arXiv subjects

Jiaju Wu

Publications and source records attributed to Jiaju Wu.

3 recordsLinked to original sources

arXiv2026

Shiva-DiT: Residual-Based Differentiable Top-$k$ Selection for Efficient Diffusion Transformers

Diffusion Transformers (DiTs) are costly at high resolution because self-attention scales quadratically with token sequence length. Existing pruning methods do not jointly provide end-to-end learnability, low training overhead, and deterministic token counts for predictable token-dependent computation. We propose Shiva-DiT, based on Residual-Based Differentiable Top-k Selection. Its forward pass executes hard top-k selection, while a residual-aware straight-through estimator propagates gradients to both token scores and the budget k without evaluating a second backbone path. A Context-Aware Router and Adaptive Ratio Policy learn layer- and timestep-dependent retention schedules under a target average budget. Experiments on SD3-Medium, Flux.1-dev, and PixArt-{\Sigma} show consistent reductions in FLOPs and measured latency. On SD3-Medium, Shiva-DiT provides four fidelity-latency operating points and reaches a 1.54x wall-clock speedup with competitive fidelity.

cs.LG↗

arXiv2025

Ultrahigh-Q chiral resonances empowered by multi-head attention deep learning

High quality (Q) factor optical chiral resonators are indispensable for many chiral photonic devices. Designing ultrahigh Q-factors in chiral metasurfaces traditionally relies on extensive parameter scanning, which is time-consuming and inefficient. While deep learning now provides a rapid design alternative, conventional models still face challenges in accurately predicting ultrahigh Q-factor spectral characteristics. In this study, we introduce a multi-head attention network (MuHAN) to accelerate the design of ultrahigh Q-factor optical chiral resonators in bilayer metasurfaces. MuHAN achieves forward spectral predictions in approximately 10ms, thousands of times faster than finite-difference time-domain simulations, boasting 99.85% and 99.9% accuracy for forward and inverse predictions, respectively. By transferring the learned physical principles, we perform inverse design of nanoscale structures with ultrahigh Q-factors (up to 2.9910E5) based on chiral quasi-bound states in the continuum (quasi-BICs) at minimal computational cost. Our rapid design tool, based on MuHAN, enables high-performance encryption imaging, bridging deep learning with high-Q chiral metasurfaces for advanced sensing, laser, and detection applications.

physics.optics↗

arXiv2025

SegQuant: A Semantics-Aware and Generalizable Quantization Framework for Diffusion Models

Diffusion models have demonstrated exceptional generative capabilities but are computationally intensive, posing significant challenges for deployment in resource-constrained or latency-sensitive environments. Quantization offers an effective means to reduce model size and computational cost, with post-training quantization (PTQ) being particularly appealing due to its compatibility with pre-trained models without requiring retraining or training data. However, existing PTQ methods for diffusion models often rely on architecture-specific heuristics that limit their generalizability and hinder integration with industrial deployment pipelines. To address these limitations, we propose SegQuant, a unified quantization framework that adaptively combines complementary techniques to enhance cross-model versatility. SegQuant consists of a segment-aware, graph-based quantization strategy (SegLinear) that captures structural semantics and spatial heterogeneity, along with a dual-scale quantization scheme (DualScale) that preserves polarity-asymmetric activations, which is crucial for maintaining visual fidelity in generated outputs. SegQuant is broadly applicable beyond Transformer-based diffusion models, achieving strong performance while ensuring seamless compatibility with mainstream deployment tools.

cs.CV↗