Multi-modal learning has enabled powerful systems that combine text, images, audio, and video for perception, reasoning, and decision-making. At the same time, it has introduced safety challenges that differ fundamentally from those in traditional uni-modal learning. This tutorial presents a structured overview of the evolving safety landscape in multi-modal AI, focusing on emerging threat models and corresponding defense strategies. It examines key risks such as compromised modality integration, modality misalignment, and fused cross-modal vulnerabilities, and reviews recent work on adversarial attacks, jailbreaks, hallucinations, and safety solutions for more reliable multi-modal systems.

Diffusion models have become a cornerstone of modern generative modeling, but their practical deployment in interactive applications is often limited by slow and computationally expensive sampling. This tutorial focuses on recent advances in accelerating diffusion models, providing a comprehensive overview of methods that enable fast and efficient generation. It covers general acceleration techniques, training-based approaches such as distillation into few-step samplers, and practical strategies for scaling to image and video generation. The tutorial further highlights how these advances enable emerging applications such as interactive world models and real-time generative systems, and provides hands-on guidance through the FastGen library.

In recent years, diffusion models have become a central paradigm in computer vision, powering advances in image synthesis, editing, and video generation. However, existing tutorials are often fragmented, focusing either on specific applications or isolated methodological perspectives without a unifying framework. This tutorial aims to present a principle-driven view of diffusion models by distilling their foundations into a small set of core ideas that unify variational, score-based, and flow-based approaches. It further emphasizes emerging directions in real-time generation through flow-map models, which enable fast and interactive visual applications. In addition, the tutorial extends this framework to discrete and tokenized diffusion models, highlighting their role in bridging continuous vision generation with multimodal and structured representations.

Vision-language models are increasingly deployed in real-world systems where images can directly influence decisions and actions, creating new security risks beyond traditional text-based attacks. This tutorial provides a hands-on introduction to attacks and defenses for vision-language systems, using a practical, end-to-end workflow that mirrors real deployment scenarios. It covers a range of vulnerabilities, including visual jailbreaks, preprocessing-induced attacks, adversarial perturbations, backdoored models, and data poisoning, along with corresponding mitigation strategies. Through interactive examples and reproducible notebooks, the tutorial emphasizes how these threats manifest in practice and how to build robust, auditable systems for multimodal AI.

Edge AI enables real-time, low-latency inference directly on devices, but achieving high performance and efficiency requires specialized optimization and deployment techniques tailored to heterogeneous hardware. This tutorial provides a hands-on guide to on-device inference, focusing on end-to-end workflows for optimizing and deploying deep learning models on leading edge platforms such as Qualcomm Snapdragon and NVIDIA Jetson. It covers key techniques including model compression, quantization, and hardware-aware optimization, along with practical tools such as SNPE and TensorRT. Through comparative analysis and real-world case studies, the tutorial highlights best practices for achieving efficient, low-latency performance in applications ranging from computer vision to multimodal AI.

Computer vision for high-stakes, real-world applications necessitates robust explanation and transparency to foster trust, accountability, and ethical deployment. Celebrating its 5th anniversary, the Explainable AI for Computer Vision (XAI4CV) workshop provides a premier forum for the entire spectrum of XAI research, from interpretable-by-design models to the challenges of multimodal foundation models. The program includes three invited talks, spotlight papers, a tutorial on concept-based explanations for the diagnosis and control of vision foundation models, and a poster session. XAI4CV accepts paper and demo submissions aimed at defining the future of trustworthy visual AI.

As deep learning systems are increasingly deployed in high-stakes applications, understanding their internal behavior is essential for ensuring trust, safety, and reliability. However, the field of interpretability remains fragmented, spanning diverse methods without a unified framework or standardized evaluation. This tutorial aims to provide a comprehensive overview of interpretability in vision models, bridging post-hoc mechanistic analysis with approaches that design inherently interpretable models. It reviews techniques for analyzing neural networks at multiple levels—from individual neurons to circuits—alongside recent advances in evaluating the faithfulness of explanations. In addition, the tutorial covers emerging methods for learning interpretable models by design, such as concept-based approaches, and highlights practical applications in debugging, model editing, and safety auditing.

Machine-learned visual systems are transforming numerous fields such as autonomous driving, biodiversity assessment, and ecological monitoring, but they hunger for vast, high-quality annotated data. Asking domain experts to manually annotate large-scale data is unrealistic; the current paradigm to scale up data annotation is to have domain experts craft annotation guidelines using visual examples and descriptions for non-expert annotators to apply. This paradigm is commonly adopted by companies which provide data labeling services. Lacking domain knowledge, ordinary annotators often produce annotations that are erroneous, subjective, biased, and inconsistent. Further, this process is labor-intensive, tedious, and costly. This workshop aims to pioneer auto-annotation, developing AI agents that can interpret expert-crafted annotation guidelines and generate labels automatically. In essence, we seek to replace ordinary human annotators with AI.

MSLR 2026 is the second edition of a rapidly growing venue on multimodal sign language recognition and translation. The program combines invited talks, a peer-reviewed track published in CVPR Workshops, and the SignEval Challenge featuring updated datasets for isolated LIS and continuous SLR. We emphasize privacy-preserving sensing (e.g., radar), healthcare accessibility, and inclusive practices with sign interpreters. Building on the success at ICCV 2025, MSLR 2026 will consolidate a global, interdisciplinary community spanning computer vision, linguistics, healthcare, and Deaf studies.

Partial differential equations (PDEs) play a central role in physics-based modeling across vision, graphics, and robotics, but conventional grid-based solvers often struggle with scalability and complex geometry. This tutorial introduces grid-free Monte Carlo methods for solving PDEs, focusing on algorithms such as walk on spheres and walk on stars that eliminate the need for spatial discretization. It presents the theoretical foundations of these methods alongside practical techniques for efficient sampling, variance reduction, and differentiable simulation. The tutorial also highlights applications in vision and robotics, including inverse problems and physics-based learning, and provides hands-on guidance for implementing Monte Carlo PDE solvers in real-world systems.

World models are emerging as a new paradigm in computer vision and multimodal learning, enabling systems to move beyond perception toward reasoning, simulation, and decision-making. This tutorial explores how world models have evolved from predictive frameworks into engines for multi-modal reasoning, capable of simulating environments, supporting counterfactual thinking, and enabling planning. It examines key approaches for learning world dynamics from visual data, including both discrete tokenization and diffusion-based methods, and highlights their role in modeling physical and causal structure. The tutorial further covers how these models support reasoning through simulation, as well as their applications in embodied agents and robotics, while discussing key challenges such as grounding, scalability, and causal understanding.

Generative AI is transforming simulation for autonomous driving, enabling data-driven and closed-loop environments that better capture the complexity of real-world scenarios. This tutorial presents an end-to-end framework for building generative simulation pipelines tailored to modern learning-based driving systems. It covers key components including world modeling and city-scale digital twins, generative synthesis of rare and safety-critical scenarios, and realistic sensor and video simulation using both graphics and neural approaches. The tutorial further discusses system-level evaluation and integration with autonomous driving stacks, providing practical guidance and open-source tools for developing scalable and reliable simulation environments.

Understanding human pose and shape through parametric body models is a key enabler of applications from AR/VR and sports analysis to human-robot interaction. This tutorial provides an in-depth overview of parametric body models and their role in Human Mesh Recovery. We cover fundamental principles and recent developments, guiding practitioners through major models (e.g., SMPL, Anny, MHR, SOMA) and their trade-offs. We then present state-of-the-art Human Mesh Recovery methods, with a focus on challenging in-the-wild settings across different input modalities, including single- and multi-view RGB, video, depth and LiDAR.