NVIDIA’s Impressive ‘Content-Aware’ AI Tool Fills Up Broken Photos Realistically

[Click here to the video in this article]

Composite image by DesignTAXI. Background images via NVIDIA

NVIDIA has been pretty progressive with its research on AI-powered visual software. Thus far, it has seamlessly combined celebrities’ photos to create wholly new faces, produced fake videos that look utterly realistic, and successfully pasted photos’ visual styles onto separate batches.

Adobe recently demonstrated an improved, AI-assisted version of Photoshop’s ‘Content-Aware Fill’, which transfers a scene from an image to another without hiccups. NVIDIA’s new tool seems to have upped the ante, however—apart from backdrops, it’s managed to accurately stitch faces without them looking like botched jobs.

According to the team of researchers, led by Guilin Liu, the new technique “can reconstruct a corrupted image, one that has holes or is missing pixels.” It can also tweak images by filling new content in gaps.

What starkly sets NVIDIA’s version apart from Photoshop’s is that it doesn’t plop content from surrounding pixels—it recognizes what’s missing from pictures, such as an eyebrow or nose, and acts accordingly.

The group trained its machine with 55,116 random masks of streaks and holes. From there, the AI would study the missing pixels and learn what to fill them with. For extra accuracy, 24,866 other masks were added during the testing stage.

“Our model can robustly handle holes of any shape, size location, or distance from the image borders,” the scientists wrote in a paper. “Previous deep learning approaches have focused on rectangular regions located around the center of the image, and often rely on expensive post-processing. Further, our model gracefully handles holes of increasing size.”

See the impressive tool at work in the two-minute video below.

Composite image by DesignTAXI. Background images via NVIDIA

Composite image by DesignTAXI. Background images via NVIDIA

Composite image by DesignTAXI. Background images via NVIDIA

[via PetaPixel, video via NVIDIA, images via NVIDIA]

You might also like
Leave A Reply

Your email address will not be published.