[15:33 Wed,19.April 2023 by Thomas Richter]
Apple&s AI team has released a new method called FaceLit that can create a 3D model just from a normal photo of a face and use it to create arbitrary new views of the face. FaceLit can also change the illumination of the face after the fact, meaning that a virtual light source can be freely positioned, which then re-lights the face.
A face in various newly generated orientations and new illumination.
Unlike existing similar algorithms LINK, FaceLit can work with normal portrait photographs and does not require specially illuminated or camera orientation data or further manual post-processing.
FaceLit takes into account both the specific shape and materiality of a face to create photorealistic portraits that are both correct in terms of lighting and 3D view. FaceLit works with photos of real people as well as, for example, classic painted portraits like the example below:
FaceLit&s code is available for download at GitHub, and the research paper can be found here. We had previously reported on a research paper published at Siggraph 2019, which could retroactively change the lighting of faces in photos via neural network.
Blackmagic has integrated a similar feature into its new DaVinci Resolve 18.5, the Relight FX AI effect allows scenes to be retroactively lit with virtual light sources. However, it works with full videos and does not produce new perspective views.
It will be interesting to see for what purpose Apple develops FaceLit - an integration into the camera or photo app would be conceivable, which would then allow portrait shots to be optimized afterwards.
more infos at bei machinelearning.apple.com
deutsche Version dieser Seite: FaceLit: Neue Apple-KI kann Gesichter auf Fotos nachträglich neu beleuchten