Google’s Imagen Video attempts to help video-generator turn into killer apps
The Trust Project is a worldwide group of news organizations working to establish transparency standards.
It didn’t take long for Google to respond to Make-a-Video from Meta. By using a text prompt, Imagen Video may produce a fantastic video. The results are a tremendous advance above the state of the art despite a number of drawbacks.
In comparison to Facebook’s Text-to-Video AI generator Make-a-Video, the results are noticeably better. However, this strategy also demanded more oversight. In contrast to Imagen Video, where the micro workers worked hard to annotate movies with written descriptions, Make-a-Scene used unlabeled videos for training.
Going into the specifics of the architecture is pointless; you should read about it in the article here. We can only confirm that 16 frames are first generated from the text embedding of the T5 encoder at a resolution of 48×24 with 3 frames per second, and that this is then upscaled by a number of diffusion models into the final movie of 128 frames at 1280×768 and 24 frames per second.
What is Imagen Video?
Imagen Video is a method for creating text-conditional videos based on a series of video diffusion models. Imagen Video produces high quality films from text prompts by combining a base video production model with a series of interlaced spatial and temporal video super-resolution models. Go over the design choices team made while scaling up the system as a high-definition text-to-video model, including the decision to v-parameterize diffusion models and the selection of fully convolutional temporal and spatial super-resolution models at specific resolutions. In addition, it validates and apply results from earlier work on diffusion-based image production to the context of video generation. Video models are then subjected to progressive distillation with classifier-free guidance for quick, high-quality sampling.
The Google research team claims that the system accepts a textual description and generates a 16-frame movie at three frames per second with a resolution of 24 by 48 pixels. The system scales and “predicts” the extra frames, creating a final video with 128 frames at 24 frames per second and 720p resolution (1280×768). There are 60 million image-text pairs and 14 million video-text pairs were used to train Imagen Video.
Imagen Video Samples
Even if merely because using AI to make video is quicker and less expensive, such technologies will undoubtedly be employed everywhere.
Interested in reading more? Here are some additional topics to check out:
Any data, text, or other content on this page is provided as general market information and not as investment advice. Past performance is not necessarily an indicator of future results.