Actor-replacement system for videos
In one aspect, an example method includes (i) estimating, using a skeletal detection model, a pose of an original actor for each of multiple frames of a video; (ii) obtaining, for each of a plurality of the estimated poses, a respective image of a replacement actor; (iii) obtaining replacement speech in the replacement actor's voice that corresponds to speech of the original actor in the video; (iv) generating, using the estimated poses, the images of the replacement actor, and the replacement speech, synthetic frames corresponding to the multiple frames of the video that depict the replacement actor in place of the original actor, with the synthetic frames including facial expressions for the replacement actor that temporally align with the replacement speech; and (iv) combining the synthetic frames and the replacement speech so as to obtain a synthetic video that replaces the original actor with the replacement actor.
1. A computing system comprising a processor and a non-transitory computer-readable medium having stored thereon program instructions that upon execution by the processor, cause performance of a set of acts comprising:
estimating, using a skeletal detection model, a pose of an original actor for each of multiple frames of a video;
obtaining, for each of a plurality of the estimated poses of the original actor, a respective image of a modified version of the original actor;
generating, using the estimated poses and the images of the modified version of the original actor, synthetic frames corresponding to the multiple frames of the video that depict the modified version of the original actor in place of the original actor, wherein the synthetic frames depict the modified version of the original actor in respective poses that align with the estimated poses of the original actor in corresponding frames of the video, and wherein the synthetic frames comprise facial expressions for the modified version of the original actor that temporally align with corresponding speech, wherein generating the synthetic frames comprises, for a given frame of the multiple frames, inserting, using an object insertion model, an image of the modified version of the original actor into the given frame at a location indicated by the estimated pose of the original actor so as to obtain a modified frame, and wherein generating the synthetic frames further comprises providing the corresponding speech and the modified frame as input to a temporal generative adversarial network having an ensemble of discriminators; and
combining the synthetic frames and the corresponding speech so as to obtain a synthetic video that replaces the original actor with the modified version of the original actor.
2. The computing system of claim 1 , wherein obtaining the images of the modified version of the original actor comprises estimating a pose of the modified version of the original actor for each of multiple frames of a sample video of the modified version of the original actor.
3. The computing system of claim 1 , wherein obtaining the images of the modified version of the original actor comprises:
recording a sample video of the modified version of the original actor; and
identifying poses of the modified version of the original actor during the sample video using a motion capture system.
4. The computing system of claim 1 , wherein obtaining the images of the modified version of the original actor comprises rendering an image of the modified version of the original actor using a target pose, an image of the modified version of the original actor, and a pose rendering model.
5. The computing system of claim 1 , wherein a difference between the original actor and the modified version of the original actor is in accordance with a storyline change.
6. The computing system of claim 1 , wherein generating the synthetic frames further comprises, for the given frame, extracting the original actor from the given frame using the estimated poses of the original actor.
7. The computing system of claim 1 , wherein the ensemble of discriminators comprises a frame discriminator, a sequence discriminator, and a synchronization discriminator.
8. A method comprising:
estimating, by a computing system using a skeletal detection model, a pose of an original actor for each of multiple frames of a video;
obtaining, by the computing system for each of a plurality of the estimated poses of the original actor, a respective image of a modified version of the original actor;
generating, by the computing system using the estimated poses and the images of the modified version of the original actor, synthetic frames corresponding to the multiple frames of the video that depict the modified version of the original actor in place of the original actor, wherein the synthetic frames depict the modified version of the original actor in respective poses that align with the estimated poses of the original actor in corresponding frames of the video, and wherein the synthetic frames comprise facial expressions for the modified version of the original actor that temporally align with corresponding speech, wherein generating the synthetic frames comprises, for a given frame of the multiple frames, inserting, using an object insertion model, an image of the modified version of the original actor into the given frame at a location indicated by the estimated pose of the original actor so as to obtain a modified frame, and wherein generating the synthetic frames further comprises providing the corresponding speech and the modified frame as input to a temporal generative adversarial network having an ensemble of discriminators; and
combining the synthetic frames and the corresponding speech so as to obtain a synthetic video that replaces the original actor with the modified version of the original actor.
9. The method of claim 8 , wherein obtaining the images of the modified version of the original actor comprises estimating a pose of the modified version of the original actor for each of multiple frames of a sample video of the modified version of the original actor.
10. The method of claim 8 , wherein obtaining the images of the modified version of the original actor comprises:
recording a sample video of the modified version of the original actor; and
identifying poses of the modified version of the original actor during the sample video using a motion capture system.
11. The method of claim 8 , wherein obtaining the images of the modified version of the original actor comprises rendering an image of the modified version of the original actor using a target pose, an image of the modified version of the original actor, and a pose rendering model.
12. The method of claim 8 , wherein a difference between the original actor and the modified version of the original actor is in accordance with a storyline change.
13. The method of claim 8 , wherein generating the synthetic frames further comprises, for the given frame, extracting the original actor from the given frame using the estimated poses of the original actor.
14. The method of claim 8 , wherein the ensemble of discriminators comprises a frame discriminator, a sequence discriminator, and a synchronization discriminator.
15. A non-transitory computer-readable medium having stored thereon program instructions that upon execution by a computing system, cause performance of a set of acts comprising:
estimating, using a skeletal detection model, a pose of an original actor for each of multiple frames of a video;
obtaining, for each of a plurality of the estimated poses of the original actor, a respective image of a modified version of the original actor;
generating, using the estimated poses and the images of the modified version of the original actor, synthetic frames corresponding to the multiple frames of the video that depict the modified version of the original actor in place of the original actor, wherein the synthetic frames depict the modified version of the original actor in respective poses that align with the estimated poses of the original actor in corresponding frames of the video, and wherein the synthetic frames comprise facial expressions for the modified version of the original actor that temporally align with corresponding speech, wherein generating the synthetic frames comprises, for a given frame of the multiple frames, inserting, using an object insertion model, an image of the modified version of the original actor into the given frame at a location indicated by the estimated pose of the original actor so as to obtain a modified frame, and wherein generating the synthetic frames further comprises providing the corresponding speech and the modified frame as input to a temporal generative adversarial network having an ensemble of discriminators; and
combining the synthetic frames and the corresponding speech so as to obtain a synthetic video that replaces the original actor with the modified version of the original actor.
16. The non-transitory computer-readable medium of claim 15 , wherein obtaining the images of the modified version of the original actor comprises estimating a pose of the modified version of the original actor for each of multiple frames of a sample video of the modified version of the original actor.
17. The non-transitory computer-readable medium of claim 15 , wherein obtaining the images of the modified version of the original actor comprises:
recording a sample video of the modified version of the original actor; and
identifying poses of the modified version of the original actor during the sample video using a motion capture system.
18. The non-transitory computer-readable medium of claim 15 , wherein obtaining the images of the modified version of the original actor comprises rendering an image of the modified version of the original actor using a target pose, an image of the modified version of the original actor, and a pose rendering model.
19. The non-transitory computer-readable medium of claim 15 , wherein a difference between the original actor and the modified version of the original actor is in accordance with a storyline change.
20. The non-transitory computer-readable medium of claim 15 , wherein generating the synthetic frames further comprises, for the given frame, extracting the original actor from the given frame using the estimated poses of the original actor.