Lossy Video Compression
Lossy video compression reduces data by allowing the reconstructed pictures to differ from the encoder input.
Lossy video coding saves data by allowing a reconstructed picture to differ from the encoder's input. The useful question is which information was approximated, and what did that change in the picture?
Which corrections survive quantization?
Suppose a predictor supplies values 100, 100, 100 for three positions. The actual input is 117, 97, 139, so the differences needed to reproduce it are 17, −3, 39.
Now use a deliberately simple quantizer that rounds those corrections to the nearest multiple of eight:
| Position | Input | Predictor | Exact correction | Rounded correction | Decoded result |
|---|---|---|---|---|---|
| 1 | 117 | 100 | 17 | 16 | 116 |
| 2 | 97 | 100 | −3 | 0 | 100 |
| 3 | 139 | 100 | 39 | 40 | 140 |
The decoder adds the rounded correction to the predictor. Its errors relative to the input are −1, +3, +1.
Original scalar example. It omits transform coding and uses arbitrary sample values to isolate quantization; it is not the exact quantizer of a named video codec.
The compressed representation says “zero correction” for position 2. That could have arisen from −3, +2, or zero. Several different inputs map to the same coded result, so the decoder has no information with which to choose the original one. This many-to-one mapping is the irreversible step.
Why approximation can reduce data
Small corrections often occur where a predictor already describes the picture well. Rounding more of them to zero leaves fewer significant values to transmit. Coarser quantization can also reduce the precision required to describe the remaining values.
In block-transform codecs, quantization operates on coefficients representing patterns across a region, rather than directly on the simple scalar corrections above. Removing or coarsening a coefficient changes the corresponding pattern when the block is reconstructed. A fine texture can weaken, or an edge can gain a faint oscillating outline.
The complete compression walkthrough connects prediction, transform, quantization, and reconstruction. Lossy behavior comes from the permitted approximation, rather than from the fact that prediction was used.
Allocate the compromise within a scene
A stationary wall may have predictable regions and small corrections. Grainy foliage in motion can have more changing information. Preserving the latter at the same fidelity needs a different allocation of bits.
An encoder makes decisions about predictors, block divisions, and quantization strength. With a constrained bitrate, difficult regions compete for that budget. With a quality target, the encoder can let size grow as the source becomes harder.
The visible consequences follow the information affected. Simplified texture looks smoother; disagreeing neighboring regions can reveal block boundaries; coarse gradients can lose distinct levels. Some differences remain invisible at the intended display size, while others affect details the viewer notices quickly.
Identify the stage being compared
For a library file that is played directly, the decoder reconstructs the representation already stored. During video transcoding, Emby Server reconstructs that source and supplies its decoded pictures to another encoder.
The second encoder's input already contains the first encode's changes. Giving the second stage more bits can preserve that input more closely; it cannot reverse the first stage's many-to-one decisions. Generation loss follows what happens when such stages are repeated.
A useful comparison keeps the same source pictures, timing, and presentation and examines the converted reconstruction against them. That reveals the additional change caused by this encode, which is the quantity its quality settings can control.