AI ML DL

One2Avatar examples
AI ML DL

Better Neural Avatars From Just Five Face Images

Many neural avatar systems of the last 18 months require extensive training data, or even full videoclips. Others are performant, but have exorbitant training demands. However, a new system from Google and the University of Minnesota is proposing a photorealistic deepfake head system that’s trained on only five images – and can work quite well from just one image; and the new system of pretraining that the framework uses throws some of the conventions regarding hyperscale training datasets into question.

Read More »
AI ML DL

The Challenge of Preventing ‘Identity Bleed’ in Face Swaps

KAIST AI has developed a new method of disentangling identity characteristics in a face-swap from secondary characteristics such as lighting, skin texture – and the original structure of the face to be ‘overwritten’ by the new identity. If such techniques can be perfected, facial replacement could be freed from having the original identity ‘bleeding through’ into the superimposed identity.

Read More »
Fake faces broken down into the spectral ranges that may reveal them
AI ML DL

Combating Stable Diffusion Face Forgery Through Frequency Analysis

For over six years, deepfake detection methods have sought to find a criteria for detection that can survive the evolution of generative systems. A new paper from Switzerland and France offers such a method, by examining the way that frequency spectra differ among real and generated images.

Read More »
Boximator examples
AI ML DL

A Text-To-Video Method That Actually Generates Some Action

Though there is currently a tidal wave of research into text-to-video systems, they tend to be criticized (if at all) because the videos are very short and/or very small. However, the biggest constriction on the current state of the art in T2V is that these videos have very little actual real movement. Now, a new system from ByteDance offers a novel box-based system capable of creating footage with significant motion, taking us a little nearer the goal of genuinely photorealistic pure neural video synthesis.

Read More »
Examples of DiffBody in action
AI ML DL

Reshaping the Human Body With CGI and Stable Diffusion

Driven by the interests of the fashion industry, new methods are emerging to automatically reshape and re-pose pictures of real people. It’s a difficult task, and yet another neural procedure that’s being aided by old-school CGI. A new method from Japan now claims to have obtained the state-of-the-art in neural body-shaping, based on just one picture of a person.

Read More »
AI ML DL

An Old Algorithm May Be Holding back New Generative AI Systems

Modern generative AI systems such as Stable Diffusion may be being held back by an out-dated but really popular loss algorithm called Fréchet Inception Distance (FID). Even though FID was only invented in 2019, it did not anticipate the growth of datasets or the advent of latent diffusion systems. New research from Google takes a critical look at FID and suggests a leaner and more modern alternative.

Read More »
AI ML DL

What’s The Difference between CGI and AI?

Generative AI promises, apparently, to replace traditional CGI techniques in Hollywood, if you believe the current punditry on the topic. In truth, for some time to come, it’s more likely that the two technologies will work in tandem to create a new standard in realism for visual effects workflows.

Read More »
AI ML DL

Solving the ‘Profile View’ Crisis in Facial Image Synthesis

Facial synthesis and deepfake systems are not very good at creating profile views of people, because there isn’t much training data of this type. Because of this, many security solutions have been built around profile (side-view) verification. However, a new and exhaustive profile view dataset could be a game-changer not only for VFX practitioners, but also for online scammers.

Read More »
AI ML DL

Using Traditional CGI to Re-Sculpt Generative AI

Editing AI images using only words is a hit-and-miss affair – but a new approach offers a novel take on this challenge, by converting a relevant part of an image into a CGI mesh and allowing the user to manipulate it with decades-old methods, before re-inserting the altered image back into the photo.

Read More »