If fine lines become colored or maze-like after demosaicing, compare algorithms on the same crop before choosing one; do not assume a switch can restore detail that the sensor did not record. In darktable 5.6, demosaicing reconstructs missing color channels from neighboring photosites. The relevant choices depend on whether the camera has a Bayer or X-Trans sensor, and a convincing-looking result is not proof that every pattern is genuine.
Identify the sensor and compare one crop
- Confirm the camera’s sensor type from the camera model or a trustworthy sensor specification. Then check the demosaic module’s available choices and compare only the candidates supported for that confirmed type. Do not treat the candidate list itself as independent confirmation, and do not infer the sensor type from filter color or visual appearance.
- Choose one crop containing the colored lines, a nearby genuine edge, and, if possible, a less detailed patch. Keep the crop position and zoom fixed while comparing. Do not compare one algorithm at fit-to-screen with another at pixel level.
- For Bayer, compare RCD, AMaZE, and LMMSE. The manual describes RCD and AMaZE as reconstructing high-frequency detail, with possible color overshoot or added noise in low-contrast areas; it identifies LMMSE as more suited to high-ISO or noisy images and as useful for moiré seen with other algorithms. These are documented tendencies, not a guaranteed ranking for your file.
- For X-Trans, compare Markesteijn 1-pass with Markesteijn 3-pass if you can wait for slower processing. The manual names 1-pass as the default and says 3-pass can offer slightly better quality at a much slower speed. Do not include Bayer candidates in this comparison.
- Record whether the colored fringe, maze-like texture, and continuity of the real edge change. Prefer the result that best preserves the edge without introducing more conspicuous false color; if differences are negligible, keep the current result rather than claiming a universal winner.
The darktable 5.6 demosaic manual explains why interpolation can produce moiré and false maze patterns, sometimes visible only at close zoom. It also notes that existing moiré- or maze-like structure in raw data may be handled differently by algorithms. The comparison above is an editorial diagnosis method, not a special test mode or a promise about what any control will recover.
Acceptance worksheet: record what changed
Use one row per candidate. The Bayer rows below are a hypothetical example for an already confirmed Bayer sensor, not a universal comparison table, a measured result, or a darktable prediction. They illustrate how to keep observations comparable; they do not prescribe numeric settings. For a confirmed X-Trans sensor, use the same fields with an empty observation row such as: X-Trans, confirmed | Markesteijn 1-pass or 3-pass | [ISO/noise note] | [fine-edge observation] | [color/maze observation] | [decision note]. Fill in only what you observe for that RAW and crop.
| Sensor label | Candidate | ISO/noise note | Fine-edge observation | Color/maze observation | Decision note |
|---|---|---|---|---|---|
| Bayer, confirmed | RCD | Hypothetical ISO 3200; visible noise | Roof wire stays continuous | Magenta fringe remains | Keep as baseline |
| Bayer, confirmed | AMaZE | Same RAW and crop | Wire looks slightly more distinct | Hypothetical stronger color fringe | Detail gain does not settle choice |
| Bayer, confirmed | LMMSE | Same RAW and crop | Wire remains visible | Hypothetical fringe is less evident | Candidate to inspect at intended viewing size |
Keep the same crop and zoom for every row. Record only visible differences and the sensor label; do not infer that a specific ISO guarantees one algorithm’s result. If no candidate changes the pattern, note that too. This worksheet is for a single-algorithm comparison, not the dual-demosaic threshold or blending-mask workflow.
Limits and what to do when comparison fails
The comparison is inconclusive when the lines remain alike across the appropriate candidates, or when available raw-data evidence already contains a similar pattern. In that case, record the unresolved result rather than calling it recovered texture. The manual says photosite_color is intended for debugging raw photosite data, not ordinary image processing; use it only if you understand that diagnostic view, not as a finished color rendering.
If the comparison is obscured by noise, judge color artifacts and edge continuity separately, and include LMMSE only in the Bayer comparison. If the crop is too small or the display scale changes between views, repeat the comparison with a fixed crop and zoom before deciding. If no rendering is acceptable, retain the least misleading one and document the unresolved pattern; a new capture or a separate specialist workflow may be needed, depending on the subject and purpose. darktable’s manual does not promise that an algorithm switch restores detail absent from the sensor data.
Common questions
#### Should I always choose AMaZE for fine detail? No. The manual describes high-frequency reconstruction strengths, but also notes possible overshoots or added noise. Compare the same area rather than choosing by name.
#### Does high ISO mean I must use LMMSE? No. It is described as better suited to noisy, high-ISO images than AMaZE or RCD, not as mandatory or guaranteed to remove every artifact.
#### Is X-Trans Markesteijn 3-pass always better? The manual says it can give slightly better quality at a much slower processing cost. Check your crop; the documentation does not establish a universal visible improvement.
Sources and next step
This guide interprets public documentation and uses explicitly hypothetical acceptance examples, not application tests. Controls and availability can change; verify the cited manual against your installed version.
For work created in Imagild, carry the input requirements and acceptance checklist into the photo editor. The external software controls discussed above belong to their respective products.