[14:57 Sun,10.August 2025 by Thomas Richter] |
Essentially, ;) Google Genie 3
An example video posted by a developer shows recordings of a drone flight, which after about 10 seconds is taken over by the user via Genie 3, who then continues to fly freely and is no longer bound to the flight path of the original recording. Genie 3 not only seamlessly generates new images of landscape aspects that were not yet visible, but also seems to be able to remember the details shown at the beginning of the video - you can see them from a different perspective after the virtual pilot has flown a loop.
This opens up fascinating new possibilities, because with Genie 3 you can, for example, "dive" into any painting or film and freely explore a specific scene and, to a lesser extent, its very special worlds, as demonstrated in the following example in all its strengths and weaknesses. Here is Edward Hopper&s famous painting "Nighthawks":
Or a classic painting like Jacques Louis David&s "The Death of Socrates":
And Genie 3 could also be interesting for filmmakers - used as a virtual camera - for example, by generating views of an already filmed scene from other perspectives, be it as a supplement to already made recordings or for experimenting when finding the optimal camera setting for a scene before the actual shooting. Genie 3 offers even more possibilities than  Runway&s Aleph, which allows the subsequent editing of video recordings and can also generate new views - but not freely navigable like Genie 3, but only roughly via text prompt. It is still unclear whether Genie 3 receives even more information for the (re)construction of the world through a longer start video and can represent it more faithfully. Apparently, however, a start video can be combined with a text prompt to influence the generation.
But one should of course keep in mind that Genie 3 is subject to the classic limitations of generative AIs: The world newly created on the basis of existing images is generated - according to probability, i.e. according to patterns that were extracted from the training material. This means that Genie 3 does not create exact reconstructions of (film) reality outside the frame, but only possible environments apart from what is visible in the source video. Therefore, in Genie 3 you cannot walk through real cities like New York or Venice in the example videos, but in an archetypal New York or Venice - which is hardly distinguishable from the original for those unfamiliar with the place. Google itself points out in the section on the limitations of the model: "Accurate representation of real places. Genie 3 is currently unable to simulate real places with perfect geographic accuracy."  more infos at bei deepmind.googledeutsche Version dieser Seite: Googles Genie 3 ermöglicht auch das Eintauchen in Filme und Bilder |
 |
|
|