See What Your Dog Would Look Like As Another Animal With This AI Tool
Dhir Acharya - Oct 31, 2019
Nvidia has a new AI-powered tool called GANimals that helps you see what your pet looks like as another animal.
- 4 Ways AI Could Change The Mobile Gaming Industry
- New ‘Deep Nostalgia’ AI Allow Users To Bring Old Photos To Life
- Pilots Passed Out Mid-Flight, AI Software Got The Aircraft Back Up Automatically
Nvidia has a new tool called GANimals that helps you see what your pet looks like as another animal. Earlier this year, the company introduced its new AI drawing tool dubbed GauGAN, which turns sketches into near photorealistic pictures. With this tool, users are required to indicate which parts of the image are mountains, trees, water, as well as other landmarks by selecting the right brush color. However, GANimal is 100 percent autonomous, which means users just need to upload a picture of their pet and the tool will produce a set of photorealistic pictures, all of which appear to feature the pet’s expressions.
As described by the researchers in a paper, the tool relies on FUNIT (Few-shot, Unsupervised Image-to-image Translation), an algorithm they have developed. The AI needs training on a lot of target images, with different light levels as well as camera angles, to generate accurate results that look genuinely like the source and targets have been merged properly.
Nevertheless, it takes a lot of time to put together such a large database and it limits the capabilities of the translation network. For example, if you train it turn a chicken into a turkey, it will only be able to do that.
FUNIT, on the other hand, can be trained with only some images of a target animal that it practices with repeatedly. As a result, it can generalize the translation required for merging two images. Once the training is done, FUNIT needs only one image of the target animal and the source, both of which can be 100% random and never analyzed or processed before, to do its job.
GANimals is available for use at AI Playground, but the results are now still in low resolution and you can’t likely use them for any other purposes. However, the researchers hope to improve the tool so that the algorithm can eventually swap faces without needing large databases with carefully curated images.
Comments
Sort by Newest | Popular