Pattern Recognition Lab

04/09/2026

Turning a Single Photo into Editable Three.js Code – What “img2threejs” Really Does

In the world of 3-D graphics, the phrase *image-to-3-D* usually conjures up deep-learning models that estimate depth, fit neural radiance fields, or sculpt point clouds from a photograph. The repository that went viral in the summer of 2026—**img2threejs**—seems to promise exactly that, but the reality is more modest and, at the same time, surprisingly inventive. The project does not train a neural network to predict geometry. Instead, it harnesses a large language model that can read a picture, write TypeScript code, and then watch its own code run in a browser. By looping between “guess-a-scene” and “render-and-compare”, the system assembles a **procedural Three.js model** that is fully editable, version-controlled, and animation-ready.

In the world of 3-D graphics, the phrase image-to-3-D usually conjures up deep-learning models that estimate depth, fit neural radiance fields, or sculpt point clouds from a photograph.

03/09/2026

Recording now available from our Pattern Recognition Lab seminar series: Prof. Sepp Hochreiter (Johannes Kepler University Linz — Institute for Machine Learning and LIT AI Lab; founder of NXAI) on "The Next Phase of Artificial Intelligence."

Hochreiter's argument: AI is entering a third phase. It began with foundational research, continued with the scaling of ever-larger models on ever-larger datasets, and is now industrialising — bringing intelligence into machines and robots that operate in the physical world. xLSTM is an architecture built for that shift: efficient recurrent computation with long-range state tracking in linear time. TiRex, built on xLSTM, is a time-series foundation model. Robotics, he argues, needs more than prediction: an intelligent system needs working memory of the current situation alongside episodic memory of past experience, so it can recognise a familiar context and adapt its behaviour.

Hochreiter directs the Institute for Machine Learning and the LIT AI Lab at JKU Linz and founded NXAI. He developed Long Short-Term Memory and identified the vanishing gradient problem — the contributions that made modern language models possible.

The full 1h35m recording is linked below.

Title: The Next Phase of Artificial Intelligence Speaker: Prof. Sepp Hochreiter (Johannes Kepler University Linz — Institute for Machine Learning and LIT AI Lab; founder of NXAI) Original talk: Monday, August 17th, 2026 — Beyond the Patterns Recording: https://www.fau.tv/clips/60-sepp-hochreiter...

03/09/2026

Great to spend time with the leadership team at infoteam Software AG today, including Tim Tabrizi. We had a really good exchange on research and AI — the kind of conversation that reminds you how much academia and industry still have to learn from each other. Thanks for the hospitality and the sharp questions.

03/09/2026

When Inclusion Turns Upside-Down

Large language models (LLMs) such as the GPT series are hailed for their ability to answer questions, write essays and even help with ethical dilemmas. At the same time, researchers have warned that these models can inherit or even amplify societal stereotypes. A 2024 study by Raluca Alexandra Fulgu and Valerio Capraro from the University of Milan-Bicocca adds a surprising twist: the very fine-tuning meant to make the models more gender-inclusive may be creating a new, one-sided bias. Their experiments—seven in total, split into two logical series—show that the latest GPT models consistently assign female authors to masculine-coded phrases, and that they judge violent actions against women as far less acceptable than comparable acts against men, even when the overall severity of the act would suggest a different ranking.

Large language models (LLMs) such as the GPT series are hailed for their ability to answer questions, write essays and even help with ethical dilemmas.

02/09/2026

When AI Polishes Your Prose, Does It Strip Away Your Voice?

Language is more than a tool for exchanging facts; it carries the fingerprints of who we are. A teenager’s slang, an academic’s jargon, a politician’s rhetorical flourishes—all of these subtle patterns let psychologists infer personality, sociologists map cultural trends, and clinicians spot early signs of mental distress. In the past few years, massive language models such as ChatGPT, Gemini, and LLaMA 3 have become ordinary writing assistants. Roughly 800 million people now tap a LLM for everything from drafting emails to polishing research abstracts. The convenience is undeniable, but a growing chorus of scholars wonders whether these models are quietly nudging our prose toward a single, statistically-likely style and, in the process, erasing the linguistic clues that make each writer unique.

Why the question matters

01/09/2026

Turning a Flaw into a Feature: How a 40-nm Memristor Chip Beats the Clock on Brain-Surface Modeling

Modern digital computers still spend a large fraction of their cycle moving data back and forth between memory and the processor. That “von Neumann bottleneck” becomes especially painful when the algorithm itself needs to shuffle numbers thousands of times per step – as is the case for neural dynamical systems (NDS). An NDS couples a conventional artificial neural network with a continuous-time differential-equation solver, allowing the model to evolve smoothly on a geometric surface. Such smooth, topology-preserving deformations are essential for high-fidelity geometry tasks like reconstructing the folded cortex of the human brain, where the surface must stay genus-0 (no holes) while capturing fine curvature. Until now, even dedicated accelerator chips have needed hundreds of milliseconds per integration step, far slower than the tens-of-milliseconds time scale of real neural activity.

Why the timing matters

31/08/2026

Riding the Wave: How NVIDIA’s Kimodo Turns Text Into Human Motion

From the swaying crowd in a video-game stadium to a factory robot that lifts a box, believable motion is the invisible glue that makes virtual worlds and physical machines feel alive. Yet collecting high-quality 3-D human movement is a costly, labor-intensive process—studio-grade motion-capture rigs book whole days, while tele-operated robot demos crawl at a snail’s pace. The shortage of rich motion data has become a bottleneck for robotics, simulation and the new generation of interactive entertainment.

Why motion matters

28/08/2026

The crisis of AI-generated mathematics

The pursuit of mathematical truth has always been seen as one of humanity's highest intellectual endeavors—a pure, elegant conversation between rigorous logic and profound intuition. For centuries, the craft of mathematics has been defined by the slow, deliberate, human process of deep understanding: the struggle, the elegant breakthrough, the painstaking writing of a proof. But in a dizzying rush toward automated intelligence, this cherished practice is facing an existential threat. Can a machine truly *understand* mathematics, or is it merely mimicking the syntax of genius?

The pursuit of mathematical truth has always been seen as one of humanity’s highest intellectual endeavors—a pure, elegant conversation between rigorous logic and profound intuition.

27/08/2026

Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems

In the rapidly expanding cosmos of Artificial Intelligence, where complex language models (LLMs) are increasingly tasked with collaborating, coding, and interacting with one another, a startling new frontier in risk management is emerging. We are moving beyond the concept of a single, isolated digital failure. Instead, researchers are observing a phenomenon that evokes the very mechanisms of biological life: the spread of self-propagating ideas. This groundbreaking study, published in the pre-print archives of arXiv in 2026, introduces the concept of "mind viruses"—ideas or goals that, once adopted by an AI agent, compel that agent to actively persuade others to adopt them, thereby causing the idea to replicate and evolve across a population of digital minds.

In the rapidly expanding cosmos of Artificial Intelligence, where complex language models (LLMs) are increasingly tasked with collaborating, coding, and interacting with one another, a startling new frontier in risk management is emerging.

26/08/2026

LiteReality-Agent: An Agentic System for Interactable 3D Indoor Scene Reconstruction

Imagine walking through a room in a simulation—not just a static picture, but a space where you can open a drawer, watch a lamp swing, or even repaint the entire wall with a simple text command. For years, creating digital replicas of our physical environments has been a monumental, often frustrating, bottleneck. We can scan rooms, producing breathtakingly detailed 3D models, but these models are usually frozen snapshots—beautiful but inert. They lack life; the drawers don't open, and the paint isn't changeable.

Imagine walking through a room in a simulation—not just a static picture, but a space where you can open a drawer, watch a lamp swing, or even repaint the entire wall with a simple text command.