Meta offers developers open source and a dataset of 180,000 images for the AI project Animated Drawings, which will help animate ordinary drawings.
The web version of the tool was first released in 2021 by the Fundamental AI Research (FAIR) team. Users can upload a drawing of a humanoid character or choose one of the figures available in the demo. When using the author’s doodle, an offer to consent to the use of the drawing for training the model appears – but you can test the tool and confirm this request.
Next, you need to select certain points on the drawn character – you can adjust them using a pen and an eraser. After confirmation, you will be provided with an animated version of the sketch. There are several available categories to choose from: simple walking, dance moves, jumping and funny animations.
Animated Drawings uses object detection models, pose estimation models, and image segmentation techniques to create a digital version of the drawing. The tool then turns to traditional computer graphics to deform and animate the image.
The demo version has been available for several months, and during that time users have granted Meta permission to use over 1.6 million images for educational purposes. Some uploaded company logos, anime characters, photos of fish or stuffed animals – despite the fact that the tool only works with humanoid characters in drawings.
Course
Excel for business
Meta says there’s been a lot of interest in the drawing-to-animation tool, and requests are already coming in to expand the toolset — including adding sound effects or text overlays. The company has developed an open-source and open-dataset project.
“Releasing the open-source model gives developers a starting point to expand the project, fostering a culture of innovation and collaboration in the community,” Meta writes in its blog.
- We will remind that in January, Meta released its AI language model LLaMA – as an open source package, which can be accessed by members of the AI community. However, a week after the company began accepting requests, a torrent to download the language model appeared on 4chan and later spread to other communities.
- Meta also created SAM, an AI image segmentation model that can respond to cues (text or cursor-pointed) to identify certain objects in an image or video, even if it hasn’t encountered them during training.