Model
Real-ESRGAN
Reconstruct a sharper, super-resolved image of a face or licence plate from many sightings.
shipped
Super-Resolution
Reconstruction4xCUDA
Overview
A vendored SRVGGNetCompact (realesr-general-x4v3) provides a 4x learned upscaler. Overseer caps and ECC-aligns a subject's crops, median-fuses them at low resolution to remove per-frame noise, then runs a single super-resolution pass so a blurry face or plate becomes legible. Weights load with no basicsr / gfpgan dependency.
| Role | Super-Resolution |
|---|---|
| Expected latency | ~0.3 s / reconstruction (GPU) |
| Hardware requirements | NVIDIA GPU recommended; CPU fallback works but is slow |
| Memory requirements | ~1 GB VRAM; ~5 MB model download |
Inputs & Outputs
Inputs
- Aligned, fused low-res crop (BGR)
- Max side hint
Outputs
- 4x upscaled BGR image
- Reconstruction metadata (frames used / offered)
Advantages & Limitations
Advantages
- Legible faces / plates from blur
- No dependency hell (vendored net)
- Fuses multiple frames for real detail, not hallucination
Limitations
- Cannot invent detail absent from all frames
- Mismatched crops are rejected to avoid worse-than-zoom output
- GPU strongly preferred for interactive latency
Example outputs
Real-ESRGAN
A readable bus destination sign from a blurry crop
Real-ESRGAN
A clarified face from 12 sightings
Real-ESRGAN
Original vs super-res comparison in the viewer
Configuration
Keys in config/default.yaml that govern this model.
| Key | Default |
|---|---|
reconstruct.super_resolution | true |
reconstruct.min_frames | 2 |
reconstruct.max_frames | 16 |
reconstruct.min_corr | 0.72 |
Benchmarks
Placeholder
Benchmark figures below are illustrative and awaiting a standardised harness. Treat them as placeholders.
| Benchmark | Dataset | Metric | Value | Hardware |
|---|---|---|---|---|
| Upscale factor | internal | x | 4 | - |
| Reconstruct | 16 crops | s | ~0.3 | RTX 3060 |