MLMC 2026 Technical Program

Abstract: The rapid advancement of generative AI (GenAI) has transformed the creation of images and videos, enabling unprecedented levels of visual realism, diversity, and controllability. Yet one pressing issue is in how to properly assess the quality of AI-generated visual content as such new types of contents poses significant challenges that extend well beyond the scope of traditional image and video quality assessment. Whereas conventional approaches have largely focused on perceptual distortions such as blur, noise, compression artifacts, and transmission degradation, assessing generated contents calls for a broader spectrum of evaluation framework—one that have to consider semantic fidelity, temporal coherence, physical plausibility, aesthetic quality, and exact alignment with human expectations and prompts. This keynote will examine the transition from traditional perception-driven quality assessment to a new paradigm for AI-generated visual content, in which human perceptual experience needs to be integrated with principled understanding of semantics, motion, and the physical world. We shall highlight some recent efforts toward interpretable and comprehensive evaluation frameworks that assess not only visual fidelity, but also motion realism, multidimensional quality, relational semantic consistency, and physical plausibility. This talk will also discuss potential future directions for next-generation GenAI quality assessment, emphasizing the critical intersection of perception and principles as the foundation for designing trustworthy generative visual systems.

Abstract: Images captured in real-world environments often suffer from diverse degradation patterns governed by underlying physical processes, such as motion blur, haze, rain, and low illumination, which lead to undesired contrast loss and appearance distortions. With the rapid development of deep generative image models, numerous image restoration methods have been proposed to effectively alleviate these degradations. As most deep restoration approaches rely on supervised learning, their performance strongly depends on the diversity and representativeness of the training data. However, for many real-world degradations—such as rain, haze, and motion blur—collecting paired degraded and clean images is expensive and often impractical, severely limiting the coverage of available training datasets. These data acquisition challenges restrict the effectiveness and generalization ability of existing restoration models. In this talk, we will present how physics-based models can be leveraged both to enrich training datasets with realistic degradation distributions and to be integrated into diffusion-based restoration frameworks. By incorporating physics-informed priors, the performance of generative restoration models can be substantially enhanced. Representative results will be demonstrated on image dehazing and deblurring tasks.

Abstract: This talk examines how AI is progressively challenging human sustainability. We use Quality of Experience (QoE) as an interpretative lens to highlight new paradigm in AI evaluation. Facing aging society, AI can be seen as technological solutions. In that context, I first trace the evolution from imitation learning to foundation models, and then to world models and embodied AI, where human behavior is increasingly abstracted into data, patterns, and simulated environments. Beyond human sustainability in that context, I also illustrate what QoE sciences can bring for Machine Experience. Second, I present the emergence of a token-based cognitive economy, and question its ressources debt and possible QoE metrics to mitigate it. Then I present the notion of AI psychological debt. AI does not only optimize tasks, it reconfigures the structure of human experience and cognitive participation. Building on this, I propose QoAI (Quality of AI Experience) as an extended framework for AI systems evaluation. QoAI aims to assess not only efficiency and accuracy, but also human agency, cognitive effort, dependency formation, attentional integrity, and long-term cognitive resilience. The QoAI Experience Dashboard operationalizes these dimensions to make visible the hidden experiential costs of AI systems. Ultimately, the question is no longer only how well AI performs, but what kind of human cognition it sustains or erodes over time.

Abstract: Building robust multimodal AI hinges on elevating static low-level perception into dynamic reasoning systems equipped with structured cognition and self-reflection. This report summarizes our series of research along this trajectory. First, to build a precise visual perception engine, we propose CTNet to resolve pixel-level ambiguity in complex scenes. Furthermore, through Singular Value Fine-tuning (SVF) and the VRP-SAM framework, we achieve open-world generalization with minimal parameter cost and visual reference prompts, establishing a powerful "visual specialist" foundation. However, Multimodal Large Language Models (MLLMs) frequently suffer from hallucinations due to the lack of precise descriptions regarding fine-grained visual attributes and object relations. To address this, we introduce the EDC framework, which leverages the aforementioned visual specialists to finely extract target attributes and transform them into high-quality image-text descriptions, significantly enhancing the visual cognition of MLLMs. Finally, targeting the pain point that large models often reason de novo and repeatedly make the same mistakes, we develop ViLoMem, a dual-stream memory framework that separately encodes logical reasoning errors and visual perception traps. This research achieves a technical leap from precise perception engines to the elimination of multimodal cognitive hallucinations, ultimately evolving into interactive reflective agents equipped with semantic memory and continual learning capabilities.