Vggface2-hq

You can access the dataset and related scripts through several academic and developer platforms:

In the standard dataset, faces can appear at any resolution. In VGGFace2-HQ, images typically undergo a size threshold filter (often retaining images with a facial resolution of 512x512 pixels or higher). Furthermore, the dataset employs advanced alignment techniques. Standard alignment crops a face based on eye coordinates, often cutting off the forehead or chin. VGGFace2-HQ utilizes whole-face alignment, preserving hair, ears, and neck regions, which are crucial for tasks like style transfer and 3D reconstruction. vggface2-hq

This results in a dataset that is smaller in volume than the original but infinitely more valuable for high-resolution training. It effectively solves the "garbage in, garbage out" problem that plaches GAN training, where low-res inputs lead to unstable generator convergence. You can access the dataset and related scripts

was born out of this necessity. It is not a completely new scrape but rather a meticulously filtered and processed subset. By leveraging automated quality assessment algorithms, researchers identified and isolated the "diamonds in the rough"—the subset of images that possess high pixel density, minimal compression artifacts, and sharp focus. Standard alignment crops a face based on eye

It would be irresponsible to discuss VGGFace2-HQ without addressing the elephant in the room: . The original VGGFace2 was scraped from Flickr under Creative Commons licenses. However, "consent" in web scraping is a grey area.