The artificial intelligence (AI) community has gotten very good at creating fake videos — just look at OpenAI's Sora, which featured a slick imaginary flythrough last month. You need to ask intelligent and practical questions. Should I stop doing all these videos?
Also: OpenAI has launched a text-to-video model and the results are surprising.see for yourself
This week, Google scholar Enric Corona and his colleagues answered, “Use the VLOGGER tool to control them.” VLOGGER can generate high-resolution videos of people talking based on a single photo. More importantly, VLOGGER can animate videos according to audio samples. This means the technology can animate videos as controlled human likenesses, or high-fidelity “avatars.”
All kinds of creations are possible with this tool. At the simplest level, Corona's team suggests that VLOGGERs could have a big impact on help desk avatars because artificially speaking humans that look more realistic can “develop empathy.” I am. They suggest that the technology could “enable entirely new use cases, including enhanced online communications, education, and personalized virtual assistants.”
VLOGGERs could also lead to a new frontier of deepfakes, which make statements and actions made by real people appear to be genuine. Corona's team will provide insight into the social impact of his VLOGGER in supplemental supporting materials. However, that material is not available on his GitHub page for the project. ZDNET contacted Corona for supporting documentation, but had not received a response at the time of publication.
Also: As AI agents become more widespread, the risks will increase, say academics.
As explained in the official paper “VLOGGER: Multimodal Diffusion for Embodied Avatar Synthesis”, Corona's team aims to overcome the inaccuracies of state-of-the-art avatars. “Creating realistic videos of humans is still complex and ripe with artifacts,” Corona's team wrote.
The research team pointed out that existing video avatars often cut off the body and hands, leaving only the face visible. VLOGGER allows you to view your entire torso along with your hand movements. Other tools typically offer only basic lip syncing, with limited variation in facial expressions and poses. VLOGGER can produce “high-resolution videos of head and upper body movements” […] Featuring a wide variety of facial expressions and gestures, it is “the first approach to generating speaking and moving humans given audio input.”
As the research team explained, “It is precisely automation and behavioral realism that [are] What we aim to do with this research: VLOGGER is a complex facial and body movement level with audio and animated visual representations designed to support natural conversations with human users. A multimodal interface to embodied conversational agents, featuring an increase in . ”
VLOGGER summarizes some of the latest trends in deep learning.
Multimodality aggregates many modes that AI tools can absorb and synthesize, such as text and audio, images and video.
Large language models such as OpenAI's GPT-4 allow you to use natural language as input to perform various types of actions, such as creating paragraphs of text, songs, or images.
Researchers have also recently discovered numerous ways to create realistic-looking images and videos by improving “spreading.” The term originates from molecular physics and refers to the way particles of matter move from being highly concentrated in a particular area to becoming more diffuse as temperature increases. By analogy, bits of digital information appear to be “spread out” to the extent that they become incoherent due to digital noise.
Also: Beyond Gemini, open source AI has its own video tricks
With the rise of AI, noise is introduced into the image, the original image is reconstructed, and a neural network is trained to find the built rules. Diffusion is at the heart of the impressive image generation processes in Stable AI's Stable Diffusion and OpenAI's DALL-E. This is also how OpenAI creates smooth videos with Sora.
For VLOGGER, Corona's team trained a neural network to associate a speaker's audio with individual frames of that speaker's video. The team used yet another recent innovation, his Transformer, to combine a diffusion process that reconstructs video frames from audio.
Transformer uses attention methods to predict video frames based on frames that have occurred in the past, in combination with audio. By predicting actions, neural networks learn how to accurately render hand and body movements and facial expressions frame by frame, in sync with the audio.
The final step uses the predictions from the first neural network, followed by a second neural network that also uses diffusion to enhance the generation of high-resolution frames for the video. This second step is also the high water mark for the data.
Also: Generative AI fails at this very common ability of human thinking.
To create the high-resolution images, Corona's team compiled MENTOR, a dataset featuring 800,000 “identities” of videos of people talking. MENTOR consists of 2,200 hours of video, which the team says is the “largest dataset ever used in terms of identity and length” and more powerful than previous comparable datasets. even he claims to be 10 times larger.
The authors found that the process could be enhanced with a subsequent step called “fine-tuning.” Having already been “pre-trained” with MENTOR, sending a full-length video to her VLOGGER allows you to more realistically capture the idiosyncrasies of human head movements, such as blinking. “By fine-tuning the diffusion model with more data, VLOGGER learns how to better capture identity on monocular videos of subjects, for example when the reference image appears to have their eyes closed. Yes, the team calls this process “personalization.”
The larger point of this approach is that it couples the predictions within a single neural network with high-resolution images, and what makes VLOGGER so provocative is that the program, like Sora, simply generates videos. It's not just that. VLOGGER links that video to actions and expressions that you can control. Its lifelike video unfolds and can be manipulated like a puppet.
Also: Nvidia CEO Jensen Huang unveils next-generation 'Blackwell' chip family at GTC
“Our aim is to bridge the gap between recent video compositing efforts that can generate dynamic videos without controlling identity or pose, and controllable image generation methods,” says Corona's team. is writing.
VLOGGER can not only be a voice-driven avatar, but also potentially lead to editing features such as changing the speaking subject's mouth or eyes. For example, you can change a virtual person in a video that blinks a lot to blink only a little or not at all. You can also narrow your wide-mouthed speech to make more detailed lip movements.
Despite achieving new cutting-edge technology for simulating humans, the question the Corona team did not address is what the world should expect from misuse of the technology. It's easy to imagine a portrait of a politician saying absolutely devastating things about, say, impending nuclear war.
Perhaps the next step in this Avatar game will be a shockingly lifelike version of the Voight-Kampf test in Blade Runner that will allow society to tell which speakers are real and which are just deepfakes. It will be a neural network that will help you distinguish in this way.