🏢 Tsinghua University

GaussianCube revolutionizes 3D generative modeling with a structured, explicit radiance representation, achieving state-of-the-art results using significantly fewer parameters.

Gaussian Graph Network (GGN) revolutionizes novel view synthesis by efficiently generating generalizable Gaussian representations from multi-view images, achieving superior rendering quality with fewe…

Near-optimal algorithms solve convex simple bilevel problems by reformulating them into functionally constrained problems, achieving near-optimal convergence rates.

Researchers developed Full Distance Attack (FDA) to generate adversarial patterns effective against pedestrian detectors across all distances, resolving the appearance gap issue between simulated and …

PepGLAD, a novel generative model, revolutionizes full-atom peptide design by leveraging geometric latent diffusion to significantly enhance peptide diversity and binding affinity.

Diffusion models, while excelling in image generation, are vulnerable to data poisoning. This paper demonstrates a BadNets-like attack’s effectiveness against diffusion models, causing image misalign…

FlowTurbo: Blazing-fast, high-quality flow-based image generation via a velocity refiner!

FedLPA: One-shot federated learning with layer-wise posterior aggregation improves model accuracy in non-IID data by efficiently aggregating layer-wise posteriors of local models using a novel approac…

Deep state space models (SSMs) gain adversarial robustness through an adaptive scaling mechanism, improving performance without overfitting issues.

SNELL: Sparse tuning with kerNElized LoRA achieves state-of-the-art parameter-efficient fine-tuning performance with drastically reduced memory usage.

Researchers introduce object-aware backdoors in Vision-and-Language Navigation, enabling malicious behavior upon encountering specific objects, demonstrating the vulnerability of real-world AI agents.

Event-3DGS: First event-based 3D reconstruction using 3D Gaussian splatting, enabling high-quality, efficient, and robust 3D scene reconstruction in challenging real-world conditions.

Revolutionizing protein mutation effect prediction, this work introduces a retrieval-augmented framework achieving state-of-the-art accuracy by efficiently incorporating similar local structure inform…

EfficientNAT: a novel approach to token-based image synthesis boosts performance and slashes computational costs by cleverly disentangling and optimizing spatial-temporal interactions between image to…

Elucidating Dataset Condensation (EDC) achieves state-of-the-art accuracy in dataset condensation by implementing soft category-aware matching and a smoothing learning rate schedule, improving model t…

DuQuant: Dual transformations distribute outliers for stronger quantized LLMs.

DCRL, a Dual Critic Reinforcement Learning framework, effectively mitigates high variance in reinforcement learning under partial observability by synergistically combining an oracle critic (with full…

Doubly Mild Generalization (DMG) improves offline reinforcement learning by selectively leveraging generalization beyond training data, achieving state-of-the-art results.

DiTFastAttn: A post-training compression method drastically speeds up diffusion transformer models by cleverly reducing redundancy in attention calculations, leading to up to a 1.8x speedup at high re…

DistDiff, a training-free data expansion framework, leverages distribution-aware diffusion models to generate high-fidelity, diverse samples that enhance downstream model performance.