loglens-nano — the benchmark models (128K–202K params)
Checkpoints from the LogLens study of fixed logarithmic memory schedules for multi-timescale change localization on fixed cameras.
Read this first: these are benchmark models — they prove what a from-scratch nano model can and cannot do. On the synthetic benchmarks they are strong (tables below). On real footage they hit a measured clutter wall (unlearnable below ~1% object-to-frame ratio against real backgrounds), and the deployable real-scene system is training-free: frozen CLIP + the same log schedule + a cosine-gap readout (0.796 on real gold, zero parameters). Code and findings in the repo.
Files
| file | what | result |
|---|---|---|
log_K8_s0.pt |
temporal LogLensNet, auto-base log schedule, K=8 (128K params) | 0.901 ± .017 TriClock (3 seeds) |
logtuned_K8_s0.pt |
boundary-aligned schedule | 0.922 ± .009 |
log_K8_s0.onnx |
ONNX export (543 KB) | 0.7 ms/query CPU-only on AGX Orin, ~3 KB memory state |
logzoom_K5_s0.pt |
tri-axis HDNet, saliency-guided log-zoom (202K params) | 0.807 ± .012 TriClock-HD |
oraclebox_K5_s0.pt |
tri-axis, oracle glimpses (ceiling reference) | 0.999 |
The two headline findings these models carry
- Schedule geometry beats capacity: same model, same budget — a log- spaced memory schedule scores 0.90 where a sliding window scores 0.48; optimal K is width-independent across a 16× parameter range.
- Coverage is information-sufficiency; guidance is learnability: full-coverage unguided tiling has every needed pixel (oracle ceiling 1.0) and learns nothing (0.25 floor); centered glimpses at the same resolution reach 0.991.
Demo: https://huggingface.co/spaces/resoajoe/loglens · Benchmark: https://huggingface.co/datasets/resoa/triclock