Re:Comparing iPhone 13 vs. iPhone 14 for Astrophotography
by chongli on Sunday May 28, 2023 @09:13PM
(#36109502)
You're focusing on the "publicly available" part but it's irrelevant to my point. Samsung's camera software is literally doctoring people's photos with fake data from its ML models. From the perspective of photographic integrity, it does not matter whether they scraped their training set off public photography sites or if they hired a team of photographers to build a dataset from scratch. To the user the effect is the same: it is not really an authentic photograph that is being created, it's a hybrid derived from (to the user) unknown sources.
[ Parent
| on HN ]
Re:Comparing iPhone 13 vs. iPhone 14 for Astrophotography
by astrange on Monday May 29, 2023 @04:27PM
(#36117716)
All photos work like this - cameras don't know what color things are, the color you see is made up by the camera.
You've never taken an authentic photo on a digital camera.
But… they don't use ML models made from publicly available images.
Re:Comparing iPhone 13 vs. iPhone 14 for Astrophotography
by chongli on Monday May 29, 2023 @07:35PM
(#36119386)
There's a difference between ICC colour profiling (which is embedded in the RAW file and can be changed by the photographer without any degradation in image quality) and what Samsung is doing. The former can only affect the image colours globally (for example, allowing adjustment of white balance) whereas Samsung's changes are local, affecting image data in a profound way that can add hallucinated details which were not present in the original scene. One rather disturbing example is that the tool added teeth to baby photos!
Re:Comparing iPhone 13 vs. iPhone 14 for Astrophotography
by astrange on Tuesday May 30, 2023 @10:00PM
(#36133644)
White balance adjustments to camera raws don't generally use ICC profiles, they do whatever the combination of camera format and raw processor want.
Which can involve local adjustments, since you may want to process people differently from skies. So object recognition/segmentation models can definitely be involved, or what's essentially an upscaling model for better demosaicing. (That wouldn't be trained on public images though, unless they were the right format of camera raw.)