JpegXlEncoder Class

Lossless output: efforts 0 to 2 code each group with a fixed color transform and a fixed predictor. From 3 on, a tree of contexts is learned for the whole frame from samples of its rows, each context with its own predictor, and the frame is coded with a palette of its colors (screenshots, graphics) or with a reversible color transform; from 4 on, each cluster of contexts codes its values in its own way; from 5 on, repeated runs of values are copied (LZ77) when a trial on part of the frame finds that copies make it smaller (screenshots, graphics; rarely photographs). The higher efforts learn from more samples, properties, predictors (the weighted predictor from 7 on, which roughly doubles the time) and contexts, and search further for copies. At every effort, the samples are written with prefix codes or with ANS, whichever is smaller. Lossy output: efforts 0 to 2 only use the 8x8 transform and quantize every block alike; from 3 on, each block has its own quantization steps (see JpegXlEncoder.Quality), each area has the 8x8 or the 16x16 transform, whichever costs less, each tile has its own factors of the chroma from luma prediction, the predictor of the alpha channel is searched, and the low-frequency image and the description of the blocks are coded with trees learned from their values; from 6 on, the contexts of the coefficients have more distributions; from 7 on, the 16x8 and 8x16 transforms are tried and the weighted predictor codes the low-frequency image where it suits it, and from 8 on the transforms of up to 64x64 samples.

Lossy output is written in three passes after the image at one eighth of its resolution; lossless output is squeezed, each resolution in the sections of its pass. The decoded pixels are those of a file that is not progressive; the file is larger (a few percent for lossy output, more for lossless output). ImageReader<TPixel>.ReadFrameRefinementInto reads such a file stage by stage.

The quality selects the quantization steps: 90 is a quality at which the difference with the original is barely visible at a normal viewing distance, and 100 is the finest quantization the encoder uses, still lossy. From the JpegXlEncoder.Effort 3 and above the quality 80, the steps of a block also follow its content in part: they are finer where the image is flat or next to an edge, where an error shows, and coarser in texture, which hides it. Below the quality 94, the decoder removes the ringing of coarse steps with the edge-preserving filter of the format, more strongly as the quality goes down.