Three ways to read one frame

The same colour negative on the same OpticFilm 8200i, read three ways: upstream SANE's genesys GL128, our plustek8200i, October, and our plustek8200i, today. The first two are the images published in the original two-driver comparison on 9 October 2026; the third is the same frame rescanned on 31 October after a month of work on the backend.

Read the columns carefully. The genesys raw data from October no longer exists, so that column cannot be re-converted — it is the published JPEG, and so is the October column beside it. Both were converted by the negative pipeline as it stood in October, which has itself changed since. So the tone and colour shift between the October and the current column is mostly the converter, not the backend. What the backend alone did is isolated in section 4, where two of our own raws taken a day apart go through one and the same conversion.

The October and genesys timings are single runs, as published. Ours today is the spread of repeat runs taken for this page — wide enough that the honest reading is that the backend is no slower than it was, not that it got faster.

Full frame, 600 dpi

genesys16.4 s
ours, Oct16.7 s
ours, today16.6–21.6 s

Full frame, 3600 dpi

genesys54.4 s
ours, Oct34.6 s
ours, today30.4–34.5 s

Output size at 3600 dpi

genesys5184 × 3519
ours, then & now5184 × 3877

Infrared channel

genesys✗ none
ours, Oct✓ one scalar
ours, today✓ flat field

1 · Whole frame, 600 dpi

All three are drawn on one shared film coordinate system, so the same piece of film lies in the same place in every view. Dark areas are where a backend's scan window did not reach. White areas are inside a window but beyond the film, where unexposed film inverts to white.

genesys 9 Oct

Genesys scan

ours, October 9 Oct

Our October scan

ours, today 31 Oct

Our current scan

Zoom with the buttons above, then scroll: all three panels move together.

2 · Scan area

Margin from the scan window to the film edge, in millimetres. Positive means the window clears the film; negative means it cuts in. Our window is unchanged since October — nothing this month touched the geometry — so one row covers both of our columns.

600 dpileftrighttopbottom
ours both+0.13+0.80+1.65+1.44
genesys+1.19−0.25+0.25+0.38
3600 dpi
ours both+0.17+0.85+1.64+1.56
genesys+1.19+0.06+0.45+0.22
What this means. Genesys's window sits 1.06 mm to the left of ours, a constant at both resolutions. It leaves a 1.2 mm dark strip on the left and runs up against the film's right edge. Its top and bottom margins are only 0.2–0.45 mm and drift with resolution (ours stay at ~1.6 mm), so a small positioning error would clip the picture.

3 · Detail at 3600 dpi

The same three crops as the original page, now with the current scan beside them. No resampling happens at 3600 dpi in any version of our backend, so this is the one place where our two columns should look alike — and they do.

the face

the face, right-hand scan
the face, left-hand scan

a tulip

a tulip, right-hand scan
a tulip, left-hand scan

ground flowers

ground flowers, right-hand scan
ground flowers, left-hand scan
Sharpness is a wash; colour is not. All three resolve the same detail, and the dust specks line up across every column. What separates them is colour balance: genesys's raw data renders greener and brighter, which comes from each backend's white balance and exposure of the raw, not from lost detail. Between our own two columns the shift is the converter, not the scan.

4 · What the backend alone changed

Two of our own raws — 30 and 31 October, the holder untouched between them, matching to ncc 0.998 — put through one and the same conversion. Nothing here is the converter.

Vertical detail: a box average replaced by Lanczos-3

The sensor samples about 2.6× finer along the scan direction than across it, so every resolution below the hardware line rate arrives with surplus lines to squash away. That was a box average, which attenuates detail well below the new Nyquist and still passes enough above it to alias. Below: the wall's mortar courses at 600 dpi, where a squash of 2.48 applies, at 3× nearest-neighbour so the pixels are visible.

box average until 31 Oct

Wall detail, box average

Lanczos-3 today

Wall detail, Lanczos-3

Band power in the picture, exposure-matched, both axes, with a noise floor from repeat scans of each:

600 dpihorizontal 0.1–0.25horizontal >0.25vertical 0.1–0.25vertical >0.25
repeat scan, same build (noise floor)1.0021.0081.0041.009
Lanczos-3 ÷ box1.0110.9741.0821.061
+8.2% in the vertical mid-band, against a noise floor under 1%, and +6.1% above it. Horizontal moves by 1–3%, as it should: the change only touches the scan axis. Measured against SilverFast on the same frame, the vertical mid-band went from 6% behind to level (1.063 → 0.989). 1200 and 3600 dpi reproduce the old scans at ncc 0.996 and 0.991, and geometry and timing are unchanged.

The infrared plane: per-column flat field

The colour pass is shading corrected in hardware; the infrared pass cannot be, because DVDSET clips IR, so the whole calibration was skipped and one scalar stood in for it. Every sensor column kept its own offset. The two calibration strips needed to fix it were already being read and discarded. These figures are from a different negative — a visibly dusty, scratched frame, the one that exposed the problem; the tulip frame above was scanned without infrared. Spread of the infrared residual, after the lamp surface and the picture’s ghost in red are fitted out:

600 dpi, before the fixours, OctoberSilverFast, same frame
pixel to pixel along a row0.008140.00820
line to line down a column0.008060.00873
column means0.019990.00211
residual MAD0.022890.01067

The random noise matched SilverFast; the column pattern was 9.5× worse and ours alone. After the flat field, measured through the shipped 0.1.36 package:

600 dpibeforeafterSilverFast
column means0.019990.001630.00211
residual MAD0.022890.008100.01067
a 2–4 px speck stands at6.49 MAD18.11 MAD13.82 MAD
The random noise was never the problem. Because dust detection thresholds on a multiple of that spread, a small speck sat under the bar: at the default sensitivity the pipeline found 28% of 2–4 px specks and now finds 98%. End to end, defect visibility above a clean-site control falls by 97.8% where it fell by 59.2% before, with high-frequency detail elsewhere moving +0.08%. Genesys exposes no infrared channel at all, so none of this is available through it.