From https://discuss.pixls.us/t/highlight-reconstruction-hard-edges/55378/11
Make this suitable for the manual:
The “visualize highlights” tool is not optimal if you want to inspect where data are clipped.
Bildschirmfoto vom 2026-01-10 06-05-57
Bildschirmfoto vom 2026-01-10 06-05-57931×431 363 KB
This (the Highlight Reconstruction module's mask button) gives much better information, it also respects a) the chosen threshold and b) the used algorithm.
2. You will immediately see the “irregular” pattern. We see that regularly when the given whitepoint (dt gets that from exif data) is not perfect.
3. In some cases (some sonys for example are notorious for that) the exif data seem to be wrong. Also with 10bit sensors this can be seen more often.
4. When you have problems with highlights reconstruction and a) see such patterns or clearly have a magenta cast but the indicator does not show clipping you should immedately decrease the threshold value.
5. A correct threshold value will make the highlights reconstruction stepping in presenting better data to the demosaicer.
6. After selecting a proper threshold you might opt for “segmentation” and modify the candidating. If there are large areas with sensor data blown out like here, segmentation is almost always the best algorithm as it does not only correct for local data but checks for valid colors at the borders of the segments. Remember - a) if the threshold is bad the segmentation analysis will be less stable b) there is no OpenCL code for segmentation so there is a performance penalty.
7. About the demosaicers. If you have such “patterns”, the demosaicers can handle this not equally well. You can find other examples in snow fields for example. Again it’s about local non-linear data. VNG4 handles those cases better than RCD, so if you chose a dual demosaicing mode, in such flat areas we chose VNG over RCD by mixing.
8. Just a reminder, the highlights module provides a raster mask presenting the blown-out parts. I very hard cases you can use it to desaturate for example.
From https://discuss.pixls.us/t/highlight-reconstruction-hard-edges/55378/11
Make this suitable for the manual:
The “visualize highlights” tool is not optimal if you want to inspect where data are clipped.
Bildschirmfoto vom 2026-01-10 06-05-57
Bildschirmfoto vom 2026-01-10 06-05-57931×431 363 KB
This (the Highlight Reconstruction module's mask button) gives much better information, it also respects a) the chosen threshold and b) the used algorithm.
2. You will immediately see the “irregular” pattern. We see that regularly when the given whitepoint (dt gets that from exif data) is not perfect.
3. In some cases (some sonys for example are notorious for that) the exif data seem to be wrong. Also with 10bit sensors this can be seen more often.
4. When you have problems with highlights reconstruction and a) see such patterns or clearly have a magenta cast but the indicator does not show clipping you should immedately decrease the threshold value.
5. A correct threshold value will make the highlights reconstruction stepping in presenting better data to the demosaicer.
6. After selecting a proper threshold you might opt for “segmentation” and modify the candidating. If there are large areas with sensor data blown out like here, segmentation is almost always the best algorithm as it does not only correct for local data but checks for valid colors at the borders of the segments. Remember - a) if the threshold is bad the segmentation analysis will be less stable b) there is no OpenCL code for segmentation so there is a performance penalty.
7. About the demosaicers. If you have such “patterns”, the demosaicers can handle this not equally well. You can find other examples in snow fields for example. Again it’s about local non-linear data. VNG4 handles those cases better than RCD, so if you chose a dual demosaicing mode, in such flat areas we chose VNG over RCD by mixing.
8. Just a reminder, the highlights module provides a raster mask presenting the blown-out parts. I very hard cases you can use it to desaturate for example.