Meta Releases New AI Models 'Emu Video' & 'Emu Edit', Allowing Easy Animation Generation From Text
On November 16th, social media giant Meta introduced its latest artificial intelligence (AI) models in an official article.
In fact, during the Meta Connect conference on September 28th, Meta unveiled the Emu (Expressive Media Universe) model. Now they have an update on it.
Firstly, "Emu Video" is capable of generating video edits based on text and image inputs. Secondly, "Emu Edit" specialises in image manipulation, significantly enhancing precision in image editing.
Effortless Video Creation
Meta divided the procedure into two stages: initially, producing images based on a text prompt, followed by generating video based on both the text and the generated image. This "factorised" or separated method for video creation enables efficient training of video generation models.
In contrast to previous methods relying on a complex series of models (such as five models for Make-A-Video), our cutting-edge technique is straightforward to execute. It utilises only two diffusion models, enabling the creation of 512x512 four-second videos at 16 frames per second.
Upgraded Product-Emu Edit
Additionally, although instructable image-generative models have progressed considerably in recent years, they encounter constraints in providing precise control.
Hence, we're introducing Emu Edit, an innovative method designed to streamline diverse image manipulation tasks and elevate the capabilities and accuracy in image editing.
In contrast to numerous current generative AI models, Emu Edit meticulously adheres to instructions, preserving the original pixels in the input image that are not relevant to the given instructions.
For example, when incorporating the text "Aloha!" onto a baseball cap, the cap's appearance should remain unaltered.
The Largest Database
To facilitate model training, Meta curated a dataset comprising 10 million synthesised samples. Each sample consists of an input image, a task description, and the corresponding targeted output image.
"We believe it’s the largest dataset of its kind to date."
Combine with FB and IG in Depth
Meta also expressed the idea ofspicing up your Instagram posts with animated static photos or creating completely fresh content.This is expected to be deeply integrated with its subsidiaries, Facebook and Instagram, in the future. However, the specific timing has not been officially announced by the authorities yet.