Quantization for Video Diffusion Models at the Edge; Technology vs. Experience in the Premium Hearing Aid Market 3 views

Deploying diffusion-based video generation models on mobile devices poses major challenges due to computational and memory limits. To enable efficient deployment, the Diffusion Transformer (DiT) architecture was enhanced with 4-bit weight quantization and mixed-precision activation strategies. The design introduced quantization-aware training, per-head quantization scales for attention, and layer-specific precision constraints, benchmarked on the iPhone 16 Pro Max. Results show that selective higher-precision activations preserve generative quality while reducing latency, making advanced video generation feasible on widely available consumer hardware. This reframes efficiency as a means of accessibility, not just optimization. Hearing aid manufacturers typically equate device value with technical complexity, while users judge it through lived experience. Drawing on a 2016 clinical trial showing no significant outcome difference between basic and premium devices, the findings reveal that added features do not guarantee satisfaction; effectiveness depends on how technology aligns with user needs. The study reminds system designers that typical performance metrics are incomplete and therefore unreliable proxies for users’ experience.

BS (Bachelor of Science)

generative media; mobile computing; machine learning systems; user experience; assistive technology

School of Engineering and Applied Science Bachelor of Science in Computer Science Technical Advisor: Rosanne Vrugtman, Brianna Morrison STS Advisor: Peter Norton

English

All rights reserved by the author (no additional license for public reuse)

2025-12-15