Select2Adapt: Semantically Constrained Preference Transfer for Single-Sample Text-to-Image Generation

Selecting a preferred image from multiple candidates does not, by itself, teach a generator to produce preferred outputs directly. We introduce Select2Adapt, a framework for transferring multi-candidate quality preferences to single-sample text-to-image generation through frozen-backbone conditioning adaptation. Our central principle is to make score-based semantic eligibility a prerequisite for quality-driven selection, rather than collapse alignment and quality into a scalar reward. An alignment-anchored teacher promotes a quality-preferred alternative only when predefined alignment-score, quality, and cross-template consistency checks pass; otherwise, it retains the alignment-selected reference. The resulting decisions supervise a bounded residual text-conditioning adapter, the only trainable component of an otherwise frozen diffusion model. A selector-supervised objective combines denoising on teacher-promoted and preservation targets with base-prediction consistency and residual regularization, while restricting the adapter’s intervention to the conditional branch. Multi-candidate selection is used to construct training supervision, whereas inference produces a single sample without candidate reranking. To distinguish the value of selection from the effectiveness of transfer, we specify an evaluation protocol that compares selectors on shared candidate pools, uses distinct selection and assessment models, and separately evaluates the adapted generator under single-sample inference. Select2Adapt reframes quality-aware selection from a terminal reranking step into semantically constrained supervision for parameter-efficient text-to-image adaptation.