In 2025, Large Language Model (LLM) services have launched a new feature-AI video chat-allowing users to interact with AI agents via real-Time video communication (RTC), just like chatting with real people. Despite its significance, no systematic study has characterized the performance of existing AI video chat systems. To address this gap, this paper proposes a comprehensive benchmark with carefully designed metrics across three dimensions: quality, latency, and internal mechanisms. Using custom testbeds, we further evaluate five mainstream AI video chatbots with this benchmark. This work provides the research community a baseline of real-world performance and identifies unique system bottlenecks. In the meantime, our benchmarking results also open up several research questions for future optimizations of AI video chatbots.

Modern real-time communication (RTC) demands both ultra-low latency and consistently high visual quality. Yet, as content becomes more dynamic and RTTs shrink, we reveal a previously overlooked problem: long-tail queuing latency in the sender's pacing queue between encoder and network. This phenomenon is rooted in a mismatch between the bursty frame stream produced by the encoder and the smooth traffic expected by the network. Existing approaches trying to smoothen the bitrate inevitably force an undesirable trade-off between latency and video quality. To address this, we propose a dual-control approach that manages both the encoding and transmission burstiness. At the sender, we dynamically adjust the bucket size of a token-based pacer to control burstiness at the granularity of frame level. Within the encoder, we introduce an adaptive complexity mechanism that smoothens frame sizes without sacrificing quality. Trace-driven emulation and real-world experiments show our solution ACE reduces end-to-end 95th percentile latency by up to 43% while maintaining superior visual quality versus the state of the art.

Facial information is widely used in security fields like identity authentication. But the large number of facial images online makes them vulnerable to unauthorized capture, posing privacy and security risks. Existing face privacy protection methods aim to mitigate these risks. However, many of these methods lack reversibility, making it impossible to restore the original face when needed. Additionally, they often neglect model intellectual property (IP) protection, leaving methods vulnerable to unauthorized stealing. Therefore, to address the shortcomings of existing face privacy protection methods in IP protection, this paper proposes an invertible privacy protection mask network with IP protection. The proposed method consists of two main parts: facial privacy protection and IP protection. For facial privacy protection, the mask generator replaces facial features with other faces and generates the mask, which is then embedded with the watermark to generate the watermarked mask. This watermarked mask conceals the original face by the putting on mask network, and the original face can be restored by the putting off mask network. For IP protection, the watermark extractor network is a key component that can extract the watermark from images of the sender, receiver and attacker to verify the method’s IP. Experimental results show that the proposed method has good effects in both privacy protection and IP protection, providing double security for face privacy protection.