Two devices arrive on a shortlist. One compresses with JPEG XS, one with JPEG 2000. Both datasheets say visually lossless, both are intra-frame — every frame coded on its own, no group-of-pictures, no motion prediction — and a buyer reasonably concludes they are the same class of product with a different badge. They are not. They sit at opposite ends of the same trade, and knowing which end a room needs is most of the specification decision.
Latency
JPEG XS — standardised as ISO/IEC 21122, developed by the JPEG committee with ITU-T — is designed around line latency rather than frame latency. Published figures put encode and decode together at the order of thirty-two lines, which lands under a millisecond. Nothing waits for a whole frame to exist.
JPEG 2000 works on frames. Roughly one to one and a half of them, commonly quoted around twenty milliseconds and tuneable lower with effort. On a 60Hz signal that is a frame or so of delay through the encode-decode pair.
For a signage wall, a lecture overflow, a campus distribution ring, that difference is invisible. For a surgeon driving an instrument from a screen, a camera operator framing on a monitor, or an XR headset, it is the whole requirement.
Compression depth
Here the trade runs the other way, and this is the part that gets missed. JPEG XS holds visually lossless quality across a relatively shallow range — roughly 2:1 to 10:1, with newer profiles reaching further. JPEG 2000 keeps its quality far deeper into compression, comfortably at 15:1 and beyond.
That difference is not academic; it decides what network the system can run on. Depth is what puts a full 4K @ 60Hz 4:4:4 stream onto the gigabit infrastructure a building already has, instead of onto a ten-gigabit fabric somebody has to buy, install and qualify first. A codec that compresses less asks for more link, and the link is usually the expensive part of the room.
Silicon, and the shape of the box
The third difference rarely reaches a datasheet and shows up in every other way. JPEG XS is deliberately light: its entropy coding is simple enough that a baseline implementation needs no external memory. JPEG 2000's arithmetic coding is far heavier, and wants substantial external memory bandwidth to work at rate.
That arrives at the specifier as cost, as power draw, as heat in a ceiling void, and as whether the encoder fits behind a wall plate or needs a shelf. It is worth asking about, because it constrains where a device can physically go.
What to write in the tender
Not visually lossless. The phrase is true of both and therefore selects neither. Write the two numbers that actually decide it: the end-to-end latency budget in milliseconds, and the link rate the network can offer. The codec then follows from the constraints rather than from whichever brochure was read most recently.
Why our networked tier is JPEG 2000
Because for the enterprise and campus estate the binding constraint is almost always the network, not the frame. Those buildings have gigabit to the outlet and no appetite for a parallel ten-gigabit fabric, and a frame of latency between a lectern and an overflow screen costs nothing anybody can perceive. Compression depth is what makes that architecture possible, so that is the trade our JPEG2000 encoders and decoders take.
Where compression itself is the risk — pixel-accurate review, sub-frame hand-eye loops, measurement — the honest answer is not a lighter codec but no codec at all, which is what the announced SDVoE tier over ten gigabit is for. Two problems, two architectures, stated rather than blurred.
Sources: ISO/IEC 21122 (JPEG XS) as published by the JPEG committee; comparative latency, compression and implementation figures as published by the standards bodies and implementers at the time of writing. Figures for any specific model are in that model's specification.