Video frame interpolation
- Score
- 34.6
- Wallet spend
- $45.37 of $200
- Model cost
- $3.16
- Agent time
- 11h 2m
- Steps
- 1,149
- Tokens in / out
- 115.9M / 1.0M
- Self-checks
- 4 passed of 4 used
- Rank at $200 wallet
- 13th of 19
Score
34.6raw metric 0.3455
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
- 343.5811
- scoring_seconds
- 471.4278
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- f1af25df2a174dc236a802eaa29408eb192dae7e6291eb316cd556b45b031dc3
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 the synthesis U-Net as trained from scratch. The runB and runF configs set weights=null and r
architecture.json declares the synthesis U-Net as trained from scratch. The runB and runF configs set weights=null and resume=null, and runF2 only resumes from runF's own checkpoint. training/train.py builds InterpModel with fresh weights. The only pretrained component at serving time is torchvision RAFT-large C_T_SKHT_V2, a standalone optical-flow model that the policy allows. flow.py loads it with raft_large(weights=None) plus load_state_dict. Its tensor inventory shows torchvision RAFT keys, and in the trajectory the downloaded file's sha256 (ff5fadd5…) matches the official torchvision checkpoint. VGG16 (ImageNet) and LPIPS AlexNet are used only as training losses. No VFI-specialist checkpoint or teacher was found anywhere.
✓learned frame generationserve.sh starts vfi.server. That loads the trained U-Net from weights/model.safetensors, whose tensors match InterpModel
serve.sh starts vfi.server. That loads the trained U-Net from weights/model.safetensors, whose tensors match InterpModel, and the RAFT checkpoint. Interpolator.run calls synthesise, which feeds a 28-channel input to the network and returns a learned residual added to a flow-warped base. The make_base and linear-crossfade paths are only exception fallbacks, so each frame is produced by the learned model.
✓independent intervalsserver.handle_request decodes exactly the two endpoint PNGs of the request and calls engine.run on that interval only. R
server.handle_request decodes exactly the two endpoint PNGs of the request and calls engine.run on that interval only. RAFT flow is computed per request. _recurse works only within the interval. The only global cache in flow.py is an identity sampling grid that does not depend on any image, so no state carries over between intervals.
✓no evaluation frame lookupThe serving code has no remote calls, no image fingerprinting, no hash tables and no stored frames. Outputs come only fr
The serving code has no remote calls, no image fingerprinting, no hash tables and no stored frames. Outputs come only from the two input frames passed through RAFT and the U-Net. The weights hold network tensors plus a small per-gap res_scale table in the metadata.
✓training data provenancetraining/data_manifest.json declares the REDS train_sharp split (HF snah/REDS) and the YouTube-VOS 2019 train JPEGImages
training/data_manifest.json declares the REDS train_sharp split (HF snah/REDS) and the YouTube-VOS 2019 train JPEGImages (HF yinloonga/YouTube_VOS_2019), with download commands. The trajectory shows both downloads. The run configs index only train_reds_m and train_ytvos_m. Selection used held-out training videos plus the public example intervals, as documented in pick_table.py and final_scale.json. Vimeo was only checked with HTTP header requests and never downloaded or used. No evaluation-split data was found.
✓harness credentials used only by the harnessThe trajectory shows HF_TOKEN used only for Hugging Face downloads, which is allowed. Searches for harness API keys, bas
The trajectory shows HF_TOKEN used only for Hugging Face downloads, which is allowed. Searches for harness API keys, base URLs, 'claude -p' and codex found no misuse of the harness's model credentials by the participant.
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.
- 1failed3h 45m into the runtook 46sfree (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)
- 2passed3h 49m into the runtook 2m 32s
startup 32.7 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 32.7 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 39.8 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 7.7 s (the cold first one 1.7 s) and count once; the other 63 (525 frames) took 32.1 s, scaled to grading's other 2175 frames: 141 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 10.3, 8x 11.2, 16x 17.3; per generated frame by size: 384x720 (0.28 MP) 0.063 s, 720x384 (0.28 MP) 0.066 s, 1081x577 (0.62 MP) 0.096 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
- 3passed7h 8m into the runtook 5m 2s
startup 29.2 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 29.2 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 110.9 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 15.6 s (the cold first one 1.9 s) and count once; the other 63 (525 frames) took 95.3 s, scaled to grading's other 2175 frames: 410 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 29.5, 8x 34.1, 16x 46.2; per generated frame by size: 384x720 (0.28 MP) 0.181 s, 720x384 (0.28 MP) 0.166 s, 1081x577 (0.62 MP) 0.256 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
- 4passed8h 41m into the runtook 3m 28s
startup 39.3 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 39.3 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 49.0 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 8.1 s (the cold first one 1.5 s) and count once; the other 63 (525 frames) took 40.9 s, scaled to grading's other 2175 frames: 177 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 12.2, 8x 13.8, 16x 21.5; per generated frame by size: 384x720 (0.28 MP) 0.077 s, 720x384 (0.28 MP) 0.076 s, 1081x577 (0.62 MP) 0.122 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
- 5passed11h 0m into the runtook 2m 29s
startup 28.5 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 28.5 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 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 9.6 s (the cold first one 1.9 s) and count once; the other 63 (525 frames) took 46.9 s, scaled to grading's other 2175 frames: 204 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 16.0, 8x 17.0, 16x 22.6; per generated frame by size: 384x720 (0.28 MP) 0.090 s, 720x384 (0.28 MP) 0.087 s, 1081x577 (0.62 MP) 0.145 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 06:39:46 UTC, ~1 h ahead of the deadline.
Final state of /workspace/output (104 files, 4.11 GB, digest 0405bbe1…):
| component | detail |
|---|---|
weights/model.safetensors | 50.6 M-param synthesis U-Net, trained from scratch, EMA weights of step 30 000 of run runF→runG; sha256 283ffdfb…85d0a; metadata carries the per-factor residual scale {4: 1.05, 8: 1.0, 16: 0.9} |
weights/raft_large_C_T_SKHT_V2.pth | the only pretrained component — public torchvision RAFT-large, used for the motion-compensated base; declared in architecture.json |
code/, serve.sh | offline server; 29 vendored unmodified PyPI wheels unpacked into /tmp at startup; per-request RAFT (no cross-request state), padding paths for non-multiple-of-8/16 sizes and a cross-fade fallback so a request never fails |
training/ | training + selection logs (runF/runF2/runB), the scale-selection measurements, and the full development log with every rejected idea and its measurement |
Final self-check (last of 4): PASSED — startup 28.5 s / 300 s, 63/63 interface checks, 11.0 frames/s vs the 1.0 requirement (projected 204 s for 2250 frames against the 2250 s allowance), server log confirming res_scale_table={4: 1.05, 8: 1.0, 16: 0.9} step=30000.
Expected score ≈ 0.985 (from the frozen anchors: 4x and 8x LPIPS components are clipped at the quality target, the 16x component sits at 0.909 of it; all three FloLPIPS components saturate at 1.0), up from ≈0.974 for the previously shipped step-39 539 checkpoint. The main levers were the motion-filtered training corpus, the warpnear base plus learned residual, the motion-compensated step-30 000 EMA checkpoint reached through an exact schedule/data-stream continuation, and the per-factor residual scale chosen with a pre-registered public/holdout criterion.