493e632c46a90339f0f
- Package: 125 files, 4.19 GB under
/workspace/output/