Excessive dynamic vary (HDR) pictures methods can precisely seize a scene’s full vary of brightness values — from these of its darkest shadows to its brightest gentle sources. Nevertheless, when the time involves view the ensuing HDR photographs, many shows are solely able to displaying a restricted vary of brightness ranges. This discrepancy signifies that HDR photographs are sometimes downgraded to suit throughout the bodily limitations of those normal dynamic vary (SDR) shows and the corresponding file codecs that retailer a restricted vary of values.
Nevertheless, an growing share of cell system and laptop shows are HDR-capable, displaying a larger vary of brightness ranges with brighter whites and darker blacks. Though a number of widespread HDR file codecs able to encoding this larger vary exist, HDR-only codecs can seem in another way relying on the show capabilities. To optimally present pictures on any show — irrespective of the dynamic vary — Android 14 launched the brand new Extremely HDR picture format in October 2023. Google Pixel 7 and newer units adopted the format, capturing and saving HDR images as Extremely HDR and affording a extra true-to-life rendition as in comparison with SDR codecs, with larger distinction and brighter highlights.
Though the format unlocked HDR picture show, its new metadata additionally elevated the complexity of picture modifying operations. Till now, should you utilized a extra advanced edit like Magic Eraser to an Extremely HDR {photograph} utilizing Google Images, your new edited picture would save in SDR, shedding its HDR rendition and the related brightness and distinction.
Right now, we introduce a brand new machine studying (ML) approach that allows advanced picture modifying for HDR images, together with these saved as Extremely HDR. On Pixel 8 and newer, our mannequin runs behind the scenes in Google Images to make sure that even should you apply advanced edits to HDR pictures, your new edited pictures stay HDR. Core to our approach is predicting the HDR picture metadata lacking after modifying, utilizing an ML mannequin educated on a big dataset of HDR pictures with full metadata.