Growing Images: Spatial Scheduling in Diffusion Inpainting

Diffusion inpainting implicitly treats the order in which masked pixels are denoised as irrelevant. We argue in this manuscript that it is not. Holding the model, prompt, and per-patch compute budget fixed in Stable Diffusion inpainting, we show that progressively filling a mask from its boundary measurably shifts reconstruction, perceptual, and distributional metrics relative to standard parallel denoising. We study this analytically by recasting inpainting as \emph{finite-depth approximate inference}. Specifically, we consider a block-structured Blume--Capel spin system: a diffusion-like stochastic interpolant tilts the prior into a time-dependent Gibbs posterior with block-dependent random fields. In this picture, schedule-induced bias turns into a predictable consequence of finite-depth inference.