Making hyperscale datasets that are both realistic and abundant is essential to the development of the computer vision and image synthesis research sectors. But it’s not easy to get both of these qualities. A new UK offering bridges this gap, providing a dataset of 250,000 photorealistic faces rendered through Stable Diffusion and ControlNet.
Researchers have proposed what may be the most radical change in deepfake software since it emerged in 2017. By removing what was formerly considered one of the most fundamental parts of the software, they’ve been able to achieve much better face-swapping results. The new approach could also have implications across other machine learning sectors which have used ‘skip connections’ (a convenient but arguably slipshod way of sorting through data, which unfortunately tends to make resulting machine learning models inflexible).
A new project from Tel-Aviv University offers a superior method of facilitating deepfake puppetry, as well as editing key facial components (such as expressions) in the latent space of a machine learning system, all via a minute’s scan-capture with a mobile phone, similar to the Apple mobile procedure for setting up a face ID login system.
One day, video-conferencing is likely to be a low-bandwidth, neurally-powered affair, with participants only sending minimal information about their facial movements, which will power locally-built, on-the-fly avatars at the other end. The question is, what are the potential security implications? A new paper offers some solutions in advance of this development.
A new research paper proposes to edit the usually rigid contents of a Neural Radiance Field using text-to-image technologies, and the controversial DreamBooth method. But does this bring us any nearer to a neural synthesis method that is both ductile and temporally consistent?
New research into Facial Effect Recognition (FER) seeks to understand facial expressions in the broader context of their environment. In much the same way we judge whether or not facial expressions are ‘appropriate’ for any given situation. the next generation of recognition systems seem set to consider multiple streams, such as text content, ambience and many other clues – and not just how the facial muscles are disposed.
Adversarial watermarking has been proposed many times as a preventative measure against deepfaking, but usually with the intention to entirely disrupt the generative process. But a new method proposes that it could be used to give individuals more control over the use of their identities in such systems.
Deepfake video is possible through an act of last-minute vandalism, in terms of neural architecture, where the two trained identities are wired into each other, long after training has finished. A new paper from China posits that if some of the training effort was expended on actually forging relationships between the two identities, results could be notably improved.
EbSynth is a non-AI system that lets animators transform video from just a handful of keyframes; but a new approach, for the first time, leverages it to allow temporally-consistent Stable Diffusion-based text-to-image transformations in a NeRF framework. Is this a step forward towards general temporal stability, or a concession that Stable Diffusion can never natively produce the video consistency that so many think is just around the corner?
A new initiative from UC Berkeley uses Transformers to achieve a new state-of-the-art in human pose extraction efficiency and accuracy, paving the way for a variety of improvements across various research lines in image synthesis, as well as offering possible benefits in other sectors, such as health and security.
