NobleBlocks

Research indexPreprint

Perceptual Priority Computing (PPC)

Jha

1 Jan 2026Open access

NobleIDark:/48914/w1/20261105/BE4CEE3CDOI10.17605/osf.io/vf83d

Abstract

Current image-generation systems allocate equal computational effort to every pixel simultaneously, requiring the entire scene to complete before the user receives any visual feedback. This paper introduces Perceptual Priority Computing (PPC) and its primary realization, Saliency-Weighted Progressive Diffusion (SWPD) — a staged generation architecture that decomposes a text prompt into a perceptual hierarchy and generates visual elements in order of human attention priority. The primary subject is generated first using a lightweight single-step diffusion model and displayed to the user immediately. The background environment is then generated as a separate conditioned inpainting pass, built around the subject's pose, lighting, and spatial position, and progressively revealed as generation completes. On CPU-only hardware, this approach reduces perceived latency by approximately 10× without increasing total computation. The paper also proposes PPC as a general computational philosophy with promising extensions to language reasoning, video generation, and other computationally intensive AI tasks.

Discussion

More

Suggest a correction or report