jpeg.h File Reference

JPEG codec, baseline and progressive. More...

#include "tinyimg/codec/codec.h"

Go to the source code of this file.

Variables
const TinyCodec	tiny_codec_jpeg
	The JPEG codec.

Detailed Description

JPEG codec, baseline and progressive.

Author

Gregory Mitchell (me@gm.nosp@m.itch.nosp@m.215.x.nosp@m.yz)

Version

1.0.0

Date

2026-09-02

Copyright

Copyright (c) 2026

Variable Documentation

tiny_codec_jpeg

const TinyCodec tiny_codec_jpeg

extern

The JPEG codec.

Decodes baseline and extended sequential Huffman streams and progressive ones, at any sampling factors, with restart markers, and for one, three or four components: grayscale, YCbCr, and CMYK or YCCK through the Adobe APP14 transform. Arithmetic coded and lossless streams are recognized and reported as an unsupported variant rather than as corrupt.

Chroma is upsampled with the triangle filter libjpeg calls fancy upsampling for the 2x cases and by replication otherwise, so the output matches what djpeg produces rather than being a nearest neighbor approximation of it.

A scaled decode is done in the DCT domain: the top left NxN coefficients of each block go through an N point inverse transform, which is both smaller and cheaper than transforming all 64 and averaging. That is a low pass reduction rather than a box average, so it differs from the other codecs' scaled decode by a fraction of a level, and the region origin is rounded down to a multiple of scale_den, since a scaled plane has no sample at a finer offset.

A region decode holds only the MCU rows the region touches, plus one row either side for the upsampler's context. The entropy coded data before them still has to be scanned, because a Huffman stream cannot be entered at an arbitrary point and DC coefficients are differential, but nothing outside the region is transformed or stored.

Progressive streams are the one exception to bounded memory. Successive approximation means no coefficient is final until the last scan that touches it, so the whole coefficient plane exists before any pixel does: two bytes per sample, which for 4:4:4 is six bytes per pixel. A region decode of a progressive file therefore costs the same as a full one, and a large enough file reports TINYIMG_ERR_MEMORY where the same image baseline coded would decode.

Encoding writes baseline or progressive streams with Huffman tables built from the image's own symbol frequencies rather than the Annex K examples, which costs one extra pass over the coefficients and is worth 2 to 6 percent. Quantization tables are the Annex K ones scaled by the libjpeg quality mapping.