AI-based computer vision assistance can be revolutionary for visually impaired people, as well as for a broad group of other users. The client interface can be highly varied depending on user needs, environmental factors, and available infrastructure. In this paper, we present a qualitative comparison of two client architectures for an AI scene description system built upon an ESP32-CAM and a Gemma-4B multimodal AI backend. The first is a server-hosted mobile app using React Native and FastAPI, designed for portability and aural feedback for users with low vision. The second is a direct-access desktop app using PyQt6, offering a live preview and direct interaction for development or sighted assistance. We compare these clients across feature sets, deployment options, network reliance, latency considerations, and user experience trade-offs. This comparison outlines the pros and cons of server-intermediated and direct-access models, offering insights into client strategies for AI vision systems. While this study is architectural and qualitative, the results highlight the significance of designing adaptable interfaces and lay the groundwork for future quantitative benchmarking and user-centered evaluations.