YAPF — Yet Another Picture Format

Lossless images that decode several times faster than PNG — on one core.

YAPF is a lossless image format for photos, UI and game assets, built for decode speed first. There is no entropy coder, no bit reader and no context model: every structure is byte aligned, and the decoder runs short loops that modern CPUs execute at memory speed — with SSE / NEON where available. The encoder does the hard work, so the decoder only follows instructions.

• 5–12× faster decoding than PNG (stb_image), single-threaded; a 4K image decodes in about 6–7 ms (around 5 GB/s). Multithreaded decoding is optional.
• Smaller than PNG on UI, icons and screenshots; 20–45 % larger on photos and natural textures — the price of having no entropy coder.
• Made for games and tools too: mipmaps, GPU format hint, sRGB / premultiplied flags, and random access to 64×64 tiles.
• A real texture format (YAPF 2): GPU-ready BC1–BC7 / BC6H / ETC2 / ASTC blocks, 16-bit and float pixels, cubemaps, arrays and volumes, channel meanings (normal X/Y, roughness, AO, …), coverage-preserving mips, small-mips-first streaming, per-tile checksums, notes and animation — see Textures.
• Tiny and portable: two C99 files, no dependencies. Plus a CLI, Python, JavaScript and plugins for the tools you already use.

Two format versions: version 1 (an 8-bit image, the first stable format) and version 2 (the texture container). Plain 8-bit images are still written as version 1, so every existing YAPF reader keeps working.

---

Use it in…

Where	What you get	Guide	Status
C / C++	yapf.h + yapf.c, drop into any project (compiles as C or C++)	below	✅ tested — see platforms
Command line	convert PNG/JPG/TGA/BMP/PSD/GIF ↔ YAPF, compare, info, bench, thumbnail	below	✅ tested on Linux and Windows
Your language	Python, C#, Rust, Go, Java, Zig, Lua, JavaScript / TypeScript, …	docs/LANGUAGES.md	✅ every binding tested
GIMP 3	open, export (flattened, optional mipmaps)	extensions/gimp	✅ tested in GIMP 3.0.8
Blender 4.2+	import, export, drag & drop, image as plane	extensions/blender	✅ tested in Blender 5.1
KDE Dolphin	thumbnails for .yapf files	extensions/kde	⚠️ tested through KDE's plugin loader (KF6); not yet inside a running Dolphin
Linux desktop	file type + thumbnails in GNOME Files, Nemo, Caja, Thunar	extensions/linux	⚠️ thumbnail command tested; not yet in a file manager
VS Code	image preview with zoom and mip picker	extensions/vscode	⚠️ packaged; preview code tested in a browser engine (Chromium); not yet run inside VS Code
Figma	import .yapf, export selection as .yapf	extensions/figma	⚠️ plugin UI tested in a browser engine with a simulated Figma; not yet run inside Figma
Photoshop	File → Scripts → Open YAPF / Save as YAPF	extensions/photoshop	⚠️ syntax-checked only; not yet run in Photoshop

Pre-built downloads (command-line tool and libraries for Windows, macOS and Linux, plus the plugin packages) are published on the Releases page; everything can also be built from source as described below.

Platforms

Platform	Status
Linux x64	✅ built and fully tested
Windows x64	✅ built and fully tested (MinGW and Zig builds, run under Wine)
Linux ARM64	✅ built and fully tested (NEON, run under qemu)
Windows ARM64	⚠️ built only, not run
macOS x64 / ARM64	⚠️ cross-compiled only, not yet run on a Mac

---

Command line

yapf photo.png photo.yapf             # encode (formats by file extension)
yapf texture.png texture.yapf --mips  # with a full mip chain
yapf photo.yapf photo.png             # decode to PNG / JPG / BMP / TGA
yapf texture.yapf mip2.png --mip 2    # export one mip level
yapf info photo.yapf                  # size, channels, mips, flags, ratio
yapf bench photo.yapf                 # decode speed, 1 thread and all cores
yapf thumbnail photo.yapf thumb.png 256
yapf compare photo.png ui.png         # size and decode speed vs the original format

Options when writing .yapf: --mips, --linear (pixels are not sRGB), --premult (alpha is premultiplied), and the texture options below.

---

Textures (YAPF 2)

Version 2 turns YAPF into a texture container for engines, designed our own way rather than copied from KTX2 or DDS (the design and the full specification are in docs/FORMAT.md):

Need	YAPF 2
GPU-compressed payloads	BC1, BC2, BC3, BC4, BC5, BC6H, BC7, ETC2 / EAC and ASTC 4×4–12×12 blocks, stored as the GPU takes them. Built-in encoders for BC1, BC3, BC4, BC5, BC7 and BC6H; or bring blocks from your own encoder.
More than 8 bits	lossless 16-bit, half and float pixels, coded as byte planes with the same fast machinery
Cubes, arrays, volumes	cubemaps and cube arrays (reflection probes, skies), texture arrays (terrain layers), 3D textures (colour-grading LUTs)
Channel meaning	per channel: normal X/Y/Z, roughness, metalness, AO, height, opacity, …; OpenGL or DirectX normal convention; linear channels inside sRGB images
Streaming	small mips are stored first: a file prefix is a complete low-resolution texture (yapf info prints the prefix sizes); single 64 × 64 tiles decode on their own for virtual texturing
Alpha-tested mips	--coverage REF keeps fences and foliage from thinning out with distance; the filter used is recorded
Entropy coding (optional)	--entropy: shared per-surface Huffman code books with a neighbourhood context — about 10 % smaller photos, about 3× smaller GPU blocks on disk; tiles stay independent
Metadata	notes (author, licence, source, tool, ICC profile, orientation, DPI, HDR light levels, your own tags), colour primaries and transfer (sRGB, Display P3, BT.2020, PQ, HLG, …)
Integrity	yapf-sum checksums on the header, tables and every tile; needs bits make an older decoder refuse cleanly what it cannot read
Animation	frames with per-frame durations and a loop count

yapf albedo.png albedo.yapf --format bc7 --mips               # GPU-ready colour
yapf normal.png normal.yapf --format bc5 --mips --normal dx   # normal map, DirectX Y
yapf orm.png orm.yapf --map ao,rough,metal --linear --mips    # packed masks
yapf foliage.png foliage.yapf --format bc7 --coverage 128     # alpha-tested mips
yapf sky.hdr sky.yapf --format bc6h --mips                    # HDR, no RGBM tricks
yapf sky.hdr sky_lossless.yapf                                # lossless half float
yapf pack cube px.png nx.png py.png ny.png pz.png nz.png probe.yapf --mips
yapf pack array grass.png rock.png sand.png terrain.yapf --format bc7 --mips
yapf pack volume lut_00.png lut_01.png ... grading.yapf --linear
yapf pack anim f0.png f1.png f2.png fx.yapf --frame-ms 50
yapf albedo.png albedo.yapf --entropy --author "Keel" --license CC-BY-4.0
yapf info albedo.yapf       # everything in the header, notes, stream prefix sizes
yapf verify albedo.yapf     # check every checksum
yapf probe.yapf face2.png --face 2 --mip 1    # export one surface

In C:

int err;
yapf_texture_t *t = yapf_tex_load("albedo.yapf", 0, 0, &err);
for (uint32_t m = 0; m < t->mip_levels; m++) {
    uint32_t s = yapf_tex_surface(t, 0, 0, 0, m), w, h;
    yapf_tex_mip_size(t, m, &w, &h, NULL);
    if (t->payload == YAPF_PAYLOAD_BLOCKS)        /* e.g. YAPF_GPU_BC7 */
        glCompressedTexImage2D(GL_TEXTURE_2D, m, gl_format(t->format, t->flags),
                               w, h, 0, (GLsizei)yapf_tex_surface_bytes(t, m), t->data[s]);
}
yapf_tex_free(t);

Every extension opens version 2 files: GIMP (at full 16-bit / float precision, faces and frames as layers), Blender (HDR as float images), KDE and Linux thumbnails, VS Code (with face / layer / frame / slice pickers), Figma and Photoshop. The Python binding (yapf.load_texture) and the JavaScript port (YAPF.decodeTexture) read every texture; the other bindings wrap the v1 functions, which also read 8-bit version 2 files (2D, cube face 0, array layer 0).

---

Benchmark

yapf compare on a mixed set of images: one binary (make, gcc -O2) for both formats, PNG decoded with stb_image, one core, AMD Ryzen 7 5700G. Every result is checked to decode back to the exact original pixels.

Image	PNG	YAPF	YAPF / PNG size	PNG decode	YAPF decode	Faster
UI screenshot 261×430	176.6 KB	143.7 KB	81 %	2.05 ms	0.17 ms	12×
UI screenshot 305×529	201.5 KB	185.0 KB	92 %	2.55 ms	0.24 ms	11×
App icon 128×128	24.9 KB	23.1 KB	93 %	0.30 ms	0.026 ms	11×
Wallpaper 3840×2160	1.77 MB	2.21 MB	125 %	73.7 ms	6.8 ms	11×
Rock texture 512×512	182.6 KB	221.8 KB	121 %	1.64 ms	0.17 ms	10×
Wood texture 512×512	185.8 KB	234.1 KB	126 %	1.64 ms	0.17 ms	10×
Canvas texture 512×512	171.5 KB	245.6 KB	143 %	1.36 ms	0.16 ms	9×
YAPF logo 512×512	60.5 KB	89.3 KB	148 %	1.25 ms	0.25 ms	5×

Smaller than PNG on UI-style images, 20–50 % larger on photos and textures (the price of having no entropy coder), and 5–12× faster to decode. More in examples/ — and run yapf compare on your own images.

---

C / C++

The library is two files. Add them to your project:

yapf.h    — public API and the complete format specification
yapf.c    — encoder and decoder

#include "yapf.h"

yapf_image_t *img = yapf_load("texture.yapf");      /* or yapf_load_memory(buf, len) */
if (img) {
    /* img->pixels: width × height × channels bytes, rows top to bottom */
    upload_texture(img->width, img->height, img->channels, img->pixels);
    yapf_free(img);
}

yapf_image_t out = { .width = w, .height = h, .channels = 4,
                     .mip_levels = 1, .pixels = rgba, .flags = YAPF_FLAG_SRGB };
yapf_save("out.yapf", &out);

Compile with any C99 compiler (-pthread on Unix for the _mt loaders):

cc -O2 -std=c99 main.c yapf.c -o app -pthread

SIMD is on by default: SSSE3 on x86 / x64 (chosen at run time with CPUID, so binaries still run on older CPUs) and NEON on ARM64, with GCC, Clang, MSVC and MinGW.

Define	Effect
-DYAPF_NO_SIMD	plain C only
-DYAPF_NO_THREADS	no threads; the _mt loaders decode on the calling thread, and only the C standard library is needed

API Reference

Types

typedef struct {
    uint32_t  width;       /* pixels per row                             */
    uint32_t  height;      /* rows                                       */
    uint8_t   channels;    /* 1 gray / 2 gray+alpha / 3 RGB / 4 RGBA     */
    uint8_t   gpu_format;  /* YAPF_GPU_* hint                            */
    uint8_t   flags;       /* YAPF_FLAG_* (premultiplied alpha, sRGB)    */
    uint8_t   mip_levels;  /* 1 = base only, up to 16                    */
    uint8_t  *pixels;      /* base level, malloc-owned                   */
    uint8_t **mips;        /* mips[0] == pixels, mips[1..] smaller levels */
} yapf_image_t;

pixels is a tightly packed, row-major byte array of width × height × channels bytes. Mip level m is max(1, width >> m) × max(1, height >> m). The layout per pixel matches the channels value:

channels	Layout per pixel
1	Y
2	Y A
3	R G B
4	R G B A

Functions

yapf_image_t *yapf_load(const char *filename);
yapf_image_t *yapf_load_memory(const void *buffer, size_t size);

Decode a YAPF file from disk or from memory (e.g. a game archive). Return a heap-allocated image on success, NULL on any failure (file not found, invalid or corrupt data, out of memory). Corrupt input is rejected, never read out of bounds.

yapf_image_t *yapf_load_mt(const char *filename, int threads);
yapf_image_t *yapf_load_memory_mt(const void *buffer, size_t size, int threads);

Same, decoding tiles on up to threads threads (0 = one per CPU core). The output is identical to the single-threaded loaders.

int yapf_save(const char *filename, const yapf_image_t *img);

Encodes an image to disk. Writes img->mip_levels levels; img->mips may be NULL when mip_levels == 1. Returns YAPF_OK (0) on success or a negative YAPF_ERR_* code on failure.

int  yapf_encode(const yapf_image_t *img, void **out_data, size_t *out_size);
void yapf_free_buffer(void *data);

Encodes into memory: *out_data receives a complete .yapf file of *out_size bytes, to be released with yapf_free_buffer().

void yapf_free(yapf_image_t *img);

Releases memory returned by the loaders. Passing NULL is safe.

Textures (version 2)

yapf_texture_t *yapf_tex_load(const char *file, int flags, int threads, int *err);
yapf_texture_t *yapf_tex_load_memory(const void *buf, size_t size, int flags, int threads, int *err);
int  yapf_tex_save(const char *file, const yapf_texture_t *t, int flags);
int  yapf_tex_encode(const yapf_texture_t *t, int flags, void **out, size_t *size);
yapf_texture_t *yapf_tex_create(const yapf_texture_t *desc);   /* zeroed storage */
void yapf_tex_free(yapf_texture_t *t);

A texture describes its shape (width, height, depth, layers, frames, cube / volume, mips), storage (pixels with u8 / u16 / half / float samples, or GPU blocks), meaning (channel map, linear mask, normal convention, primaries, transfer, mip filter, alpha reference) and holds data[s] per surface s = yapf_tex_surface(t, frame, layer, face, mip). Load flag YAPF_LOAD_TRUST skips checksum checks; encode flag YAPF_ENC_ENTROPY allows the entropy stage.

yapf_texture_t *yapf_tex_open(const void *buf, size_t avail, int flags, int *err);
uint64_t yapf_tex_stream_bytes(const yapf_texture_t *t, uint32_t first_mip);
int  yapf_tex_decode_surface(yapf_texture_t *t, const void *buf, size_t avail, uint32_t s, int threads);
int  yapf_tex_decode_tile(const yapf_texture_t *t, const void *buf, size_t avail, uint32_t s,
                          uint32_t tx, uint32_t ty, uint32_t slice, void *out, size_t stride,
                          uint32_t *x, uint32_t *y, uint32_t *w, uint32_t *h);

Streaming and virtual texturing: open a file from its first bytes, ask how many bytes hold mip m and smaller, decode surfaces as their bytes arrive (YAPF_ERR_TRUNCATED until then), or decode single tiles.

int yapf_tex_build_mips(yapf_texture_t *t, uint32_t levels, uint8_t alpha_ref);
int yapf_tex_compress(const yapf_texture_t *src, uint8_t format, yapf_texture_t **out);
int yapf_blocks_decode(uint8_t format, const void *blocks, uint32_t w, uint32_t h, void *out);
int yapf_tex_add_note(yapf_texture_t *t, const char tag[4], const void *data, uint32_t size);
const yapf_note_t *yapf_tex_find_note(const yapf_texture_t *t, const char tag[4]);

Mip chains (linear-light sRGB, optional coverage preservation), the built-in BCn encoders, BC1–BC7 decoding for previews, and notes. The complete list, with sizes and helpers, is in yapf.h.

Return codes

Constant	Value	Meaning
YAPF_OK	0	Success
YAPF_ERR_INVALID	–1	NULL or logically invalid argument
YAPF_ERR_IO	–2	File could not be opened or written
YAPF_ERR_OOM	–3	Out of memory
YAPF_ERR_CORRUPT	–4	Corrupt or not a YAPF file
YAPF_ERR_UNSUPPORTED	–5	Needs a feature this decoder lacks
YAPF_ERR_CHECKSUM	–6	A checksum does not match (v2)
YAPF_ERR_TRUNCATED	–7	The data is not (yet) all there

Channel constants

YAPF_CHANNELS_GRAY        1
YAPF_CHANNELS_GRAY_ALPHA  2
YAPF_CHANNELS_RGB         3
YAPF_CHANNELS_RGBA        4

---

Building

System	Command	Output
Linux, macOS, BSD	make · make test · sudo make install	build/: libyapf.a, libyapf.so / .dylib, yapf
NixOS / Nix	nix-build (or nix-env -if .) — see extensions/linux	result/: bin/yapf, lib/, include/, MIME type, thumbnailer
Linux, macOS (LLVM)	./build.sh	dist/<platform>/ incl. macOS universal binaries
Windows (LLVM)	build.bat	dist\windows-x64\, dist\windows-arm64\: yapf.dll, yapf.lib, libyapf.a, yapf.exe
Windows (Visual Studio)	cl /O2 /c yapf.c and add yapf.obj to your project
All platforms at once	tools/build_release.sh (needs zig)	dist/release/: every download on the Releases page

The command-line tool is tools/yapf_cli.c + yapf.c; it uses the public domain stb headers in third_party/.

---

How it works

1. Two colour transforms — per 64×64 tile, YCoCg-R or subtract-green (G, R−G, B−G), whichever codes smaller. Both run mod 256, so every plane stays one byte per sample.
2. Twelve row filters — per band of 8 rows: Left, Gradient (L + A − D), Up-right, Up-average, or the row 1–8 rows above (catches repeating textures and dither patterns). None depends on the previous sample by more than one add, and the vertical ones decode as plain vector code.
3. Groups of 8 — residuals are packed in groups of 8 at the group's bit width: 8 values × w bits = exactly w bytes, so every group is byte aligned and unpacks with one 64-bit load (or one SIMD shuffle).
4. 4-bit control codes — one per group, or one per run of up to 64 groups repeating the previous width; the encoder picks the cheapest sequence with a small dynamic program.
5. Validate first, then run — all sizes in a tile are checked before the hot loops start, so the loops need no bounds checks. Corrupt files are rejected, never read out of bounds.
6. Constant and stored tiles — a plane that is constant in a tile costs one byte; a tile that would not shrink is stored raw (noise never grows).
7. Independent tiles — any tile can be decoded on its own, in any order, on any thread.

The version 1 specification is the comment at the top of yapf.h; version 2 (textures) is specified in docs/FORMAT.md. Version 1:

HEADER (20 bytes)   "YAPF", version 1, channels, GPU hint, mip count,
                    flags, 3 reserved, width u32, height u32
MIP INDEX           per level: offset u32, size u32
LEVEL DATA          tile_size u32 × tile_count, then the tiles
TILE (64×64)        header byte: stored raw, or constant-plane mask + colour transform
                    per plane: filter byte per 8 rows, control nibbles,
                               w bytes per group of 8 values

---

Testing

make test                      # C suite: 2631 checks, plus JavaScript if node is installed
node bindings/js/test.js       # JavaScript: round-trips, sample file, corrupt input

What has been verified:

• exact round-trips for 1–4 channels, odd sizes from 1×1 up, mip chains, single- and multi-threaded decoding, in-memory and file encoding;
• the same tests pass on Linux x64, Windows x64 (run under Wine) and Linux ARM64 (NEON, run under qemu), with and without SIMD, compiled as C and as C++, and under AddressSanitizer, UndefinedBehaviorSanitizer and ThreadSanitizer;
• thousands of corrupted and truncated files are rejected without crashing;
• version 2: lossless round-trips of u8 / u16 / half / float pixels with 1–4 channels, with and without entropy coding; cubes, cube arrays, arrays, volumes and animations with notes and every description field; GPU block payloads; checksum, needs-bit and truncation handling; streaming from file prefixes and single-tile decoding; BCn encoder quality; linear-light and coverage-preserving mips; fuzzing of entropy-coded and packed-block files;
• the BC1–BC7 decoders were checked against an independent reference decoder (bcdec) on 300,000 random blocks: BC6H and BC7 bit-exact, BC1–BC5 within the ±1 rounding the formats allow (a development check, not part of make test);
• every language binding runs its example; JavaScript, Python, C++, C#, Rust, Go, Java and Lua re-encode the sample to byte-identical output;
• the GIMP plugin in GIMP 3.0.8 and the Blender extension in Blender 5.1;
• the KDE thumbnailer through KDE's own plugin loader.

Not yet verified: running on a Mac (macOS builds are cross-compiled only), Windows on ARM, and running inside Photoshop, VS Code, Figma or a live Dolphin / GNOME file manager. The version 2 code has been run on Linux x64 (with and without SIMD, as C and C++, under ASan, UBSan and TSan); its NEON path is compiled for ARM64 but has not been run yet, and the Windows builds have not been rerun. ETC2 / EAC / ASTC blocks are stored and loaded, but there is no built-in encoder or preview decoder for them.

---

Project layout

yapf.h, yapf.c          the library (v1 spec in yapf.h)
docs/FORMAT.md          the full format specification, version 2
tools/yapf_cli.c        the `yapf` command-line tool
tools/build_release.sh  builds every download for all platforms
tests/test_yapf.c       C test suite
examples/               speed and size comparison, C example
docs/LANGUAGES.md       using YAPF from other languages
bindings/               cpp, python, csharp, rust, go, java, zig, lua, js
extensions/gimp/        GIMP 3 plugin
extensions/blender/     Blender extension
extensions/kde/         Dolphin thumbnail plugin (KF6)
extensions/linux/       MIME type and freedesktop thumbnailer
extensions/vscode/      VS Code preview
extensions/figma/       Figma plugin
extensions/photoshop/   Photoshop scripts
third_party/            stb_image / stb_image_write (public domain, CLI only)
other/YAPF.YAPF         sample image (the YAPF logo)

---

License

Apache License 2.0 — see LICENSE. third_party/ contains stb_image and stb_image_write by Sean Barrett (public domain / MIT).