Beijing Film Academy

Swarm UAV autonomous flight for Embodied Long-Horizon (ELH) tasks is crucial for advancing the low-altitude economy. However, existing methods focus only on specific basic tasks due to dataset limitations, failing in real-world deployment for ELH tasks. ELH tasks are not mere concatenations of basic tasks, requiring handling long-term dependencies, maintaining embodied persistent states, and adapting to dynamic goal shifts. This paper presents U2UData+, the first large-scale swarm UAV autonomous flight dataset for ELH tasks and the first scalable swarm UAV data online collection and algorithm closed-loop verification platform. The dataset is captured by 15 UAVs in autonomous collaborative flights for ELH tasks, comprising 12 scenes, 720 traces, 120 hours, 600 seconds per trajectory, 4.32M LiDAR frames, and 12.96M RGB frames. This dataset also includes brightness, temperature, humidity, smoke, and airflow values covering all flight routes. The platform supports the customization of simulators, UAVs, sensors, flight algorithms, formation modes, and ELH tasks. Through a visual control window, this platform allows users to collect customized datasets through one-click deployment online and to verify algorithms by closed-loop simulation. U2UData+ also introduces an ELH task for wildlife conservation and provides comprehensive benchmarks with 9 SOTA models.

Generative AI is experiencing high-paced iterations and is steadily transforming the operational paradigms of embedded systems on the computational level. This transformation depends on reconfigured neural structure customised as per the requirements of real-time data processing. The algorithmic enhancements of the Stable Video Diffusion (SVD) model combined with the Audio2Face tool have been thoroughly studied by researchers to obtain a perfect control of the facial semantics of virtual agents, but to pay attention to the performance of computational efficiency and the actual utilisation efficiency of hardware devices. Stable Video Diffusion (SVD) has the capability to produce highly realistic dynamic sequences using given text specifications; Audio2Face has the capability to perform an exact match of synchronisation of acoustic animations with contextually relevant semantics. The hybrid architecture of the research team consisting of various modules incorporates spatiotemporal transformers and ControlNet adapters to provide micro-expression control of virtual agents, thus reaching high fidelity of images and low data processing latency of data processing in various embedded devices. The model training phase relies on the FER2013 data, combined with MEAD, primarily aiming at the quality of the emotion recognition and the accuracy of the mapping of the AU intensity, with variational autoencoders and perceptual loss functions making specific modifications to the latent space in order to minimise the likelihood of the occurrence of distortion phenomena, including temporal discrepancies. According to various test results, the model is accurate to the level of 0.85, which surpasses the majority of similar benchmark models and costs the overall processing workflow a reduction of 20-30 percent due to a reasonable distribution of GPUs. This structure reduces bottlenecks in the process, maximises the reality of content presentation during virtual content previsualization, rendering, and post-processing phases, and is scaled to different applications of embedded AI. Problems like leakage of identity information, content defects can also be resolved in the solution attaining scalable models supporting real-time functionality, and capable of adjusting to various application conditions, including mobile communication and edge computing. Ablation experiments confirm the soundness of the model that emphasises the profound fusion implementation of embedded technologies and AI innovation in the digital realm.

This study takes the children's animation IP “Brave Brother Ah Niu” as a case study, where the animation IP refers not merely to a single work, but rather to a core content asset capable of multi-dimensional development and possessing long-term commercial value, such as a well-known character, story, or worldview. This case is used to explore the path for children's animation IP to empower the deep integration of cultural tourism based on Bourdieu's field theory. In response to the current challenges of fragmentation, superficial symbolization, and capital feedback disruption in development, a three-field transformation framework of “cultural production-media practice-industrial consumption” is constructed, and an integration mechanism of “empowerment-translation-feedback” is proposed. This study provides a replicable theoretical model and practical path for the Ah Niu IP and similar regional children's animation IPs to overcome the bottleneck of cultural tourism integration, achieve cultural inheritance, and enhance the industrial value.

This paper delves into the audio-video synchronization generation mechanism of Google Veo 3 within the field of computer vision, emphasizing its role in producing synchronized video clips aligned with audio inputs. Veo 3 employs diffusion models to achieve high-fidelity video synthesis, incorporating generative adversarial networks (GANs) along with audio embeddings to boost lip-sync precision. We introduce a novel synchronization algorithm that features a lip attention mechanism designed to alleviate lip-sound discrepancies, leveraging Veo 3’s diffusion backbone enhanced by attention modules and a regression head for accurate synchronization scoring. The evaluation of this model is conducted on the Lipreading Dataset sourced from Kaggle, revealing outstanding performance in key metrics like F1-score and PSNR. Results from experiments indicate that the proposed approach attains over 90% synchronization accuracy, marking a 5-10% enhancement over baseline methods, primarily attributed to the attention mechanism. This methodology not only refines alignment in generated multimedia but also substantiates quantitative benchmarks such as PSNR surpassing 30 dB and F1-score exceeding 0.85, guaranteeing strong generalization capabilities. Ultimately, this research advances AI-driven multimedia generation, especially in contexts demanding exact audio-visual coherence, thereby contributing significantly to the evolving landscape of generative technologies.

In the era of information explosion and increasingly fierce brand competition, champion image endorsement has become one of the important strategies for enterprises to enhance their brand image and their market influence. Champions, as the top figures in their respective fields, they not only have excellent skills and extraordinary achievements, but also carry the spiritual core of perseverance, the courage to challenge and the pursuit of excellence. These characteristics are exactly the brand values that enterprises are eager to convey to consumers. As the spokesperson, the sports stars have a unique charm and influence. They not only have a wide fan base but also have a high image and credibility in the public mind. The fighting spirit, competitive level, and healthy image of sports stars can often inject more positive energy and trust into the brand. Therefore, choosing sports stars as spokesmen has a significant effect on improving brand recognition and exposure. This paper is committed to studying the relationship between the characteristics of sports star spokesmen and brand identity and brand exposure, analyzing the commercial activities and commercial value of star endorsement of CHAGEE, and studying the spiritual/material pursuit of consumers, which will provide a direction for businesses to choose spokesmen.

Liveness, as the core audience experience of live performance, is a valuable tool for evaluating the quality of both live and mediatized performance. Since most of the current research work on liveness mainly focus on media design and there is no widely-accepted definition of liveness, we propose a conceptual framework of liveness of theater performance. The conceptual framework defines the key factors of liveness, which are collective engagement, collective memory, performance-audience interactivity, audience copresence and theater presence. To verify this conceptual framework, a comprehensive experiment was designed and carried out to evaluate the level of liveness of a live performance and two media contents based on this live performance. The live performance was recorded and replayed in the form of the traditional video and the panoramic video. The liveness level of these three different viewing modes of performance are measured by the electroencephalography (EEG) and explicit methods including the questionnaire, interview and focus group. By comparison, we found that people in the virtual reality viewing mode have a higher sense of presence but the collective community is not as well as the traditional video group. We hope that the proposed framework could help the digital media community to improve the audience experience with the emerging media technology.

Abstract The creation of the dance drama film A Tapestry of a Legendary Land (zhici qinglv, 2024) was inspired by the painting Thousand Miles of Rivers and Mountains (qianli jiangshangtu). Combining artistic expression means such as dance and film, it promotes the beauty of Chinese culture. Based on the aesthetics in the Huizong period of the Northern Song Dynasty, the film renders the main color tone of “green and blue,” creating a new realm for the modern presentation of the aesthetics of the Song Dynasty on the screen. The film also makes a vivid and elaborate display of splendid chapters like “painting study” and “listening to the rain” with the help of Cai Jing’s postscript and the lifestyle preference of ancient scholars for “listening to the rain,” conveying the concept of the symbiosis of Tian, Di, Ren (heaven, earth and humanity) contained in Thousand Miles of Rivers and Mountains.