Mosaic & pixelation restoration, explained
The short version: mosaic hides detail rather than erasing it. Restoration stages the surviving information back together, so light to moderate mosaics can come back close to the original picture; only heavy or wide occlusion falls back to a plausible guess.
Three stages
- Detection. Find the degraded regions frame by frame. Faster presets trade hit rate for speed; accurate presets are slower but catch more.
- Temporal aggregation. The same region shifts slightly from frame to frame. Align those views and the usable information grows.
- Reconstruction. A model rebuilds detail on top of the aggregated signal, then blends it back so the sequence stays stable.
How far can it go?
| Mosaic severity | What survives | Typical result |
|---|---|---|
| Light (small blocks, simple texture) | Plenty of structure | Close to the original; natural detail |
| Moderate (medium blocks, movement) | Substantial structure | Clearly improved; often close to the original |
| Heavy / wide occlusion | Very little | Plausible inference: natural-looking, not guaranteed pixel-accurate |
Why moderate mosaic often comes back close to the original
Mosaic is downsampling followed by upscaling, which removes mainly high-frequency detail. Low-frequency structure, edges and motion survive — and because neighbouring frames sample the scene slightly differently, those differences remain available too. Restoration exploits exactly that residual signal, which is why moderate footage frequently returns close to the original rather than being invented from nothing.
Choosing detection quality
- Scan fast, then go accurate. Run a fast preset over the whole clip, then re-run the important shots at higher quality.
- Match the content. Close-ups of faces deserve the accurate preset; landscapes and motion shots usually look fine on fast.
- Watch clip length. For long jobs, cap the per-segment length to keep memory and stability in check.
Frequently asked questions
Does restoration just paint over the mosaic?
No. It aggregates evidence across neighbouring frames and reconstructs detail from it. Mosaic is downsampling, so low-frequency structure, edges and motion survive — that is the evidence being used.
Can moderate mosaic really approach the original image?
Often, yes. The more residual signal survives, the closer the reconstruction gets to the original picture.
When does quality drop?
Heavy mosaic, very wide occlusion, or fast motion combined with occlusion — there is simply too little left to work with, so the model can only infer a plausible image.
How should I choose the detection preset?
Scan the whole clip with the fast preset, then re-run key shots with the accurate one. Faces favour accurate; landscapes and motion shots are usually fine on fast.
More topics on the guides hub, or see what Lumasce does.