Compression · Phidelity
Keep every measurement, in less space.
Phidelity compresses 8 to 16-bit sensor images and video, losslessly or with a guaranteed maximum error. Every file checks itself: an error is always detected, never returned silently.
Infrared cameras of the MAST tokamak: 4.3 times smaller than raw, versus 1.9 times for the archive’s current storage.
Against FFV1 and JPEG-LS, lossless; −36 % against JPEG XL at maximum effort.
Smaller than the reference format of every domain; smaller than JPEG XL e9, or equal, in 6 out of 10.
Than JPEG XL at maximum effort, for files of similar size; the light mode, 56 to 190 times faster, also decodes 4 to 5 times faster.
Lossless, domain by domain: size and time
Size: Phidelity difference, negative = smaller. Time: ×10 = 10 times faster, ×0.8 = slightly slower. Against JPEG XL at maximum effort (e9), Phidelity’s compact mode, the smallest; against JPEG-LS, its light mode, which targets the same speed (* colour photos: compact mode, the light mode is grey only). Against the domain format: compact mode. Public data, every file decoded and checked; PC, one core, time of the call only.
| Against JPEG XL e9 — Phidelity compact | Against JPEG-LS — Phidelity light | ||||||
|---|---|---|---|---|---|---|---|
| Domain | Size | Encoding | Decoding | Size | Encoding | Decoding | Against the domain format |
| Infrared video, fixed cameras (LTIR) | −36 % | ×16 | ×0.6 | −35 % | ×1.2 | ×1.4 | −37 % (FFV1) |
| Radiometric infrared (ARCO) | equal | ×10 | ×0.8 | −14 % | ×0.85 | ×1.15 | −40 % (PNG) |
| Hyperspectral (AVIRIS, ROSIS) | −22 % | ×26 | ×0.8 | −21 % | ×1.4 | ×1.6 | −6 % (CCSDS 123) |
| Radiography (CR) | −7 % | ×15 | ×0.8 | −10 % | ×1.0 | ×1.2 | −13 % (JPEG 2000) |
| Raw side-scan sonar (EdgeTech, Klein) | −3 % | ×35 | ×0.9 | −24 % | ×1.1 | ×1.6 | −23 % (deflate) |
| 8-bit colour photos | equal | ×11 | ×1.6 | −7 % * | ×0.1 * | ×0.4 * | −2 % (WebP) |
| CT | +4 % | ×8 | ×0.8 | −44 % | ×0.95 | ×1.45 | −47 % (JPEG 2000) |
| MRI | +4 % | ×9 | ×0.8 | −52 % | ×1.0 | ×1.3 | −54 % (JPEG 2000) |
| Ultrasound | +11 % | ×6.5 | ×0.8 | +3 % | ×0.75 | ×0.8 | −19 % (JPEG 2000) |
| Gridded bathymetry (NOAA) | +2 % | ×11 | ×0.8 | −1 % | ×0.8 | ×1.0 | −40 % (LERC) |
Detailed results, domain by domain
A one-page PDF sheet per domain: the data used, every codec compared (lossless and bounded-error size, speed, memory), embedded figures where measured, and the limits.
What sets it apart
A file that checks itself
Every file holds what it takes to check its own decoding. A damaged file is refused; a wrong decode never goes unnoticed. Format 1.7 frozen: every version must decode the reference files and reproduce them byte for byte. Test campaign on three series of public datasets (infrared, sonar, photos, hyperspectral, CT and MRI): no failure.
A guaranteed error, not an estimate
In bounded-error mode no sample moves more than k from the original, and the “no measurement” pixels (0) of depth or bathymetry maps stay exact.
Built for streams
Sequences from fixed cameras are its best ground: −37 % against FFV1 on infrared video, −35 % on tokamak movies.
From server racks to microcontrollers
Light mode faster than JPEG-LS, or within 5 %, in 8 domains out of 11 for encoding and faster in 10 out of 11 for decoding, and since format 1.5 within 4 % of the compact mode on CT, MRI and sonar; on Cortex-M7, 73 KB of code (encoder and decoder, grey, video, colour) and 40 to 90 KB of working memory; byte-identical streams on PC (SSE, AVX2) and ARM (NEON, Cortex-M7).
Where it pays
Fusion and large instruments
Diagnostic infrared movies: on MAST, −35 % against the best FFV1 and ×2.3 over current storage.
Fixed thermography
Site monitoring, processes, test benches: the full 16-bit stream kept, half the size of deflate.
Sonar and hydrography
Raw side-scan sonar 23 % smaller than the original archives, lossless.
Space and hyperspectral
Hyperspectral cubes 6 % smaller than the CCSDS 123 space standard.
3D scanners and depth
Depth maps: on par with JPEG XL at maximum effort, 18 times faster, 2.5 times smaller than FFV1.
Radiography
−13 % against JPEG 2000 and −7 % against JPEG XL.
What it does not do better
- On CT, MRI and ultrasound, JPEG XL at maximum effort stays 4 to 11 % ahead; on bathymetry, 2 % — while encoding 6 to 11 times more slowly.
- For decoding, the compact mode is 13 to 38 % slower than JPEG XL (except on colour photos); the light mode decodes 4 to 5 times faster.
- On depth maps, the light mode (×14.7) stays behind the compact mode (×24.9).
- Embedded figures are instruction counts in emulation, not yet timings on a board.
Floating-point data rather than images?
Phidelity Sci compresses simulation and measurement fields in single or double precision with a guaranteed maximum error: at equal quality, 9 to 58 % more compact than SZ3, QoZ and HPEZ on cosmology, weather, climate and scientific imaging.
Try it on your data
- You send one representative file or movie, raw or in your archive format.
- Within a week you receive the measurements: size, speed, memory, bit-exact check.
- Under NDA if needed; nothing is published.