ARMD: Adaptive Regional Mixture of Diffusers for Hallucination-Reduced Creative Upscaling

Creative upscaling occupies a distinct regime between faithful restoration and unconstrained image synthesis:users often want to preserve the large-scale identity of an input image while still allowing diffusion models to inventricher local structure, texture, and detail. In practice, tiled diffusion methods remain necessary for high resolutions,yet they expose an unresolved trade-off. Independent tile pipelines often suffer from visible seams, color drift,or unstable structural agreement, whereas shared-canvas methods derived from Mixture of Diffusers preservecontinuity more effectively but can become semantically unstable when a single global prompt is applied uniformlyto heterogeneous image regions. We present ARMD (Adaptive Regional Mixture of Diffusers), a training-freeregional conditioning layer for creative upscaling that combines shared-canvas tiled denoising with per-regionprompts, optional vision-language-model captioning, weighted accumulation, and optional spatial control payloads.The core idea is simple: keep the strong inter-tile communication of Mixture of Diffusers-style generation, butreplace prompt uniformity with region-aware conditioning. We formulate the method independently of anyparticular node graph or software package, discuss its practical implementation in open workflow environmentssuch as ComfyUI, and evaluate it through qualitative comparison figures and a blank-canvas multi-promptdemonstration. The current evidence focuses on high-denoise creative upscaling with SDXL and shows thatregional conditioning can reduce local hallucinations without reintroducing visible seams or tile-to-tile color drift.