Prior Subspace Recomposition: A Unified Prior Initialization Framework for Conditional Generation

Initial noise plays a crucial yet underexplored role in continuous conditional generative models. While existing golden-noise methods demonstrate that carefully selected or optimized noises can improve generation quality, they are often tailored to specific generation regimes or predefined quality objectives, limiting their adaptability across tasks and evaluation criteria. In this paper, we propose Prior Subspace Recomposition (PSR), a unified framework that reformulates noise initialization as metric-induced prior discovery followed by condition-aware anchor recomposition. Given a downstream evaluation suite, PSR first empirically identifies a metric-induced superior region in the Gaussian prior space and distills it into a reusable set of semantic-agnostic prior anchors. To handle complex conditions, PSR further estimates concept-conditioned anchor fitness and converts multi-dimensional metric responses into relative anchor preferences, which supervise a lightweight local ranker based on first-step denoising velocity fields. At inference time, local ranking predictions are resolved into a global anchor ordering through graph-based optimization, and the top-ranked anchors define a condition-specific local recomposition region on the Gaussian prior shell. The final initialization is obtained via geometry-preserving spherical averaging, integrating complementary anchor priors while preserving the prior distribution geometry. Extensive experiments across text-to-image, text-to-video, and image-to-video generation demonstrate that PSR consistently improves generation quality without modifying the underlying generative backbones.