Could Gaussian Splatting become the hottest new deepfake technology since 2017? The massive surge of interest from the research sector suggests it might – and the latest innovation not only brings full controllability to neural or deepfaked faces, but also lets you become someone else at an unprecedented level of photorealism and efficiency.
CLIP is the new darling of the computer vision research, and of image-based generative AI, with wide uptake of the image/text analysis framework across the sector. However, new research indicates that CLIP’s efficiency and usefulness is negatively affected by badly-compressed images. Though this should not be a problem in the modern high-speed broadband age, it is – because so much essential data and methodologies still in use data back several decades.
Though it’s possible to detect deepfake video by studying movement in a deepfaked face, it takes a lot of resources to do it, and it’s not easy or cheap to use it to detect potential ‘live’ deepfakes in video streaming calls. However, the most popular video codec in the world, H.264, has already done a lot of the work necessary to discern deepfake content, and new research offers a way to cheaply use these ‘motion vectors’ to discern simulated faces.
A new generation of fake image detectors hope to discern AI-generated photos from real ones. But how do they cope when real photos exhibit qualities normally found in fake photos? A new an seminal database of ‘impossible-but-true’ photos hopes to stir deeper interest in the nascent field of photo verification.
Bad rendering of eye direction in neural facial synthesis can make the difference between an image looking photoreal and looking like bad CGI. We are experts in eye contact, but machine learning systems only have small pixel spaces in which to attempt to recreate our survival-level understanding of eye gaze, and often make rudimentary mistakes, such as conflating head position with eye position. But a new paper is using sound as an extra factor to guess where people in a video are really looking, offering potential for improved gaze estimation in downstream applications.
Researchers from Germany have developed a new method of restoring obstructed faces using only Generative Adversarial Networks (GANs), instead of needing to undertake expensive and time-consuming fine-training of existing models.
New research specifically targets deepfake faces created by latent diffusion models such as Stable Diffusion, in a security sector that seems stuck on GANs and 2017-era autoencoders.
Though Stable Diffusion is an impressive generative system, it has difficulty performing the same operation twice, which makes it difficult to render the same subject consistently across frames – a major obstacle to the development of latent diffusion models aimed at producing coherent video. Now, a new system is offering a method that can output image concepts in a way that’s more consistent and repeatable, as CGI is.
Loss functions, which determine the ways that a machine learning network should develop during the course of training, are at the heart of generative AI and computer vision systems. However, the loss functions commonly used in facial neural synthesis are not specifically designed for faces, which can be a retarding and even confusing factor in framework development. Now, Disney Research and EFL Zurich are proposing an novel architecture-agnostic loss function designed specifically for facial workflows, and which is based on 1990s research into the way humans perceive 3D faces based on shading.
New work from Google Research rewrites the terrible alternative captions that are often found in images on the web, and which subsequently get used as pertinent information for generative models such as Stable Diffusion, so that they contain richer and more useful information. The authors found that fine-tuning Stable Diffusion on these better-annotated images wrought significant improvements in the generated images.
