ShootFlow develops camera raw files into 16-bit TIFF masters without going through Adobe's DNG Converter. Its Nikon reader was checked on 18 NEFs from 16 bodies, every sensor value equal to LibRaw's. The to-do list said a Coolpix .NRW reader was waiting for a public sample file.
There were samples all along. raw.pixls.us (CC0) has NRW files from twelve Coolpix bodies. Running them through the same LibRaw comparison before changing anything:
| body | sensor values wrong |
|---|---|
| P6000 | 12.8 M of 13.5 M |
| P7000 | 9.9 M of 10.0 M |
| P1000 | 15.8 M of 16.0 M |
| P330, P340, P7700, P7800 | all but a few hundred of 12.2 M |
| B700 | 20.1 M |
E8800 (an older Coolpix .NEF) | 8.0 M |
Nothing raised an error. A Coolpix raw is a Nikon TIFF with a colour-filter image in a SubIFD, which is everything the NEF reader checks for, so it read these files as if a D-series body had written them. The README even said Coolpix files were "refused". They weren't. Since the card copy started keeping .NRW files two days ago, a Coolpix shooter would have got a confident develop of static.
Five layouts and no tag that names them
The compacts store their sensor data in ways the D-series bodies don't:
- 12-bit values packed MSB first, as in an uncompressed D200 file (P340, E8800). The P340's raw IFD says compression 34713, Nikon's Huffman coding, but the data isn't Huffman-coded at all.
- The same packing inside little-endian 32-bit words (P6000, P7000, P7100, P950, P1000, B700). Each four bytes have to be reversed before unpacking.
- One 16-bit word per sensel, little-endian (A1000) or big-endian with the 12-bit value in the top bits (P330, P7700, P7800).
- Even rows first, then odd rows (E8800).
dcraw and LibRaw pick the layout from per-model tables. ShootFlow uses the data instead. The strips' byte count tells packed from 16-bit. For the byte order, the reader decodes a band of eight rows from the middle of the frame both ways. Done right, sensels of the same colour two columns apart differ by a little, because it's a picture. Done wrong, they differ by a lot, because it's noise. The other order has to be at least twice as smooth to win, so ordinary NEFs keep their usual order. That held on all seven D-series files from the original check, including the uncompressed D200.
The smaller faults
- Black level. Seven bodies came out at 0. Three of them (P950, P1000, B700) store the level in a maker-note block that starts with the text
NRW 0104. The four 4032-pixel-wide bodies store no level at all and use LibRaw's per-model 200. In those four files no sensel reads below 200 (the lowest are 201 to 217). - White balance. Four bodies have no as-shot white-balance tag (0x000C). The same maker-note block carries the balance, with red and blue stored at half value.
- The P6000's colour-filter tag is wrong. It says GBRG. In the sensels, the sites at (0,1) and (1,0) average 709.9 and 710.0 with a correlation of 0.996, while the other diagonal's two sites average 507 and 500. Green is on the first diagonal, so the layout is RGGB, which is also how LibRaw reads it. Where the tag and the sensels disagree, the sensels win.
Where it ended
All 13 Coolpix files (twelve NRW and E-series bodies plus the Coolpix A, which was already right) now match LibRaw 0.22 in every sensor value, black level, white balance and colour-filter pattern, on both the C and Python decoders. One honest footnote: LibRaw shifts the P330's 16-bit words down four bits but keeps the P7700's and P7800's at full size, though all three store the same thing. ShootFlow keeps all three at full size, so the P330 is compared at LibRaw's scale. The measurements are in eval/nikon-nrw-reader-2026-11-18.md.
What's still missing is colour. Like every NEF, an NRW needs the camera's .dcp profile, which Adobe's free DNG Converter installs. That profile is the one step left before a Coolpix frame develops correctly.
The lesson that carries over: "no public sample yet" was a note from an earlier run, and nobody had checked it since. It took one directory listing to disprove.