arXiv subjects

Xingzhong Hou

Publications and source records attributed to Xingzhong Hou.

5 recordsLinked to original sources

arXiv2026

LookME: Lookup-Based Multimodal Embeddings for Layer Injection in Vision-Language Models

Vision-Language Models (VLMs) have achieved strong progress in multimodal understanding. However, scaling dense or sparse Mixture-of-Experts (MoE) models to improve performance limits deployment in resource-constrained environments due to the trade-off between high memory usage from full loading and increased latency from on-demand loading. Recently, the Per-Layer Embedding (PLE) architecture addresses this by scaling models with large external embedding tables stored in read-only memory (ROM) and performing lightweight lookup to retrieve relevant embeddings to enhance token representations. Nevertheless, existing PLE-style methods are primarily designed for text embeddings due to the convenience of ID-based retrieval, limiting their effectiveness in VLMs where multimodal embeddings contain richer information for visual tasks. In this paper, we propose LookME, the first framework that enables lookup-based enhancement for multimodal embeddings in VLMs while supporting partitioned storage and on-demand loading. To efficiently lookup arbitrary continuous multimodal embeddings from large-scale embedding tables, we propose a hierarchical two-level lookup method employing a coarse-to-fine strategy that performs lookups from the scene-level to the intra-scene primitive-level. Furthermore, we integrate the lookup method with a sparse injection strategy, which adaptively prioritizes critical embeddings over voluminous multimodal embeddings within layers, and facilitates embedding table reuse across neighboring layers, improving the trade-off among efficiency, model size, and performance. Experiments on multiple visual benchmarks show that LookME outperforms text-only PLE-style methods, validating the effectiveness of lookup-based multimodal embedding enhancement.

cs.CV↗

arXiv2025

Towards Seamless Borders: A Method for Mitigating Inconsistencies in Image Inpainting and Outpainting

Image inpainting is the task of reconstructing missing or damaged parts of an image in a way that seamlessly blends with the surrounding content. With the advent of advanced generative models, especially diffusion models and generative adversarial networks, inpainting has achieved remarkable improvements in visual quality and coherence. However, achieving seamless continuity remains a significant challenge. In this work, we propose two novel methods to address discrepancy issues in diffusion-based inpainting models. First, we introduce a modified Variational Autoencoder that corrects color imbalances, ensuring that the final inpainted results are free of color mismatches. Second, we propose a two-step training strategy that improves the blending of generated and existing image content during the diffusion process. Through extensive experiments, we demonstrate that our methods effectively reduce discontinuity and produce high-quality inpainting results that are coherent and visually appealing.

cs.CV↗

arXiv2024

See Further When Clear: Curriculum Consistency Model

Significant advances have been made in the sampling efficiency of diffusion models and flow matching models, driven by Consistency Distillation (CD), which trains a student model to mimic the output of a teacher model at a later timestep. However, we found that the learning complexity of the student model varies significantly across different timesteps, leading to suboptimal performance in CD.To address this issue, we propose the Curriculum Consistency Model (CCM), which stabilizes and balances the learning complexity across timesteps. Specifically, we regard the distillation process at each timestep as a curriculum and introduce a metric based on Peak Signal-to-Noise Ratio (PSNR) to quantify the learning complexity of this curriculum, then ensure that the curriculum maintains consistent learning complexity across different timesteps by having the teacher model iterate more steps when the noise intensity is low. Our method achieves competitive single-step sampling Fréchet Inception Distance (FID) scores of 1.64 on CIFAR-10 and 2.18 on ImageNet 64x64.Moreover, we have extended our method to large-scale text-to-image models and confirmed that it generalizes well to both diffusion models (Stable Diffusion XL) and flow matching models (Stable Diffusion 3). The generated samples demonstrate improved image-text alignment and semantic structure, since CCM enlarges the distillation step at large timesteps and reduces the accumulated error.

cs.CV↗

arXiv2021

An Efficient and Balanced Graph Partition Algorithm for the Subgraph-Centric Programming Model on Large-scale Power-law Graphs

The subgraph-centric programming model is a promising approach and has been applied in many state-of-the-art distributed graph computing frameworks. However, traditional graph partition algorithms have significant difficulties in processing large-scale power-law graphs. The major problem is the communication bottleneck found in many subgraph-centric frameworks. Detailed analysis indicates that the communication bottleneck is caused by the huge communication volume or the extreme message imbalance among partitioned subgraphs. The traditional partition algorithms do not consider both factors at the same time, especially on power-law graphs. In this paper, we propose a novel efficient and balanced vertex-cut graph partition algorithm (EBV) which grants appropriate weights to the overall communication cost and communication balance. We observe that the number of replicated vertices and the balance of edge and vertex assignment have a great influence on communication patterns of distributed subgraph-centric frameworks, which further affect the overall performance. Based on this insight, We design an evaluation function that quantifies the proportion of replicated vertices and the balance of edges and vertices assignments as important parameters. Besides, we sort the order of edge processing by the sum of end-vertices' degrees from small to large. Experiments show that EBV reduces replication factor and communication by at least 21.8% and 23.7% respectively than other self-based partition algorithms. When deployed in the subgraph-centric framework, it reduces the running time on power-law graphs by an average of 16.8% compared with the state-of-the-art partition algorithm. Our results indicate that EBV has a great potential in improving the performance of subgraph-centric frameworks for the parallel large-scale power-law graph processing.

cs.DC↗

arXiv2020

Fast and Efficient Parallel Breadth-First Search with Power-law Graph Transformation

In the big data era, graph computing is widely used to exploit the hidden value in real-world graphs in various scenarios such as social networks, knowledge graphs, web searching, and recommendation systems. However, the random memory accesses result in inefficient use of cache and the irregular degree distribution leads to substantial load imbalance. Breadth-First Search (BFS) is frequently utilized as a kernel for many important and complex graph algorithms. In this paper, we describe a preprocessing approach using Reverse Cuthill-Mckee (RCM) algorithm to improve data locality and demonstrate how to achieve an efficient load balancing for BFS. Computations on RCM-reordered graph data are also accelerated with SIMD executions. We evaluate the performance of the graph preprocessing approach on Kronecker graphs of the Graph500 benchmark and real-world graphs. Our BFS implementation on RCM-reordered graph data achieves 326.48 MTEPS/W (mega TEPS per watt) on an ARMv8 system, ranking 2nd on the Green Graph500 list in June 2020 (the 1st rank uses GPU acceleration).

cs.DS↗