Workshops and Tutorials
Workshop
Speaker: Xinyu Zhang (The University of Adelaide); Lingqiao Liu (The University of Adelaide); Chang Xu (The University of Sydney); Yujun Cai (The University of Queensland)
Anton van den Hengel (The University of Adelaide); Dong Gong (The University of New South Wales)
Title: Visual Generative Models: Past, Current and Future
Abstract
Recent breakthroughs in generative adversarial networks, diffusion and autoregressive models have dramatically advanced the state of visual content generation, including widespread applications in generating images, videos, 3D objects, and more. These advancements not only push the frontiers of synthesis quality and scalability but also unlock new applications in design, entertainment, vision, scientific domains, and even improving or reformulating the vision tasks. However, several fundamental and practical challenges remain, e.g., improving controllability, enhancing fidelity and realism, scaling across modalities, ensuring alignment with human values, and achieving efficient, safe deployment. This workshop aims to provide a broad forum for exploring the past breakthroughs, current developments, and future directions of visual generative models, with particular emphasis on foundational innovations, emerging challenges, and practical applications.
Tutorial 1
Speaker: Feng Liu (University of Melbourne and RIKEN AIP); Zesheng Ye (University of Melbourne)
Title: Neural Network Reprogrammability: A Unified Theme on Model Reprogramming, Prompt Tuning and Prompt Instruction
Abstract
This tutorial introduces neural network reprogrammability as a unifying framework that coherently organizes the fragmented landscape of parameter-efficient adaptation research. We identify the common principle underlying diverse approaches: the inherent sensitivity of neural networks to input perturbations can be systematically harnessed through strategic information manipulation and output alignment to induce desired behaviors for new tasks, without architectural or parameter changes. We also present a taxonomy that examines model reprogramming, prompt tuning, and prompt instruction through this integrative lens, uncovering adaptation principles that generalize across techniques, modalities, and applications.
Tutorial 2
Speaker: Zhi Chen (The University of Southern Queensland); Jingcai Guo (The Hong Kong Polytechnic University)
Title: On the Element-wise Representation and Reasoning in Zero-shot Recognition
Abstract
The tutorial will offer a comprehensive review of traditional attribute-based ZSR methods and transition into recent trends emphasizing fine-grained, element-wise interactions between visual and semantic spaces. We will discuss research advances, covering topics such as alignment strategies, interpretability, and recent neural architectures tailored for cross-modal matching. Emphasis will be placed on practical understanding.