Offline Ambient-Controlled Latent Diffusion: Architecture, Telemetry, and On-Device Evaluation

Details

Most on-device image generation research focuses on speed or quality; this paper focuses on provenance. The researchers built an Android latent diffusion app that runs without a cloud connection, using the device's ambient light sensor as its prompt instead of a text input. Each generated image is bound to the sensor reading, runtime path, and seed that produced it — creating a per-artifact audit trail for offline analysis.\n\nThe system runs on a single Samsung foldable under the Android Neural Networks API (NNAPI). Across 373 artifacts, the log-lux ambient input correlates positively with output luminance (Pearson r=0.532, 95% CI [0.455, 0.601]), confirming the sensor signal survives the full denoising pipeline and VAE decoding. Latency ranges from 552 to 1334ms across three quality tiers.\n\nThe contribution is architectural rather than algorithmic: the authors are not proposing a new diffusion method but demonstrating a telemetry-first, offline-first design pattern. The ambient sensor input keeps generation deterministic and auditable from a known environmental signal — an unusual constraint that has potential applications in edge AI contexts where auditability matters.\n\nFor practitioners interested in edge AI or privacy-preserving generation, this demonstrates what current mobile hardware can achieve with a full latent UNet/VAE pipeline without cloud dependency. The limitation is a single-device measurement (373 artifacts); generalization to diverse hardware is untested.