In recent years, technologies for creating aerial displays, such as drone light shows, have garnered significant attention. However, drone light shows have the disadvantage of requiring many drones and being expensive. To overcome this disadvantage, our group had proposed a cost-effective method for aerial displays using a spherical screen suspended by drones and a high-brightness projector. However, the previous methods faced several challenges in achieving large-scale, long-distance projections. This study proposes a new method for projecting onto large spherical screens at long distances by developing a new infrared marker structure and a processing method. This method uses highspeed vision to detect the position of a moving screen using the developed IR marker and project images precisely at that position to create an aerial display. The gaze direction of the high-speed camera and the projection direction of the projector was controlled with a high-speed optical gaze controller. The novel arrangement of the two IR light sources enabled dynamic projection onto larger screens at long distances. To validate the proposed system, we conducted a field experiment. In this evaluation, a large spherical display with a diameter of 2.3 meters was suspended from a drone and maneuvered at a distance of 100 meters from the projection equipment. The system successfully demonstrated stable projection onto the moving screen.
We present a lightweight and interpretable framework for analyzing anime-style facial expressions using sparse, semantically meaningful feature-line annotations. By deforming annotated contours such as eyelids, eyebrows, and lips into a canonical frontal pose, our method removes pose variation while preserving expressive differences. The canonicalized space enables consistent expression clustering and transfer across views, providing a practical foundation for stylized face normalization in anime production.
This study connects visual rhetoric with computational and quantitative methods to reveal the cultural norms embedded in film poster design across languages and regions. We constructed a large-scale dataset of over 100,000 posters from The Movie Database (TMDB), covering foreign films released between 1960 and 2024 that also had Japanese versions. Using image processing techniques, we extracted visual features such as colour, text, faces, and objects, and conducted cross-linguistic comparisons. The analysis revealed that Japanese posters display substantially more visual information than their original counterparts. By introducing quantitative and statistical perspectives to a field that has largely relied on case studies, this research provides new academic contributions.
We present a Neural Radiance Cache (NRC) integrated into the global illumination and reflection passes of a hybrid renderer targeting mobile devices. Real-time Monte Carlo rendering on mobile remains challenging due to limited compute and memory bandwidth, despite the emergence of hardware-accelerated ray tracing on mobile GPUs. Ray tracing workloads often exhibit incoherent memory access patterns that can bottleneck GPU performance. In contrast, our NRC implementation shifts this cost toward predictable, compute-centric neural inference. Our design features a fused MLP architecture implemented entirely in compute shaders, along with an efficient path-traced training pipeline decoupled from inference. While introducing more overhead than traditional radiance cache techniques, our mobile-optimized NRC achieves higher perceptual quality while maintaining real-time frame rates across a range of scenes on modern mobile hardware.
The cooperative project The Theater Machine develops an interactive, AI-based installation for the foyer of a theater. It aims to lower barriers to performing arts by enabling participatory engagement with theatrical content. Users interact via camera and touch-screen, being integrated in real time into dynamically generated theater scenes. The system operates locally, using a high-performance workstation running a generative model based on Adversarial Diffusion Distillation within a high-throughput StreamDiffusion pipeline. This paper outlines the conceptual framework, system architecture, interaction design, and sustainability model.
With the advancement of embodied interaction technologies, traditional musical instruments are increasingly viewed as interactive media for digital cultural expression. However, the symbolic dimensions of lesser-known traditional instruments such as the Suona, often remain difficult to externalize through conventional auditory-focused interfaces. This paper presents When Birds Bow, an interactive audio-visual installation that reimagines collaborative tangible interaction with the Suona into a narrative of birds offering feathers and the phoenix's rebirth. By mediating cultural translation through embodied, multisensory engagement, the system fosters collective participation and cultural resonance. Preliminary results suggest that this metaphor-driven interaction model enhances user engagement and deepens understanding of the Suona's symbolic significance, offering a new framework for reimagining traditional musical interfaces as vessels for cultural storytelling.
We present a filming process with synchronized cameras and lights which allows a scene to be photographed in two lighting conditions nearly simultaneously with a global shutter cinema camera. The technique enables novel applications for relighting, matting, and direct-indirect reflectance component separation for both technical and creative fields.
This paper presents a VR system for neuroanatomy education targeting challenges in spatial complexity, knowledge integration, and long-term retention. The system combines immersive VR with AI-assisted visualizations and gamified interactions to encourage multimodal engagement with anatomical structures. A small formative probe suggests potential benefits for spatial understanding and motivation. Future work will conduct a systematic evaluation of educational effectiveness and explore lightweight personalization.
Unmanned aerial vehicles (UAVs) enable scalable capture of cultural landmarks, yet monocular video alone remains geometrically ambiguous and difficult for language models to interpret. We present AeroVis3R, a unified framework that couples geometric reconstruction with multimodal reasoning. Our system reconstructs dense 3D structures from UAV videos and distills them into compact geometry representations that can be jointly processed with visual and textual inputs. To support research in this direction, we collect a new dataset of 92 landmark videos paired with automatically aligned textual descriptions. By integrating explicit geometry into the reasoning process, AeroVis3R achieves geometry-grounded multimodal understanding that bridges UAV video capture, 3D reconstruction, and language-based analysis. The dataset will be made publicly available upon acceptance to encourage further advances in geometry-aware multimodal learning.
Video colorization remains challenging to produce vivid and semantically appropriate colors with temporal consistency. Existing diffusion approaches typically learn in an RGB-based latent space, which bridge a domain mismatch for downstream colorization applications. In this work, we present ChromaFlow, a novel framework for video colorization that adheres to the fundamental colorization formulation: directly predicting chrominance components conditioned on luminance. Our key innovation is a DCT-based latent compression mechanism in the YCbCr space that naturally decouples luma and chroma components, significantly reducing computational cost while maintaining perceptual quality aligned with colorization objectives. ChromaFlow employs a Diffusion Transformer with linear attention and asynchronous flow matching, enabling efficient processing of long video sequences and natural color propagation from reference frames. Extensive experiments demonstrate that ChromaFlow achieves superior performance in color accuracy, temporal coherence, and controllability compared to state-of-the-art methods, while supporting conditioning modalities including text prompts and reference frames. Project page: https://chroma-flow.github.io.
This work pioneers the application of computational acoustic simulation to assistive jewelry design for hearing enhancement. While acoustic modeling has been extensively used in traditional hearing aid development and architectural acoustics, its application to jewelry-scale assistive devices represents an unexplored design space. We present the first systematic methodology integrating acoustic simulation into the design process for assistive earwear jewelry, examining how material selection influences sound propagation in our originally designed Green Voice Smart Jewelry Device. Through computational analysis of four jewelry-grade metals-silver, titanium, platinum, and white gold-we demonstrate how acoustic simulation can inform both aesthetic and functional decisions in assistive jewelry design, establishing a new paradigm where computational acoustics guides the creation of hearing assistance devices that users want to wear.
This paper presents a workflow that reimagines the traditional CGI pipeline for product visualization by leveraging FLUX. 1 Kontext and WAN 2.2 diffusion models. Our approach streamlines the creation of high-quality product videos, eliminating time-intensive 3D workflow while seamlessly merging visual elements and harmonizing lighting and aesthetics. This enables rapid look development and iterative pre-visualization, offering faster turnaround and greater creative flexibility than conventional render-based workflows.
Sparse NanoVDB volumes are increasingly used for interactive effects and scientific visualization. However, fast primary-visibility (first-hit) computation typically requires either compute-centric pipelines with multi-draw indirect and transient geometry buffers, or pure ray marching/HDDA with significant empty-space work. We introduce a rasterization-centric pipeline that maps the NanoVDB hierarchy directly onto modern task and mesh shaders, eliminating intermediary buffers while tightly coupling visibility culling with shading. A task shader expands only potentially visible Level 2 Upper Internal nodes into candidate Level 1 Lower Internal nodes for the current view; a mesh shader materializes just the faces likely to contribute to the frontmost G-buffer hit, passing each tile's world-space bounds to the fragment stage to shorten per-pixel traversal. To further reduce empty-space stepping, we build an acceleration structure over Level 1 and use it to jump HDDA between nodes; the same structure is queried with Vulkan ray queries for shadows and selective skipping. Our implementation requires no modifications to NanoVDB data (only runtime auxiliaries such as the tile AS and per-frame bounds) and runs on commodity GPUs. Across a diverse set of scenes, including open Vox and VDB datasets, our approach achieves 1.5x end-to-end speedups on average over vertex-shader baselines while preserving first-hit G-buffer fidelity. We argue that task-mesh pipelines provide a practical, intermediary-buffer-free pathway to interactive first-hit NanoVDB rendering suitable for integration into existing Vulkan engines.
This study presents a new example of visual expression on water surfaces, realized by controlling the positions of drifting objects-such as aquatic plants and fallen leaves-through bubble-induced flow. As a proof-of-concept prototype, we built a 30 cm square water tank equipped with sixteen pumps, and successfully demonstrated the display of a single letter using floating plants.
We present an augmentation method that makes Large Language Models (LLMs) both data aware and visually aware. Unaugmented LLMs can provide high-quality information about the broad context of a visualization, but are unaware of the visual content and thus cannot provide accurate, visualization-specific answers. We address this limitation by providing LLMs with structured metadata generated from a combination of extracted visual information and textual descriptions. Our LLM-agnostic approach preprocesses the visualization to extract features from it using a vision model, combines them with textual information about the data, and generates a compact JSON file that then augments the LLM during user interactions. This highly structured file provides the LLM with the necessary multimodal context without requiring any fine-tuning or costly multimodal prompts, and applies to any existing prerendered visualization paired with descriptive text. We demonstrate our method using geospatial datasets from the Science On a Sphere project. A user study confirms our system's accuracy and appeal to users.
This study tackles construction labor shortages by introducing a robotic brick assembly system using a mobile manipulator. It bridges CAD models and real-world construction through a Design-to-Construction workflow. The proposed Graph Q-Learning (GQL) framework enhances automated brick assembly in complex designs by overcoming the limitations of traditional graph search methods. A Graph Attention Network + Deep Q-Network (GAT-DQN) architecture dynamically models inter-brick relationships, prioritizing structurally critical connections to optimize assembly placement and improve overall stability.
School Legacy is an interactive Extended Reality (XR) experience that transforms personal school memories into a shared, AI-augmented virtual classroom. By scanning a physical memory book, users unlock digital artifacts and generate 3D "time capsules" from personal photos using generative AI. These capsules serve as emotional keys to enter a multiplayer VR environment where alumni reconnect through voice, gesture, and nostalgic rituals-such as decorating graduation caps and signing digital yearbooks. The system blends tangible interaction, real-time co-presence, and AI-driven content creation to create a living archive of youth, turning private nostalgia into a collective, embodied experience.
Deep learning-based segmentation models have revolutionized matte extraction, yet their rasterized outputs are often ill-suited for the precise, vector-based workflows of visual effects artists. Compositors frequently discard these intermediate mattes because they lack the editability and scalability inherent to spline-based representations. While bespoke models can be trained for specific productions to minimize artist touch-ups, the creation of large-scale, high-quality matte datasets remains a significant bottleneck, imposing prohibitive costs and time commitments on studios. Introducing RotoShop, a novel framework that automatically converts raster segmentation sequences into production-ready, spline-based animations. Our key idea is to leverage existing pose estimation models to intelligently rig a set of canonical 2D splines to the subject. We then introduce a differentiable splining process that optimizes the control points of these splines to precisely match the segmentation masks. This optimization is enabled by a differentiable vector rasterizer, which allows us to directly learn the spline representation from pixel-based data. The resulting animated splines can be efficiently stored and interpolated between keyframes, significantly reducing data overhead and providing artists with an intuitive starting point for final adjustments. RotoShop bridges the gap between raster-based deep learning and vector-based visual effects workflows. By automatically generating editable, animated splines from segmentation data, our method drastically reduces the manual effort required for creating high-quality segmentation datasets.
