Why moderate mosaic often comes back close to the original
Mosaic hides information; it does not delete all of it. What disappears is mostly high-frequency detail. Low-frequency structure, edges, motion and the small differences between neighbouring frames survive — and that is what restoration works from.
What is left after mosaic
| Information | After mosaic | Usable? |
|---|---|---|
| Low-frequency structure (blocks of light and shade, composition) | Largely intact | ✅ directly |
| Edges and outline direction | Coarsened but readable | ✅ guides reconstruction |
| Motion | Preserved | ✅ used for alignment |
| Inter-frame sampling differences | Preserved | ✅ aggregated, raises effective detail |
| High-frequency texture (pores, hair) | Removed | ⚠️ reconstructed or inferred |
Stitching it back together
- Align — line up the same region across neighbouring frames, cancelling motion.
- Aggregate — stack those views; effective sampling density goes up and finer detail becomes resolvable.
- Reconstruct — rebuild texture on the aggregated signal and blend it back so the sequence does not flicker.
When it can only infer
- Very large mosaic blocks — even low-frequency structure is destroyed.
- Very wide occlusion — there is nothing nearby to reference.
- Fast motion plus occlusion — frames will not align, so aggregation fails.
In those cases the model produces a plausible image: natural-looking, but not guaranteed to match the original pixel for pixel.
Frequently asked questions
Does mosaic erase the picture information completely?
No. It removes mostly high-frequency detail. Low-frequency structure, edges, motion and inter-frame sampling differences remain, and restoration works from those.
Why does using multiple frames help?
Neighbouring frames sample the same region slightly differently. Aligning and stacking them raises the effective sampling density, so finer detail becomes resolvable.
Are there cases that cannot be recovered?
Yes — very heavy mosaic, very wide occlusion, or fast motion that prevents frame alignment. There is too little left, so the result is a plausible inference.
More topics on the guides hub, or see what Lumasce does.