We eliminate “texture sticking” in GANs through a comprehensive overhaul of all signal processing aspects of the generator, paving the way for better synthesis of video and animation.

Unaufhaltsam entwickeln sich KI-Systeme weiter. Bilder, die damit erstellt werden, können täuschend echt erscheinen. Das verunsichert einige Menschen. Doch eine Untersuchung zeigt, dass man seinen Bli...

Date Submitted: Jan 13, 2025. Open Peer Review Period: Jan 15, 2025 - Mar 12, 2025.

Contribute to AtomicNixon/Stylegan3-Audio development by creating an account on GitHub.

A few minutes of focused training can help people see what artificial intelligence often hides in plain sight. Researchers from the Universities of Reading, Greenwich, Leeds and Lincoln have found that just five minutes of guided practice can noticeably improve a person’s ability to tell whether a face was created by AI or captured from real life. The work involved 664 participants and tested how humans respond to images made by StyleGAN3, one of the most advanced face-generation systems available when the study was carried out. The researchers grouped participants into two categories. One consisted of “super-recognizers,” people whose natural face recognition skills are far stronger than average. The second group included typical observers with no special ability beyond normal vision and memory. Each participant had to decide whether faces on screen were real or AI-generated. At first, even the most talented participants struggled. Super-recognizers managed to correctly identify fake faces 41 percent of the time. Typical participants did worse, at only 31 percent. In both cases, those numbers were below chance level. In other words, many people would have done better by guessing. This reflects what scientists call AI hyperrealism, a phenomenon where computer-generated faces appear so natural that they seem more believable than genuine human photographs. To see if skill could be improved, researchers created a short training exercise that showed participants examples of common visual errors produced by generative models. The tutorial pointed out small but telling flaws such as irregular hair strands, awkward tooth patterns, and mismatched details near the edges of the face. Participants were then given a short practice round with immediate feedback after each choice. The entire session took about five minutes. The results were striking. After training, super-recognizers improved their detection accuracy to 64 percent. Typical participants rose to 51 percent, nearly reaching chance level but still a clear improvement. The gains appeared consistent across most of the images tested, not just the easiest or most flawed ones. According to the data, more than half of the synthetic faces used in the trials saw accuracy improvements greater than ten percent once training was introduced. These changes suggest that a small burst of attention training can shift how people look at faces. It does not simply make them more suspicious. Instead, the study found that trained participants became more sensitive to real structural cues rather than randomly flagging all images as fake. That distinction matters because the researchers used signal detection methods to confirm that the training enhanced perception, not bias. While the overall numbers may seem modest, they carry weight for digital security. Synthetic faces have already appeared in fake social media profiles, scam accounts, and identity fraud attempts. AI hyperrealism gives these images a trust advantage, especially when users or verification systems rely on instinctive judgments. The new findings show that even a small amount of human instruction can make that instinct more reliable. The researchers also noted differences in how people with exceptional recognition ability approach such tasks. Super-recognizers tended to take longer before deciding, which could reflect more deliberate visual processing rather than hesitation. Their advantage did not come from simple caution, since response time and accuracy were not closely related. Instead, it points to an underlying perceptual skill that can be further strengthened with proper guidance. What stands out from the data is that training benefited both groups by a similar margin. This means super-recognizers were not merely better at spotting technical rendering mistakes. They likely used deeper visual cues beyond texture or artifact recognition. Typical participants, once trained, also reduced their bias toward judging every image as real. In both cases, five minutes of exposure helped them recalibrate their sense of what an authentic human face looks like. The study used StyleGAN3 because it represented a major leap in how synthetic faces are rendered. Earlier versions, such as StyleGAN2, produced more obvious distortions. The newer system generated faces so convincing that untrained people frequently misjudged them as genuine. This shows how quickly AI realism has progressed and why human detection skills need continuous updating. Researchers warn that as newer models become more refined, the obvious cues may fade away. Hair strands may align perfectly, teeth may appear uniform, and backgrounds may blend seamlessly. In such conditions, short visual tutorials might need to evolve or be combined with machine-based detectors. Future experiments will test whether the benefits of this brief training last over time and whether groups of trained observers can outperform automated detection systems when working together. For now, the study highlights a practical defense that costs almost nothing but awareness. A few minutes of instruction can help ordinary people, and even experts, resist the illusion of AI realism. As synthetic media continues to flood the internet, the ability to spot what looks “too perfect” may turn out to be one of the most valuable human skills left in a world full of digital faces. Notes: This post was edited/created using GenAI tools. Read next: Lower App Store Fees Didn’t Mean Lower Prices for Users in Europe

I’m working on a project involving: 1. **Face swapping** (e.g., swapping faces in images while preserving expressions) 2. **Celebrity face matching** (similarity scoring across datasets) 3. **Text-driven background replacement** 4. **Image-to-video generation** (e.g., animating a single image with text prompts) 5. **Dual-image video generation** (e.g., creating a handshake video from two portraits) I’ve researched models like **DeepFaceLab, StyleGAN3, RunwayML, and Disco Diffusion** , but I’m struggling to find **direct comparisons** for these niche tasks. Have you encountered studies, benchmarks, or repositories that evaluate models for these specific functionalities? Additionally, if you’ve built similar pipelines, which open-source tools/libraries (e.g., PyTorch3D, Transformers) did you find most effective?

Date Submitted: Jan 13, 2025. Open Peer Review Period: Jan 15, 2025 - Mar 12, 2025.

Date Submitted: Jan 13, 2025. Open Peer Review Period: Jan 15, 2025 - Mar 12, 2025.

Date Submitted: Jan 13, 2025. Open Peer Review Period: Jan 15, 2025 - Mar 12, 2025.