Learn/Remade with Rust

rusty_tiff, rusty_image and rusty_icc: Same Bytes, Less Work

The 0.1.1 releases of rusty_tiff, rusty_image and rusty_icc cut copies and instructions with byte-identical output. Where each wins, and where each trails.

Signed by M·
A painted mountain path at morning where a courier walks with a slim leather satchel, the same bundle of sealed letters visible through its open flap
What changed in rusty_tiff, rusty_image and rusty_icc 0.1.1?

Speed, not output. Every 0.1.1 output is byte-identical to 0.1.0. The release cut memory copies and instructions: rusty_tiff LZW decode runs 42% fewer instructions on 8-bit files, and rusty_image decodes BMP-24 2.11 times faster than the image crate.

Three crates took their second step on 20 November 2026, from 0.1.0 to 0.1.1: rusty_tiff, rusty_image and rusty_icc. The work between those versions chased two costs, memory copies and instruction counts, under one rule: no output may change. The rule held. Every byte the new versions produce matches what the first versions produced. Below is what that bought, including the rows where another library still wins. All three live in the remade_ffmpeg_rs monorepo, with public cards for the TIFF codec, the small-format codecs and the color engine on Remade with Rust.

Trust is a speed claim printed beside its method. Security is three parsers that stayed pure Rust, with no unsafe, while they got faster. Incentive is a safe decoder that no longer asks anyone to accept a slow one. The NIST Secure Software Development Framework asks for memory-safe parsers, and this campaign was about making that choice cheaper.

How the 0.1.1 Numbers Were Measured

Two kinds of number appear here. "Ir" means instruction counts from callgrind, taken on 1536 by 1024 crops of a real photo. An instruction count does not drift with CPU temperature or a background task, so it shows whether a change removed work, even a small amount. Each percentage is the drop between the two versions.

Clock numbers are wall time from pinned, interleaved runs in ABBA order, 20 rounds each, so neither arm always goes first. They come from the harness of ArtCraft, the team behind the storytold creative suite, which benchmarks these crates. A clock figure marked "ours" came from our own runs. A ratio above 1 means the Remade crate is faster.

rusty_tiff Did Less Work per File

The archive codec carried the largest share of the campaign, on both decode and encode.

LZW, Uncompressed and G4 Decode

LZW decode dropped 42% in Ir on 8-bit files and 45% on 16-bit. Uncompressed decode dropped 40%. CCITT G4 decode, the black-and-white scan format, dropped 73%. On our clock, the final LZW change by itself took 8-bit decode from 0.24 to 0.12 seconds and 16-bit from 0.40 to 0.18 seconds. LZW decode already beat the tiff crate by 1.15–1.55× before that change. Uncompressed decode now ties tiff 0.11 and beats libtiff by 1.10–1.12×. G4 decode is 4–6× faster than tiff and 1.3–1.6× faster than libtiff.

Uncompressed, LZW, PackBits and Deflate Encode

Writing got the same treatment. Uncompressed encode fell 71% in Ir, LZW encode 49%, PackBits encode 67% and Deflate encode 25%. On the clock, PackBits encode is 3.9–4.2× faster than the tiff crate. Written files are byte-identical to what the first version wrote, so an archive saved by either version reads back the same.

Damaged TIFFs That Now Open

Speed was not the only thing 0.1.1 changed. On a set of 18,024 damaged TIFFs, 141 files had decoded under tiff 0.11 yet been refused by rusty_tiff's first release. Of those, 128 now decode. Another 806 files that both libraries refused now decode too, and nothing panicked across the set. Files that already decoded in rusty_tiff still produce the same bytes; fewer files are refused. A broken scan either opens or returns an error, and the process keeps running. On a Home Computer holding decades of scans, that separates one bad file from a crashed import.

Small Formats and Color Against image and LittleCMS

The other two crates were timed against libraries people already run: the image crate 0.25.10, and LittleCMS.

BMP, TGA, HDR and the QOI Jump Table

In Ir, QOI decode dropped 38% for RGB and 35% for RGBA, while QOI encode dropped 30% and 45%. BMP 8-bit palette decode fell 37%, and BMP-32 decode 33%. On the clock against image, BMP-24 decode is 2.11× faster, TGA RLE decode 2.06×, Radiance HDR encode 3.12× and BMP encode 2.35–2.58×. QOI RGB decode went from 0.82× to 1.07×. The old deficit came from the decoder's op dispatch, which compiled into a jump table that mispredicted 59% of the time.

Where the image Crate Still Leads

QOI encode remains behind: 0.81× for RGB and 0.89× for RGBA. Both figures predate the last two QOI changes, so they are the latest clock numbers available rather than a final verdict. They are published anyway. A speed table that lists only wins is an advertisement, and the EFF has argued for years that security claims need evidence a reader can check.

The Color Engine: Faster Than Before, Behind on 8-Bit

In Ir, the matrix-shaper path dropped 22%, with 12% of that coming from the last commit. 8-bit RGB to CMYK dropped 44%, and profile parse and encode dropped 66%. On the clock, the new version is 1% to 37% faster than the first across the transforms timed. Against LittleCMS it is 1.40× faster on 16-bit matrix-shaper and 1.14× faster on Lab to sRGB in float. LittleCMS still leads on 8-bit paths, because its fixed-point arithmetic would change the Rust engine's output, and the crate keeps its results. Earlier notes in the same spirit cover rusty_jpeg against zune-jpeg and rusty_vorbis on ArtCraft's file. The pipeline all of these feed is remade_ffmpeg_rs, and the full family sits in MATA's rebuild catalog.

FAQ

Short answers already in this article.

How were the rusty_tiff, rusty_image and rusty_icc 0.1.1 numbers measured?

Instruction counts are callgrind Ir on 1536 by 1024 crops of a real photo. Clock numbers are pinned, interleaved ABBA runs with 20 rounds, on the ArtCraft team's harness unless marked as ours.

How much faster is rusty_tiff 0.1.1 LZW decode?

LZW decode runs 42% fewer instructions on 8-bit files and 45% fewer on 16-bit. On our clock, the final change alone took 8-bit decode from 0.24 to 0.12 seconds and 16-bit from 0.40 to 0.18 seconds.

Did rusty_tiff 0.1.1 open more damaged TIFF files?

Yes. On an 18,024-file damaged-TIFF set, 128 of the 141 files that tiff 0.11 decoded but 0.1.0 refused now decode, and 806 files that both refused now decode. There were no panics.

Where does rusty_image 0.1.1 still trail the image crate?

QOI encode runs at 0.81 times the image crate's speed for RGB and 0.89 times for RGBA, measured before the last two QOI changes. Every other path timed in this note is at or ahead of image 0.25.10.

Why does LittleCMS still beat rusty_icc 0.1.1 on 8-bit transforms?

LittleCMS uses fixed-point arithmetic on its 8-bit paths. Copying it would change rusty_icc's output, so rusty_icc keeps its results and trails there, while leading on 16-bit matrix-shaper at 1.40 times and Lab to sRGB float at 1.14 times.