arXiv subjects

Xu Duan

Publications and source records attributed to Xu Duan.

14 recordsLinked to original sources

arXiv2025

Cosine-Similarity Methods for Efficient Training and Sampling in High-Dimensional Latent Spaces

Latent generative models are increasingly shifting from traditional VAEs toward representation autoencoders and semantically aligned latent spaces, which lift images into higher-dimensional feature domains where semantic factors become more separable. Yet these spaces also contain geometric regularities that existing methods do not fully exploit--particularly in the directional relationships between features. We introduce a cosine-similarity-based mechanism that improves both training and sampling by selecting couplings that produce cleaner, less entangled velocity fields. This simple alignment reduces gradient noise, accelerates convergence, and improves sample fidelity. Building on this idea, we develop cosine-similarity-based fine-tuning and time-scheduling strategies that reduce the FID of an 800-epoch RAE from 11.99 to 8.60. Furthermore, by formulating an optimal-transport coupling using a cosine cost, a single-epoch fine-tuning step at the 20-epoch checkpoint reaches 3.30 FID-matching the performance of the 80-epoch baseline.

math.OC↗

arXiv2025

DM-MPPI: Datamodel for Efficient and Safe Model Path Integral Control

We extend the Datamodels framework from supervised learning to Model Predictive Path Integral (MPPI) control. Whereas Datamodels estimate sample influence via regression on a fixed dataset, we instead learn to predict influence directly from sample cost features, enabling real-time estimation for newly generated samples without online regression. Our influence predictor is trained offline using influence coefficients computed via the Datamodel framework across diverse MPPI instances, and is then deployed online for efficient sample pruning and adaptive constraint handling. A single learned model simultaneously addresses efficiency and safety: low-influence samples are pruned to reduce computational cost, while monitoring the influence of constraint-violating samples enables adaptive penalty tuning. Experiments on path-tracking with obstacle avoidance demonstrate up to a $5\times$ reduction in the number of samples while maintaining control performance and improving constraint satisfaction.

eess.SY↗

arXiv2025

Trajectory-Optimized Density Control with Flow Matching

Optimal transport (OT) and Schr{\"o}dinger bridge (SB) problems have emerged as powerful frameworks for transferring probability distributions with minimal cost. However, existing approaches typically focus on endpoint matching while neglecting critical path-dependent properties -- particularly collision avoidance in multiagent systems -- which limits their practical applicability in robotics, economics, and other domains where inter-agent interactions are essential. Moreover, traditional density control methods often rely on independence assumptions that fail to capture swarm dynamics. We propose a novel framework that addresses these limitations by employing flow matching as the core modeling tool, where the flow model co-evolves with the control policy. Unlike prior methods that treat transport trajectories as mere interpolations between source and target distributions, our approach explicitly optimizes over the entire transport path, enabling the incorporation of trajectory-dependent costs and collision avoidance constraints. Our framework bridges optimal transport theory with mean field control, providing a principled approach to multiagent coordination problems where both endpoint alignment and path properties are critical. Experimental results demonstrate that our method successfully generates collision-free transport plans while maintaining computational efficiency comparable to standard flow matching approaches.

math.OC↗

arXiv2025

RAGSR: Regional Attention Guided Diffusion for Image Super-Resolution

The rich textual information of large vision-language models (VLMs) combined with the powerful generative prior of pre-trained text-to-image (T2I) diffusion models has achieved impressive performance in single-image super-resolution (SISR). However, existing methods still face significant challenges in generating clear and accurate regional details, particularly in scenarios involving multiple objects. This challenge primarily stems from a lack of fine-grained regional descriptions and the models' insufficient ability to capture complex prompts. To address these limitations, we propose a Regional Attention Guided Super-Resolution (RAGSR) method that explicitly extracts localized fine-grained information and effectively encodes it through a novel regional attention mechanism, enabling both enhanced detail and overall visually coherent SR results. Specifically, RAGSR localizes object regions in an image and assigns fine-grained caption to each region, which are formatted as region-text pairs as textual priors for T2I models. A regional guided attention is then leveraged to ensure that each region-text pair is properly considered in the attention process while preventing unwanted interactions between unrelated region-text pairs. By leveraging this attention mechanism, our approach offers finer control over the integration of text and image information, thereby effectively overcoming limitations faced by traditional SISR techniques. Experimental results on benchmark datasets demonstrate that our approach exhibits superior performance in generating perceptually authentic visual details while maintaining contextual consistency compared to existing approaches.

cs.CV↗

arXiv2025

SCALAR: Scale-wise Controllable Visual Autoregressive Learning

Controllable image synthesis, which enables fine-grained control over generated outputs, has emerged as a key focus in visual generative modeling. However, controllable generation remains challenging for Visual Autoregressive (VAR) models due to their hierarchical, next-scale prediction style. Existing VAR-based methods often suffer from inefficient control encoding and disruptive injection mechanisms that compromise both fidelity and efficiency. In this work, we present SCALAR, a controllable generation method based on VAR, incorporating a novel Scale-wise Conditional Decoding mechanism. SCALAR leverages a pretrained image encoder to extract semantic control signal encodings, which are projected into scale-specific representations and injected into the corresponding layers of the VAR backbone. This design provides persistent and structurally aligned guidance throughout the generation process. Building on SCALAR, we develop SCALAR-Uni, a unified extension that aligns multiple control modalities into a shared latent space, supporting flexible multi-conditional guidance in a single model. Extensive experiments show that SCALAR achieves superior generation quality and control precision across various tasks. The code is released at https://github.com/AMAP-ML/SCALAR.

cs.CV↗

arXiv2025

FLUX-Text: A Simple and Advanced Diffusion Transformer Baseline for Scene Text Editing

Scene text editing aims to modify or add texts on images while ensuring text fidelity and overall visual quality consistent with the background. Recent methods are primarily built on UNet-based diffusion models, which have improved scene text editing results, but still struggle with complex glyph structures, especially for non-Latin ones (\eg, Chinese, Korean, Japanese). To address these issues, we present \textbf{FLUX-Text}, a simple and advanced multilingual scene text editing DiT method. Specifically, our FLUX-Text enhances glyph understanding and generation through lightweight Visual and Text Embedding Modules, while preserving the original generative capability of FLUX. We further propose a Regional Text Perceptual Loss tailored for text regions, along with a matching two-stage training strategy to better balance text editing and overall image quality. Benefiting from the DiT-based architecture and lightweight feature injection modules, FLUX-Text can be trained with only $0.1$M training examples, a \textbf{97\%} reduction compared to $2.9$M required by popular methods. Extensive experiments on multiple public datasets, including English and Chinese benchmarks, demonstrate that our method surpasses other methods in visual quality and text fidelity. All the code is available at https://github.com/AMAP-ML/FluxText.

cs.CV↗

arXiv2024

Origin of Interstitial Doping Induced Coercive Field Reduction in Ferroelectric Hafnia

Hafnia-based ferroelectrics hold promise for nonvolatile ferroelectric memory devices. However, the high coercive field required for polarization switching remains a prime obstacle to their practical applications. A notable reduction in coercive field has been achieved in ferroelectric Hf(Zr)$_{1+x}$O$_2$ films with interstitial Hf(Zr) dopants [Science 381, 558 (2023)], suggesting a less-explored strategy for coercive field optimization. Supported by density functional theory calculations, we demonstrate the $Pca2_1$ phase, with a moderate concentration of interstitial Hf dopants, serves as a minimal model to explain the experimental observations, rather than the originally assumed rhombohedral phase. Large-scale deep potential molecular dynamics simulations suggest that interstitial defects promote the polarization reversal by facilitating $Pbcn$-like mobile 180$^\circ$ domain walls. A simple pre-poling treatment could reduce the switching field to less than 1 MV/cm and enable switching on a subnanosecond timescale. High-throughput calculations reveal a negative correlation between the switching barrier and dopant size and identify a few promising interstitial dopants for coercive field reduction.

cond-mat.mtrl-sci↗

arXiv2023

Emergent superconductivity in doped ferroelectric hafnia

Superconductivity and ferroelectricity,representing two distinct forms of ordered states, are typically not found together in the same system, making it even more difficult to create a connection between them. Here, supported by first-principles calculations, we propose that Anderson-Blount's ferroelectric-like metal can be manifested in electron-doped ferroelectric Pca2$_1$ HfO$_2$. In this system, polar phonons and consequently ferroelectricity are not affected by the presence of itinerant electrons. We find that a nonpolar optical phonon, being strongly coupled to doped electrons, can acquire a pronounced electron-phonon coupling strength to activate conventional Bardeen-Cooper-Schrieffer superconductivity. The displacements of polar oxygen atoms in Pca2$_1$ HfO$_2$ create a link between ferroelectricity and superconductivity, enabling a tunable superconducting temperature ranging approximately from 10 to 30 Kelvin. Owing to hafnia's compatibility with silicon, we suggest HfO$_2$-based ferroelectric superconductors present an opportunity to construct high-performing hybrid integrated systems utilizing switchable quantum states.

cond-mat.supr-con↗

arXiv2021

Two-dimensional multiferroic metal with voltage-tunable magnetization and metallicity

We design a multiferroic metal that combines seemingly incompatible ferromagnetism, ferroelectricity, and metallicity by hole doping a two-dimensional (2D) ferroelectric with high density of states near the Fermi level. The strong magnetoelectric effect is demonstrated in hole-doped and arsenic-doped monolayer α-In2Se3 using first-principles calculations. Taking advantage of the oppositely charged surfaces created by an out-of-plane polarization, the 2D magnetization and metallicity can be electrically switched on and off in an asymmetrically doped monolayer. The substitutional arsenic defect pair exhibits an intriguing electric field-tunable charge disproportionation process accompanied with an on-off switch of local magnetic moments. The charge ordering process can be controlled by tuning the relative strength of on-site Coulomb repulsion and defect dipole-polarization coupling via strain engineering. Our design principle relying on no transition metal broadens the materials design space for 2D multiferroic metals.

cond-mat.mtrl-sci↗

arXiv2021

On-demand quantum spin Hall insulators controlled by two-dimensional ferroelectricity

The coexistence of ferroelectric and topological orders in two-dimensional (2D) atomic crystals allows non-volatile and switchable quantum spin Hall states. Here we offer a general design principle for 2D bilayer heterostructures that can host ferroelectricity and nontrivial band topology simultaneously using only topologically trivial building blocks. The built-in electric field arising from the out-of-plane polarization across the heterostrucuture enables a robust control of the band gap size and band inversion strength, which can be utilized to manipulate topological phase transitions. Using first-principles calculations, we demonstrate a series of bilayer heterostructures are 2D ferroelectric topological insulators (2DFETIs) characterized with a direct coupling between band topology and polarization state. We propose a few 2DFETI-based quantum electronics including domain-wall quantum circuits and topological memristor.

cond-mat.mtrl-sci↗

arXiv2020

Designing Xenes with Two-Dimensional Triangular Lattice

Xenes, graphene-like two-dimensional (2D) monoelemental crystals with a honeycomb symmetry, have been the focus of numerous experimental and theoretical studies. In comparison, single-element 2D materials with a triangular lattice symmetry have not received due attention. Here, taking Pb as an example, we investigate the triangular-lattice monolayer made of group-IV atoms employing first-principles density functional theory calculations. The flat Pb monolayer supports a mirror-symmetry-protected spinless nodal line in the absence spin-orbit coupling (SOC). The introduction of an out-of-plane buckling creates a glide mirror, protecting an anisotropic Dirac nodal loop. Both flat and buckled Pb monolayers become topologically trivial after including SOC. A large buckling will make the Pb sheet a 2D semiconductor with symmetry-protected Dirac points below the Fermi level. The electronic structures of other group-IV triangular lattices such as Ge and Sn demonstrate strong similarity to Pb. We further design a quasi-3D crystal PbHfO$_2$ by alternately stacking Pb and 1T-HfO$_2$ monolayers. The new compound PbHfO$_2$ is dynamically stable and retains the properties of Pb monolayer. By applying epitaxial strains to PbHfO$_2$, it is possible to drive an insulator-to-metal transition coupled with an anti-ferroelectric-to-paraelectric phase transition. Our results suggest the potential of the 2D triangular lattice as a complimentary platform to design new type of broadly-defined Xenes.

cond-mat.mtrl-sci↗

arXiv2020

Coexistence of nontrivial topological properties and strong ferromagnetic fluctuations in $A_2$Cr$_3$As$_3$ ($A$=Na, K, Rb and Cs)

Superconductivity in crystals without inversion symmetry has received extensive attention due to its unconventional pairing and possible nontrivial topological properties. Using first-principles calculations, we systemically study the electronic structure of noncentrosymmetric superconductors $A_2$Cr$_3$As$_3$ ($A$=Na, K, Rb and Cs). Topologically protected triply degenerate points connected by one-dimensional arcs appear along the $C_{3}$ axis, coexisting with strong ferromagnetic (FM) fluctuations in the non-superconducting state. Within random phase approximation, our calculations show that strong enhancements of spin fluctuations are present in K$_2$Cr$_3$As$_3$ and Rb$_2$Cr$_3$As$_3$, and are substantially reduced in Na$_2$Cr$_3$As$_3$ and Cs$_2$Cr$_3$As$_3$. Symmetry analysis of spin-orbit coupling $g_{k}$ suggests that the arc surface states might remain stable in the superconducting state, giving rise to possible nontrivial topological properties.

cond-mat.supr-con↗

arXiv2018

Tunable Electronic Structure and Topological Properties of $LnPn$ ($Ln$=Ce, Pr, Gd, Sm, Yb; $Pn$=Sb, Bi)

We have performed systematic first principles study of the electronic structure and band topology properties of $LnPn$ compounds ($Ln$=Ce, Pr, Gd, Sm, Yb; $Pn$=Sb, Bi). Assuming the $f$-electrons are well localized in these materials, both hybrid functional and modified Becke-Johnson calculations yield electronic structure in good agreement with experimental observations, while generalized gradient approximation calculations severely overestimate the band inversions. From Ce to Yb, a systematic reduction of band inversion with respect to the increasing $Ln$ atomic number is observed, and $\mathcal{Z}_2$ for Ce$Pn$ and Yb$Pn$ are [1;000] and [0;000], respectively. In both hybrid functional and modified Becke-Johns calculations, a topologically nontrivial to trivial transition is expected around SmSb for the antimonides and around DyBi for the bismuthides. Such variation is related with lanthanide contraction, but is different from simple pressure effect.

cond-mat.str-el↗

arXiv2016

A quantum gravity tensor equation formally integrating general relativity with quantum mechanics

Extending black-hole entropy to ordinary objects, we propose kinetic entropy tensor, based on which a quantum gravity tensor equation is established. Our investigation results indicate that if N=1, the quantum gravity tensor equation returns to Schrodinger integral equation. When N becomes sufficiently large, it is equivalent to Einstein field equation. This illustrates formal unification and intrinsic compatibility of general relativity with quantum mechanics. The quantum gravity equation may be utilized to deduce general relativity, special relativity, Newtonian mechanics and quantum mechanics, which has paved the way for unification of theoretical physics.

physics.gen-ph↗