Video frame interpolation
- Score
- 56.8
- Wallet spend
- $5.48 of $10
- Model cost
- $1.08
- Agent time
- 10h 35m
- Steps
- 436
- Tokens in / out
- 43.4M / 365k
- Self-checks
- 3 passed of 3 used
- Rank at $10 wallet
- 9th of 19
Score
56.8raw metric 0.5677
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
- 347.0265
- scoring_seconds
- 495.9750
- evaluation_sha256
- 375276285d8168e061004e5873daeecbb7e8db7c6e8354ef72d50ad47496e236
- metric_bundle_sha256
- 98e13949365e1541fbec7972317356abb77c00956c7126d0d81dd99f045d5fdd
- calibration_sha256
- 8c72e23cb352711cfddff07fea74a4935a99f753ea7b498dbd83a999278630ab
- prediction_manifest_sha256
- b0ddd477383e2a490f2e873578ab9434f410e5c169b7d5f220ca80cc90999e56
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 initializationThere is no specialist (VFI) initialization. The only pretrained component is torchvision RAFT-large (C_T_SKHT_V2). It i
There is no specialist (VFI) initialization. The only pretrained component is torchvision RAFT-large (C_T_SKHT_V2). It is a standalone optical-flow model, it stays frozen, and it is declared in architecture.json with a sha256. trajectory#412 shows the file was copied unchanged from the torch hub cache to weights/raft_large.pth, and that file's tensor inventory contains only RAFT modules (feature_encoder.* and so on). SynthNet is freshly initialized in training/train.py with no teacher and no VFI checkpoint. The later scheduled-sampling stages (train_ss.py, train_ss_deep.py) start only from the participant's own checkpoints. The weights/synth.pt inventory contains only model.* SynthNet tensors (stem through head, base 56), which matches the architecture in vfi_lib.py. No VFI checkpoint downloads or from_pretrained calls appear in the trajectory.
✓learned frame generationserve.sh installs the vendored wheels and runs serve.py. serve.py loads weights/synth.pt (EMA weights, or the model weig
serve.sh installs the vendored wheels and runs serve.py. serve.py loads weights/synth.pt (EMA weights, or the model weights as fallback) into SynthNet and loads the frozen RAFT from weights/raft_large.pth. In vfi_lib.py, Interpolator.interpolate estimates bidirectional RAFT flows and turns them into constant-velocity intermediate flows (flow_inputs). It backward-warps both endpoints and has the learned SynthNet predict a blend mask and a residual (mask*w0 + (1-mask)*w1 + res). Intermediate frames come from this recursive midpoint synthesis at t=0.5, so the output frames are generated by the trained network.
✓independent intervalshandle() in serve.py decodes only frames[0] and frames[T-1] of the current request and passes them to Interpolator.inter
handle() in serve.py decodes only frames[0] and frames[T-1] of the current request and passes them to Interpolator.interpolate. The optional polish step uses only neighbouring frames generated within the same request. The only state kept across requests is frames_done, rate and max_pixels, which control compute and quality settings and never carry image content. No other requests or intervals are used.
✓no evaluation frame lookupNeither serve.py nor vfi_lib.py contains image caches, hashing or fingerprinting of inputs, stored reference frames, ret
Neither serve.py nor vfi_lib.py contains image caches, hashing or fingerprinting of inputs, stored reference frames, retrieval by content, or remote calls. The only bundled data is weights and wheels: the weight inventories show model tensors only, and no frame datasets are shipped. Outputs are computed by the network on every request.
✓training data provenanceDATA.md, training/prep.py and training/get_vimeo.py declare the training corpora. They are UCF101 (the official trainlis
DATA.md, training/prep.py and training/get_vimeo.py declare the training corpora. They are UCF101 (the official trainlist01 split, via HF quchenyuan/UCF101-ZIP, 5,000 videos), DAVIS-2017 (train.txt sequences only) and a general raw-video corpus from HF danjacobellis/vimeo2k (150 videos, seed 11). The internal validation set is an md5-bucket split of these training clips. The download, preparation and training commands are kept in DATA.md and the scripts. The full download logs are not all kept, so the downloads themselves are partly unverified. A Vimeo90K triplet download was attempted and blocked (trajectory#12 and trajectory#13), and that data was not used. The trajectory shows no use of SNU-FILM, GoPro or other evaluation benchmark data. Public task examples were used only for engineering checks and variant selection, which the policy allows.
✓harness credentials used only by the harnessThe trajectory searches found no use of ANTHROPIC/OPENAI/OPENROUTER/GEMINI keys or *_BASE_URL variables by the participa
The trajectory searches found no use of ANTHROPIC/OPENAI/OPENROUTER/GEMINI keys or *_BASE_URL variables by the participant. Early steps show only spin CLI help text. HF_TOKEN appears only as a read-only bearer token for Hugging Face dataset downloads (get_vimeo.py and trajectory#12), which the policy allows.
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.
- 1passed53m 43s into the runtook 5m 42s
startup 68.0 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 68.0 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 65.9 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 10.3 s (the cold first one 0.5 s) and count once; the other 63 (525 frames) took 55.6 s, scaled to grading's other 2175 frames: 241 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 10.0, 8x 18.0, 16x 36.6; per generated frame by size: 384x720 (0.28 MP) 0.104 s, 720x384 (0.28 MP) 0.102 s, 1081x577 (0.62 MP) 0.192 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 13m into the runtook 3m 21s
startup 55.9 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 55.9 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 72.2 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 0.7 s) and count once; the other 63 (525 frames) took 59.8 s, scaled to grading's other 2175 frames: 260 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 20.3, 16x 39.1; per generated frame by size: 384x720 (0.28 MP) 0.114 s, 720x384 (0.28 MP) 0.125 s, 1081x577 (0.62 MP) 0.217 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
- 3error8h 32m into the runtook 1h 0mfree (stopped before any GPU time)
The self-check could not finish (TimeoutError: Grading resource operation timed out: inference sandbox start did not return within 3630 s).
1 gates
- ✓review prerequisites
- 4passed9h 40m into the runtook 3m 21s
startup 57.1 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 57.1 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 83.9 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 12.5 s (the cold first one 0.6 s) and count once; the other 63 (525 frames) took 71.4 s, scaled to grading's other 2175 frames: 308 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 13.0, 8x 23.4, 16x 46.5; per generated frame by size: 384x720 (0.28 MP) 0.135 s, 720x384 (0.28 MP) 0.135 s, 1081x577 (0.62 MP) 0.218 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 06:12:49 UTC. Final state of the work:
What was built
A two-endpoint VFI system that generates all intermediate frames recursively at t=0.5, from two frozen, permitted components and one network trained here from scratch:
- RAFT (
torchvisionraft_large, C_T_SKHT_V2, unmodified, hash-declared) — bidirectional flow only. - SynthNet (25.5M-param U-Net, base 56, 19→4 channels) — trained from scratch in three stages on UCF101 fold-01 train, DAVIS 2017 train, and a raw Vimeo corpus: (1) real-pair midpoint training with serving-matched gap sampling, (2) scheduled sampling feeding its own generated anchors back as endpoints, (3) a deeper scheduled-sampling stage training pairs with both endpoints network-generated — the input the deepest 16x recursion levels actually consume.
No VFI checkpoint, descendant, teacher, or fallback was used anywhere.
Final measured results (shipped stage-3 EMA, step 8,000)
| probe | 4x | 8x | 16x |
|---|---|---|---|
| public LPIPS | 0.01873 | 0.03507 | 0.06215 |
| public normscore (LPIPS-only) | 0.837 (started 0.550) | ||
| hard proxy vs linear blend | 0.696 | 0.716 | 0.940 |
| smooth proxy vs blend | 0.548 | 0.643 | 0.962 |
Each stage improved every factor of every probe; the deep stage alone moved public normscore 0.818 → 0.837 and hard-16x 0.963 → 0.940.
Serving
serve.sh → stdlib HTTP server, vendored cu128 wheels, fp16 weights. Self-check on the grading image passed on the exact submitted digest: startup 57.1 s (limit 300), 7.30 frames/s (required 1), projected 308 s for all 2,250 frames (limit 2,250), healthy after requests. Verified separately end-to-end through the real serve.py on GPU: byte-exact rgb8 bodies at all three factors.
Rejected by measurement: flip/TTA averaging, RAFT polish, coarse top-level, composed and multi-scale flows, sharpening, blend shrinkage, extra RAFT updates (12 beat 20 and 28 on the public score), and a further 4k-step continuation (a wash).