In one embodiment, a method includes initializing latent codes respectively associated with times associated with frames in a training video of a scene captured by a camera. For each of the frames, a system (1) generates rendered pixel values for a set of pixels in the frame by querying NeRF using the latent code associated with the frame, a camera viewpoint associated with the frame, and ray directions associated with the set of pixels, and (2) updates the latent code associated with the frame and the NeRF based on comparisons between the rendered pixel values and original pixel values for the set of pixels. Once trained, the system renders output frames for an output video of the scene, wherein each output frame is rendered by querying the updated NeRF using one of the updated latent codes corresponding to a desired time associated with the output frame.