Diffusion-based world models have shown strong potential for unified world simulation, but the iterative denoising remains too costly for interactive use and long-horizon rollouts. While feature caching can accelerate inference without training, we find that policies designed for single-modal diffusion transfer poorly to world models due to two world-model-specific obstacles: token heterogeneity from multi-modal coupling and spatial variation, and non-uniform temporal dynamics where a small set of hard tokens drives error growth, making uniform skipping either unstable or overly conservative. We propose WorldCache, a caching framework tailored to diffusion world models. We introduce Curvature-guided Heterogeneous Token Prediction, which uses a physics-grounded curvature score to estimate token predictability and applies a Hermite-guided damped predictor for chaotic tokens with abrupt direction changes. We also design Chaotic-prioritized Adaptive Skipping, which accumulates a curvature-normalized, dimensionless drift signal and recomputes only when bottleneck tokens begin to drift. Experiments on diffusion world models show that WorldCache delivers up to 3.7× end-to-end speedups while maintaining 98% rollout quality, demonstrating the vast advantages and practicality of WorldCache in resource-constrained scenarios. Our code is released in https://github.com/FofGofx/WorldCache.

Large-scale video diffusion models (V-DMs) have achieved remarkable text-to-video generation quality, yet their massive computational complexity makes deployment costly. Post-Training Quantization (PTQ) offers an appealing route to accelerate inference without retraining, but existing diffusion PTQ methods remain fragile on modern V-DMs. A key reason is that contemporary V-DMs are intrinsically multi-resolution due to multi-stage training, while most prior PTQ pipelines calibrate at a fixed resolution, causing suboptimal calibration signals and biased distributions under resolution changes. To address this gap, we propose S $^{2}$ Q-VDiT $^+$, a multi-resolution co-design PTQ framework from data, supervision, and quantizer perspectives. First, Denoising-Prior Based Multi-Resolution Sampling constructs resolution-consistent noisy latents by mapping to the clean space and re-noising, together with a trajectory-aware resolution policy across timesteps. Second, Structure-Aware Multi-Resolution Distillation enhances structural alignment via window-wise distillation and transfers resolution-aware spatial dependencies via multi-scale attention distillation. Third, Debiased Modulated Quantization mitigates skewed distributions using asymmetric weight quantization and a fuseable activation debiasing scheme. Extensive experiments on multiple state-of-the-art video generation models demonstrate that S$^{2}$ Q-VDiT$^+$ consistently outperforms strong PTQ baselines under W4A6 and W4A4, delivers up to $2.08\times$ end-to-end speedup, and reduces model storage and inference memory by up to $3.8\times$ and $2.1\times$, respectively.

Diffusion models have demonstrated remarkable performance on vision generation tasks. However, the high computational complexity hinders its wide application on edge devices. Quantization has emerged as a promising technique for inference acceleration and memory reduction. However, existing quantization methods do not generalize well under extremely low-bit (2-4 bit) quantization. Directly applying these methods will cause severe performance degradation. We identify that the existing quantization framework suffers from the outlier-unfriendly quantizer design, suboptimal initialization, and optimization strategy. We present MPQ-DMv2, an improved Mixed Precision Quantization framework for extremely low-bit Diffusion Models. For the quantization perspective, the imbalanced distribution caused by salient outliers is quantization-unfriendly for uniform quantizer. We propose Flexible Z-Order Residual Mixed Quantization that utilizes an efficient binary residual branch for flexible quant steps to handle salient error. For the optimization framework, we theoretically analyzed the convergence and optimality of the LoRA module and propose Object-Oriented Low-Rank Initialization to use prior quantization error for informative initialization. We then propose Memory-based Temporal Relation Distillation to construct an online time-aware pixel queue for long-term denoising temporal information distillation, which ensures the overall temporal consistency between quantized and full-precision model. Comprehensive experiments on various generation tasks show that our MPQ-DMv2 surpasses current SOTA methods by a great margin on different architectures, especially under extremely low-bit widths.

Autoregressive video diffusion models enable open-ended generation through local attention and KV caching. However, existing training-free long-video optimization methods mainly focus on stable extension under a single prompt, making them difficult to handle interactive scenarios involving prompt switching, old scene forgetting, and historical scene recall. We identify the core bottleneck as the functional entanglement of historical KV states: stable anchors and recent dynamics are handled by the same cache policy, leading to outdated background contamination, delayed response to new prompts, and loss of long-range memory. To address this issue, we propose Echo-Forcing, a training-free scene memory framework specifically designed for interactive long video generation with three core mechanisms: (1) Hierarchical Temporal Memory, which decouples stable anchors, compressed history, and recent windows under relative RoPE; (2) Scene Recall Frames, which compresses historical scenes into spatially structured KV representations to support long-term recall; and (3) Difference-aware Memory Decay, which adaptively forgets conflicting tokens according to the discrepancy between old and new scenes. Based on these designs, Echo-Forcing uniformly supports smooth transitions, hard cuts, and long-range scene recall under a bounded cache budget. Extensive evaluations on VBench-Long further demonstrate that Echo-Forcing achieves the best overall performance in both long-video generation and interactive video generation settings. Our code is released in https://github.com/mingqiangWu/Echo-Forcing

Unified multimodal object tracking has achieved remarkable robustness by leveraging complementary sensor data (e.g., RGB, Thermal, Depth), yet the heavy computational burden of state-of-the-art models hinders their deployment on resource-constrained edge devices. In this work, we identify the prediction head as a critical but often overlooked efficiency bottleneck. By strategically streamlining the decoder architecture, we unlock the potential for real-time inference but simultaneously introduce a capacity gap between the lightweight student and the heavy teacher. To resolve this, we conduct a systematic analysis of 17 distillation strategies and introduce a Dual-Alignment Distillation framework. Our key insight is that effective compression requires decoupling knowledge transfer into two complementary streams: (1) Spatial Representation Alignment, which employs feature distillation to sharpen the student's spatial focus on foreground targets ("Where to track"); and (2) Semantic Distribution Alignment, which utilizes logit-based distillation to align decision boundaries and transfer discriminative dark knowledge ("What to track"). Extensive experiments across five benchmarks demonstrate that our approach significantly outperforms complex state-of-the-art methods. Notably, our distilled model achieves 91.5

SAM3D enables scalable, open-world 3D reconstruction from complex scenes, yet its deployment is hindered by prohibitive inference latency. In this work, we conduct the first systematic investigation into its inference dynamics, revealing that generic acceleration strategies are brittle in this context. We demonstrate that these failures stem from neglecting the pipeline's inherent multi-level heterogeneity: the kinematic distinctiveness between shape and layout, the intrinsic sparsity of texture refinement, and the spectral variance across geometries. To address this, we present Fast-SAM3D, a training-free framework that dynamically aligns computation with instantaneous generation complexity. Our approach integrates three heterogeneity-aware mechanisms: (1) Modality-Aware Step Caching to decouple structural evolution from sensitive layout updates; (2) Joint Spatiotemporal Token Carving to concentrate refinement on high-entropy regions; and (3) Spectral-Aware Token Aggregation to adapt decoding resolution. Extensive experiments demonstrate that Fast-SAM3D delivers up to 2.67× end-to-end speedup with negligible fidelity loss, establishing a new Pareto frontier for efficient single-view 3D generation. Our code is released in https://github.com/wlfeng0509/Fast-SAM3D.

Learning-based 3D reconstruction models, represented by Visual Geometry Grounded Transformers (VGGTs), have achieved remarkable progress with large-scale transformers. Their prohibitive computational and memory costs severely hinder real-world deployment. Post-Training Quantization (PTQ) has emerged as a common practice to compress and accelerate models. However, we empirically observe that PTQ faces unique obstacles when compressing billion-scale VGGTs: the data-independent special tokens induce heavy-tailed activation distributions, while the multi-view nature of 3D data makes calibration sample selection highly unstable. This paper proposes the first **Quant**ization framework for **VGGT**s, namely **QuantVGGT**. This mainly relies on two technical contributions: First, we introduce *Dual-Smoothed Fine-Grained Quantization*, which integrates pre-global Hadamard rotation and post-local channel smoothing to robustly mitigate heavy-tailed distributions and inter-channel variance. Second, we design *Noise-Filtered Diverse Sampling*, which filters outliers via deep-layer statistics and constructs frame-aware diverse calibration clusters to ensure stable quantization ranges. Comprehensive experiments demonstrate that QuantVGGT achieves the state-of-the-art results across different benchmarks and bit-width, surpassing the previous state-of-the-art generic quantization method with a great margin. We highlight that our 4-bit QuantVGGT can deliver a **3.7$\times$** memory reduction and **2.5$\times$** acceleration in real-hardware inference, while preserving over **98\%** reconstruction accuracy of the full-precision counterparts. This demonstrates the vast advantages and practicality of QuantVGGT in resource-constrained scenarios.

Diffusion transformers exhibit remarkable video generation capability, yet their prohibitive computational and memory costs hinder practical deployment. Model quantization and attention sparsification are two promising directions for compression, but each alone suffers severe performance degradation under aggressive compression. Combining them promises compounded efficiency gains, but naive integration is ineffective. The sparsity-induced information loss exacerbates quantization noise, leading to amplified attention shifts. To address this, we propose **QuantSparse**, a unified framework that integrates model quantization with attention sparsification. Specifically, we introduce *Multi-Scale Salient Attention Distillation*, which leverages both global structural guidance and local salient supervision to mitigate quantization-induced bias. In addition, we develop *Second-Order Sparse Attention Reparameterization*, which exploits the temporal stability of second-order residuals to efficiently recover information lost under sparsity. Experiments on HunyuanVideo-13B demonstrate that QuantSparse achieves 20.88 PSNR, substantially outperforming the state-of-the-art quantization baseline Q-VDiT (16.85 PSNR), while simultaneously delivering a **3.68$\times$** reduction in storage and **1.88$\times$** acceleration in end-to-end inference.

Forecasting high-dimensional chaotic systems, such as weather and turbulence, is fundamentally limited by uncertainty. Traditional deterministic models fail to capture this, collapsing to blurry, physically implausible mean-state predictions and offering no measure of confidence. We introduce Prism, a generative spatiotemporal framework that directly addresses this by learning the probability distribution of future states. Instead of regressing a single outcome, Prism, built on a conditional diffusion model, generates a diverse ensemble of high-fidelity, physically consistent future trajectories. This ensemble enables principled uncertainty quantification and improves the capture of low-probability, high-impact events. On standard benchmarks like WeatherBench and 2D turbulence, Prism significantly outperforms state-of-the-art baselines on probabilistic metrics such as the Continuous Ranked Probability Score (CRPS). Our work establishes generative modeling as a powerful paradigm for building trustworthy AI tools for scientific discovery.

The success of large-scale visual language pretraining (VLP) models has driven widespread adoption of image-text retrieval tasks. However, their deployment on mobile devices remains limited due to large model sizes and computational complexity. We propose Adaptive Multi-Modal Multi-Teacher Knowledge Distillation (AMMKD), a novel framework that integrates multi-modal feature fusion, multi-teacher distillation, and adaptive optimization to deliver lightweight yet effective retrieval models. Specifically, our method begins with a feature fusion network that extracts and merges discriminative features from both the image and text modalities. To reduce model parameters and further improve performance, we design a multi-teacher knowledge distillation framework to pre-train two CLIP teacher models. We decouple modalities by pre-computing and storing text features as class vectors via the teacher text encoder to enhance efficiency. To better align teacher and student outputs, we apply KL scatter for probability distribution matching. Finally, we design an adaptive dynamic weighting scheme that treats multi-teacher distillation as a multi-objective optimization problem. By leveraging gradient space diversity, we dynamically adjust the influence of each teacher, reducing conflicts and guiding the student toward more optimal learning directions. Extensive experiments on three benchmark datasets demonstrate that AMMKD achieves superior performance while significantly reducing model complexity, validating its effectiveness and flexibility.

Knowledge Distillation (KD) aims to improve a light-weight student network supervised by a large teacher network. The core idea of KD is to explore valuable knowledge from the teacher. Previous works often extract information from a single sample, but ignore relation modeling among multiple samples between student and teacher. Therefore, we propose Cross-Layer Graph Knowledge Distillation (CLGKD) that conducts graph-augmented feature and relation distillation assisted by graph neural networks. We further propose a meta-learning mechanism to optimize cross-layer matching weights for promoting GKD among all student and teacher layers. Experimental results on image classification and object detection demonstrate that CLGKD achieves state-of-the-art performance compared to other KD methods. Our code is available at https://github.com/cynmzzz/ICASSP2025-CLGKD

3D point cloud segmentation faces practical challenges due to the computational complexity and deployment limitations of large-scale transformer-based models. To address this, we propose a novel Structure- and Relation-aware Knowledge Distillation framework, named SRKD, that transfers rich geometric and semantic knowledge from a large frozen teacher model (>100M) to a lightweight student model (<15M). Specifically, we propose an affinity matrix-based relation alignment module, which distills structural dependencies from the teacher to the student through point-wise similarity matching, enhancing the student's capability to learn contextual interactions. Meanwhile, we introduce a cross-sample mini-batch construction strategy that enables the student to perceive stable and generalized geometric structure. This aligns across diverse point cloud instances of the teacher, rather than within a single sample. Additionally, KL divergence is applied to align semantic distributions, and ground-truth supervision further reinforces accurate segmentation. Our method achieves state of the art performance with significantly reduced model complexity, demonstrating its effectiveness and efficiency in real-world deployment scenarios. Our Code is available at https://github.com/itsnotacie/SRKD.

Diffusion transformers have emerged as the mainstream paradigm for video generation models. However, the use of up to billions of parameters incurs significant computational costs. Quantization offers a promising solution by reducing memory usage and accelerating inference. Nonetheless, we observe that the joint modeling of spatial and temporal information in video diffusion models (V-DMs) leads to extremely long token sequences, which introduces high calibration variance and learning challenges. To address these issues, we propose **$S^2$Q-VDiT**, a post-training quantization framework for V-DMs that leverages **S**alient data and **S**parse token distillation. During the calibration phase, we identify that quantization performance is highly sensitive to the choice of calibration data. To mitigate this, we introduce *Hessian-aware Salient Data Selection*, which constructs high-quality calibration datasets by considering both diffusion and quantization characteristics unique to V-DMs. To tackle the learning challenges, we further analyze the sparse attention patterns inherent in V-DMs. Based on this observation, we propose *Attention-guided Sparse Token Distillation*, which exploits token-wise attention distributions to emphasize tokens that are more influential to the model's output. Under W4A6 quantization, $S^2$Q-VDiT achieves lossless performance while delivering $3.9\times$ model compression and $1.3\times$ inference acceleration. Code will be available at https://github.com/wlfeng0509/s2q-vdit.

Existing online Knowledge Distillation (KD) often perform probability-based predictions from independent data samples for knowledge transfer. However, these online KD methods neglect valuable relational information across multiple networks. To address this problem, we propose Online Relational Knowledge Distillation (ORKD). ORKD includes a discriminative loss to construct meaningful feature space and a relational distillation loss to guide structured knowledge transfer among multiple networks. Beyond feature-level distillation, we further construct an ensemble teacher by aggregating probability predictions from multiple networks. The virtual teacher is used to supervise a specific network to enhance its accuracy and avoid the cohort homogenization problem. Experimental results on CIFAR-100 and ImageNet classification demonstrate that ORKD achieves the best performance among state-of-the-art online KD methods over various network architectures. The qualitative visualization shows that ORKD can help the network to learn a more discriminative feature space, resulting in better classification performance.

Deep neural networks (DNNs) exhibit noticeable performance degradation when exposed to out-of-distribution samples during testing. This degradation occurs due to the fact that the statistical characteristics (mean and standard deviation) of the features reflect the domain-specific properties of the training data, causing DNNs to be biased towards unseen distributions. One prominent approach to mitigate this bias is training networks with style augmented data, which helps alleviate the reliance on specific style distributions. However, existing methods either have limitations in terms of the number of styles they can handle or rely on intricate pipelines. In this paper, we propose a novel and simple approach called Style Factorization (StylF) to generate novel styles by identifying the directions that exhibit the most significant variations in style. We formulate this problem as a constrained optimization task and decompose the feature statistics matrix of the training data to generate meaningful and diverse novel styles. Through extensive experiments conducted on three publicly available domain generalization benchmarks (PACS, OfficeHome, DomainNet), we demonstrate that our proposed method achieves SOTA performance.

We propose a novel regularization method to effectively train a neural network for avoiding overfitting, thus improving the performance. The core idea is to bridge the gap between predictive distributions derived from two popular image mixture techniques Mixup and CutMix by an ensemble distribution in a class-wise manner. Consistent optimization towards these three distributions is conducted by mutual distillation to guide the model to alleviate over-confidence predictions and robustly learn discriminative features as the classification evidence. Experiments across various image classification tasks show that our method significantly achieves better performance than previous data augmentation Mixup+CutMix and Self-KD methods.

In this report, we describe the technical details of our submission to the EPIC-SOUNDS Audio-Based Interaction Recognition Challenge 2023, by Team "AcieLee" (username: Yuqi\_Li). The task is to classify the audio caused by interactions between objects, or from events of the camera wearer. We conducted exhaustive experiments and found learning rate step decay, backbone frozen, label smoothing and focal loss contribute most to the performance improvement. After training, we combined multiple models from different stages and integrated them into a single model by assigning fusion weights. This proposed method allowed us to achieve 3rd place in the CVPR 2023 workshop of EPIC-SOUNDS Audio-Based Interaction Recognition Challenge.