Computer Animation and Virtual Worlds最新文献

Recent advances in Score Distillation Sampling (SDS) have enabled text-driven 3D human generation, yet the standard classifier-free guidance (CFG) framework struggles with semantic misalignment and texture oversaturation due to limited model capacity. We propose a novel framework that decouples conditional and unconditional guidance via a dual-model strategy: A pretrained diffusion model ensures geometric stability, while a preference-tuned latent reward model enhances semantic fidelity. To further refine noise estimation, we introduce a lightweight U-shaped Swin Transformer (U-Swin) that regularizes predicted noise against the reward model, reducing gradient bias and local artifacts. Additionally, we design a time-varying noise weighting mechanism to dynamically balance the two guidance signals during denoising, improving stability and texture realism. Extensive experiments show that our method significantly improves alignment with textual descriptions, enhances texture details, and outperforms state-of-the-art baselines in both visual quality and semantic consistency.

This study explores how environmental design elements in library spaces influence human psychophysiological responses using virtual reality (VR). Thirty participants experienced VR simulations of library reading areas, with variations in wall color, flooring material, and lighting intensity, whereas electroencephalography (EEG) and galvanic skin response (GSR) measured physiological reactions alongside subjective ratings. Moderate lighting (20,000–30,000 cd) minimized arousal and supported attention, while white walls enhanced relaxation via increased alpha brain activity. Green plant walls slightly boosted attention-related beta activity, and wood flooring was rated highest for comfort and naturalness. VR enabled precise control of design variables, advancing environmental psychology research. These findings offer evidence-based guidelines for designing public spaces like libraries to enhance user well-being and cognitive performance, with implications for educational and public buildings.

This research introduces a novel approach to generate three-dimensional objects using human voice as input for a machine learning model. The spoken input is first converted into text using a Google API, which then guides the creation of the desired three-dimensional object. This approach is particularly beneficial for individuals without design expertise, paving the way for broader participation in the evolving Metaverse—a virtual reality that transcends our physical realm. The three-dimensional object generation model comprises three primary components: Neural Radiance Fields (NeRF), Low-Rank Adaptation (LoRA), and Stable Diffusion. When combined, these components facilitate the creation of a diverse range of three-dimensional objects. Our method presents an innovative approach to harnessing speech recognition for generating three-dimensional objects within the Metaverse, while demonstrating competitive quality and practical efficiency.

Text-driven human motion generation aims to synthesize coherent, natural human actions from natural-language descriptions. While current techniques are effective for short sequences, they often suffer from abrupt transitions, detail loss, and temporal misalignment when extended to long sequences. We present SD-T2LM, a unified framework that couples transition optimization, detail enhancement, and temporal alignment to produce natural, detail-rich long motion sequences. First, a dynamic SLERP interpolation selects transition length conditioned on start/end pose similarity to suppress discontinuities while preserving physical plausibility. Second, a dynamic masking mechanism progressively reduces the mask ratio from 60% to 30% during training, enabling staged learning-global structure early, fine-grained details later. Third, a Temporal Flexible Alignment loss based on Soft-DTW aligns generated and ground-truth rhythms without enforcing point-wise synchrony. On the HumanML3D dataset and KIT-ML dataset, SD-T2LM showed significant improvements in FID and diversity compared to the baseline model, while reducing transition distance and jerk.

Reliable estimation of torso direction is essential for virtual locomotion as it controls the locomotion direction in virtual environments. Existing approaches rely on external sensors to track torso direction or use the poses and/or movements from the head, hands, or feet to control the direction of travel. In addition, deep learning methods based on a head-mounted display (HMD) and its controllers were proposed for torso direction estimation but are unable to produce continuous estimates or that the estimation errors were not systematically evaluated. On the other hand, 3D full-body human motion generation methods based on an HMD and its controllers are capable of estimating torso direction but this capability was largely ignored for its use in virtual steering. We proposed a new method called TCNPoser that can perform torso direction estimation and generate 3D full-body human motion. Through offline evaluation and a user study, we show that the proposed method achieves the state-of-the-art (SOTA) performance for real-time torso direction estimation.

The COVID-19 pandemic has profoundly disrupted traditional face-to-face education, accelerating the shift towards digital learning environments and prompting the adoption of innovative pedagogical approaches. Among these, Augmented Reality (AR) has emerged as a transformative educational tool, offering immersive and interactive platforms that enhance students' understanding of complex concepts. This study investigates the integration of AR technology into university-level beverage preparation courses, with the aim of fostering creativity and strengthening professional competencies. The findings indicate that AR has considerable potential to enhance creative performance, reshape pedagogical practices, and promote self-directed learning. By addressing prevalent learning challenges and providing actionable insights into students' developmental progress, AR supports the design of tailored teaching strategies while cultivating industry-relevant skills. This research highlights the role of AR in advancing inclusive, equitable, and high-quality education, in alignment with Sustainable Development Goal 4 (SDG 4), and underscores its broader significance in redefining educational paradigms in the post-pandemic era.

The convergence of deep learning and the Metaverse represents a pivotal frontier in the evolution of intelligent digital ecosystems. This paper presents a comprehensive survey of how deep learning techniques spanning convolutional, generative, transformer-based, and reinforcement architectures collectively enable perception, creation, cognition, and governance within immersive virtual worlds. Building upon this synthesis, we propose the Deep Learning-Empowered Metaverse Intelligence (DL-MI) framework, which unifies sensory intelligence, generative world-building, adaptive reasoning, and ethical-social governance into a cohesive architecture. The study illustrates how deep learning facilitates realistic avatar synthesis, dynamic environmental rendering, emotion-aware interaction, and predictive personalization, thereby transforming the Metaverse from reactive systems to anticipatory, self-evolving spaces. Key challenges such as data privacy, algorithmic bias, and computational sustainability are critically examined alongside emerging paradigms, including quantum-augmented AI and federated collaboration. By integrating technical, ethical, and societal dimensions, this survey provides a structured foundation for developing scalable, transparent, and human-centered Metaverse intelligence.

The proposed MotionBlend GAN model marks a significant step forward in video synthesis by blending the motion from a source video with the appearance of a target person's image. As training progresses, the model improves video creation by enhancing the smoothness and natural flow of motion, resulting in more coherent and lifelike videos. Using advanced techniques like MoBConv blocks of EfficientNet-B7, OpenPose for precise pose detection, ResNet blocks for feature integration, and a 3D CNN discriminator, the model produces high-quality videos that maintain both spatial and temporal consistency. After 200 epochs, the model achieved an adversarial loss of 0.2265, with metrics like PSNR at 20.246, SSIM at 0.867, and LPIPS at 0.178. The high PSNR and SSIM values, along with the low LPIPS, show that the generated frames are well aligned and preserve important details. These results highlight the model's strong performance over time, consistently generating visually convincing videos of human activities using a reference image and source video. The model effectively transfers motion from video to image, creating realistic videos of human activity in comparison to existing models.