Connect with us

Civil Engineering

New Hybrid AI Tool Generates High-Quality Images Faster Than State-of-the-Art Approaches

Researchers developed a hybrid AI approach that can generate realistic images with the same or better quality than state-of-the-art diffusion models, but that runs about nine times faster and uses fewer computational resources. The tool uses an autoregressive model to quickly capture the big picture and then a small diffusion model to refine the details of the image.

Avatar photo

Published

on

The ability to generate high-quality images rapidly is crucial for producing realistic simulated environments that can be used to train self-driving cars to avoid unpredictable hazards. However, current generative AI techniques have drawbacks. Diffusion models can create stunningly realistic images but are too slow and computationally intensive for many applications. On the other hand, autoregressive models, which power LLMs like ChatGPT, are much faster but produce poorer-quality images.

Researchers from MIT and NVIDIA have developed a new approach that combines the best of both methods. Their hybrid image-generation tool, known as HART (Hybrid Autoregressive Transformer), uses an autoregressive model to quickly capture the big picture and then a small diffusion model to refine the details of the image. This innovative approach enables HART to generate images that match or exceed the quality of state-of-the-art diffusion models but do so about nine times faster.

The generation process consumes fewer computational resources than typical diffusion models, allowing HART to run locally on a commercial laptop or smartphone. A user only needs to enter one natural language prompt into the HART interface to generate an image.

HART has a wide range of potential applications, including helping researchers train robots to complete complex real-world tasks and aiding designers in producing striking scenes for video games. If you are painting a landscape, and you just paint the entire canvas once, it might not look very good. But if you paint the big picture and then refine the image with smaller brush strokes, your painting could look a lot better. That is the basic idea with HART.

The researchers encountered challenges in effectively integrating the diffusion model to enhance the autoregressive model. They found that incorporating the diffusion model in the early stages of the autoregressive process resulted in an accumulation of errors. Instead, their final design of applying the diffusion model to predict only residual tokens as the final step significantly improved generation quality.

Their method uses a combination of an autoregressive transformer model with 700 million parameters and a lightweight diffusion model with 37 million parameters. This allows HART to generate images of the same quality as those created by a diffusion model with 2 billion parameters but do so about nine times faster, using about 31 percent less computation than state-of-the-art models.

The future applications of this technology are vast and exciting. In the future, one could interact with a unified vision-language generative model, perhaps by asking it to show the intermediate steps required to assemble a piece of furniture. The researchers want to go down this path and build vision-language models on top of the HART architecture, since HART is scalable and generalizable to multiple modalities. They also want to apply it for video generation and audio prediction tasks.

Civil Engineering

The Sinking Cities of America: A Study Reveals Widespread Land Movement Across 28 Major U.S. Metropolises

A new study of the 28 most populous U.S. cities finds that all are sinking to one degree or another. The cities include not just those on the coasts, where relative sea level is a concern, but many in the interior. Furthermore, using newly granular data, the study finds that some cities are sinking at different rates in different spots, or sinking in some places and rising in others, potentially introducing stresses that could affect buildings and other infrastructure.

Avatar photo

Published

on

By

The study, published in Nature Cities, reveals that all 28 most populous U.S. cities are experiencing some degree of land movement, with the majority sinking at varying rates due to a combination of factors including groundwater extraction, climate change, and human activities such as construction and urbanization.

Lead author Leonard Ohenhen, a postdoctoral researcher at Columbia Climate School’s Lamont-Doherty Earth Observatory, notes that as cities continue to grow, subsidence can become more pronounced, producing stresses on infrastructure that may exceed safety limits. “We will see more cities expand into subsiding regions,” he says.

The study uses satellite data to map land movements in the 28 cities, including Houston, which is experiencing some of the most rapid sinking, with over 40% of its area subsiding more than 5 millimeters per year. Other Texas cities, Fort Worth and Dallas, are also among the fastest-sinking, while areas around New York’s LaGuardia Airport and parts of Las Vegas, Washington, D.C., and San Francisco are experiencing localized fast-sinking zones.

Researchers found that groundwater removal for human use was responsible for 80% of overall sinkage, with compaction below ground level causing subsidence at the surface. Climate-induced droughts in some areas will likely worsen subsidence in the future, says Ohenhen.

The study also reveals that natural forces are at work in some areas, such as the weight of ancient ice sheets that once covered much of interior North America. Even today, some cities like New York, Indianapolis, Nashville, Philadelphia, Denver, Chicago, and Portland are still subsiding due to these bulges, with rates ranging from 1 to 3 millimeters per year.

The researchers emphasize that continued population growth and water usage will likely exacerbate subsidence in the future. They recommend that cities focus on solutions such as land raising, enhanced drainage systems, and green infrastructure to mitigate flooding, and retrofitting existing structures to address tilting hazards.

Ohenhen concludes, “We have to move to solutions.” The study was coauthored by researchers from various institutions and provides a valuable resource for policymakers and urban planners to address the challenges posed by subsidence in major American cities.

Continue Reading

Artificial Intelligence

The Hidden Barrier to Advanced Robotic Touch

Researchers argue that the problem that has been lurking in the margins of many papers about touch sensors lies in the robotic skin itself.

Avatar photo

Published

on

By

The development of advanced robotic touch has been hindered by a seemingly innocuous yet critical issue – the insulating layer in robotic skin. Researchers at Northwestern University and Tel Aviv University have successfully overcome this barrier, paving the way for low-cost solutions that enable robots to mimic human touch.

In their study, the researchers observed that inexpensive silicon rubber composites used to make robotic skin host an insulating layer on both top and bottom surfaces. This prevents direct electrical contact between the sensing polymer and the monitoring surface electrodes, making accurate and repeatable measurements impossible. By eliminating this error, cheap robotic skins can now allow robots to sense an object’s curves and edges, essential for proper grasping.

The research team, consisting of electrical engineers and polymer materials scientists, shed light on this problem in a paper published in Advanced Electronic Materials. The study highlights the importance of validating electrical contacts, which might unknowingly obscure device performance.

“A lot of scientists misunderstand their sensor response because they lump together the behavior of the contacts with the behavior of the sensor material, resulting in inconsistent data,” said Matthew Grayson, professor of electrical and computer engineering at Northwestern’s McCormick School of Engineering. “Our work identifies the exact problem, quantifies its extent both microscopically and electrically, and gives a clear step-by-step trouble-shooting manual to fix the problem.”

The researchers detected that adding electrically conducting fillers like carbon nanotubes to rubber composites creates an ideal candidate for touch sensors. However, this material needs electrical signals, which are blocked by the insulating layer. By sanding down the ultrathin insulation layer, the team achieved a stronger electrical contact and calibrated the thickness of the insulating layer.

The collaboration between Northwestern University and Tel Aviv University is essential in addressing the “contact preparation” challenge. The researchers relied on each other’s expertise to prepare materials and study their properties, leading to consistent results across various variables.

As awareness spreads among researchers about the issue of reproducibility in touch sensing literature, new publications can be more rigorously relied upon to advance the field with new capabilities. The research was supported by various organizations, including the U.S. National Science Foundation, Northwestern University, and Tel Aviv University through the Center for Nanoscience & Nanotechnology.

The breakthrough has significant implications for robotics development, enabling robots to sense and interact with their environment more effectively. By overcoming this critical barrier, researchers have opened up new possibilities for advanced robotic touch, paving the way for future innovations in robotics and beyond.

Continue Reading

Aerospace

“Challenging the Classics: Researchers Reveal New Insights into Material Deformation under Stress”

Scientists have expanded on a longstanding model governing the mechanics behind slip banding, a process that produces strain marks in metals under compression, gaining a new understanding of the behavior of advanced materials critical to energy systems, space exploration and nuclear applications.

Avatar photo

Published

on

By

Researchers at the University of California, Irvine (UCI) have made a groundbreaking discovery in the field of materials science. By expanding on a classic model developed over 70 years ago, scientists in UCI’s Samueli School of Engineering have gained new insights into the behavior of advanced materials critical to energy systems, space exploration, and nuclear applications.

The traditional Frank-Read theory attributed slip band formation to continuous dislocation multiplication at active sources. However, the UC Irvine team found that extended slip bands emerge from source deactivation followed by the dynamic activation of new dislocation sources. This process was observed at the atomic scale through mechanical compression on micropillars made of a chromium-cobalt-nickel alloy.

Using advanced microscopy techniques and large-scale atomistic modeling, researchers were able to visualize the confined slip band as a thin glide zone with minimal defects and the extended slip band with a high density of planar defects. This understanding has shed new light on collective dislocation motion and microscopic deformation instability in advanced structural materials.

Deformation banding, where strain concentrates in local zones, is a common phenomenon in various substances and systems, including crystalline solids, metals, granular media, and even geologic faults under compressive stress. The discovery of extended slip bands challenges the classic model developed by physicists Charles Frank and Thornton Read in the 1950s.

“This foundational knowledge will accelerate the discovery of materials with tailored and predictable mechanical properties to meet the rising demand for advanced materials resilient to extreme environments across energy and aerospace sectors,” said corresponding author Penghui Cao, UC Irvine associate professor of mechanical and aerospace engineering.

The research was funded by the U.S. Department of Energy, UC Irvine, and the National Science Foundation (through the UC Irvine Center for Complex and Active Materials). The project involved graduate students, research specialists, and other professors from UCI’s Department of Mechanical and Aerospace Engineering and Department of Materials Science and Engineering.

Continue Reading

Trending