CL 8519837 reported a JPEG XL
codestream's intensity_target as CLLI MaxCLL unconditionally.
But encoders write 10000 cd/m² (the PQ nominal peak) by default, and
255 means the field is absent from the codestream - neither value states
the content's peak. Typical PQ files (e.g. the
jpegxl.info HDR
test page) were suddenly tone mapped from 10000 instead of the
1000 cd/m² default and rendered dim. Follow-up change: only an
informative value (above 255 and below 10000) is reported as content peak.
Three JXL files hold the same flat 588 cd/m² BT.2100 PQ patch; only the
stated intensity_target differs. Each is drawn to an SDR canvas
and the mean of the red channel is printed.
| file | intensity_target | mean red |
|---|
| expectation | result |
|---|
broken M155..M157 pre-fix: the 10000 and 255 files tone map from their stated value - 10000 is dim, 255 is clipped - and the second expectation fails.
fixed With the follow-up fix (and in M154 and earlier): 10000 and 255 fall back to the 1000 cd/m² default and render identically; 600 is still honored as a deliberate peak.
Luca's jixel sample (P3 PQ, intensity_target 10000) rendered by
headless content shell builds of the same revision, with and without the fix.
Tone mapping to SDR from a claimed 10000 cd/m² peak dims the HDR
highlights - the neon tubes and letters - while the shadows barely move.
Crop mean brightness: broken 72.5, fixed 77.6. Full-frame RMSE 2.9%.
The difference is confined to the HDR highlights.
Note: an HDR-headroom capture would show no difference at all - with a 10000 cd/m² output target neither build tone maps. The regression is visible exactly where headroom is scarce: SDR displays and HDR displays with less headroom than the claimed peak.
Broken build: dim, crushed highlights. Fixed: tone mapped from the 1000 cd/m² default again.
From jpegxl.info HDR test page.
From jpegxl.info HDR test page.