PRISM: Precise region-aware instance style manipulation via text-guided test-time optimization
Junhao Chen, Peng Rong, Xiang Li, Jingbo Sun, Boran Zhang, Hao Zhao, Ruqi Huang, Fei Ma, Qi Tian · Pattern Recognition · 2026
Image style transfer is a core task in computer graphics and multimedia, yet recent text-guided stylization and image editing systems still offer limited control over which instance is stylized and often perturb non-target regions. We address this gap with PRISM , a test-time optimization (TTO) framework for text-guided, instance-level style transfer that stylizes only a user-specified object while preserving the background with pixel-level fidelity. PRISM is designed for the practical setting in which an instance mask is available from a text-guided segmenter or user interaction but remains imperfect at fine boundaries. To make such masks usable for high-fidelity stylization, we introduce two core components: (1) Potential Target Region Recovery (PTR) , which formulates under-segmentation correction as a maximum-a-posteriori crop classification problem in CLIP feature space with dual-threshold voting and a dynamically updated target prototype; and (2) a boundary-aware background consistency loss that combines pixel-level MSE with gradient-domain regularization to suppress seams at object boundaries. We further contribute PRISMBench , a benchmark comprising 248 images, 100 style terms, and 24,800 evaluation triplets, together with both CLIP-based and CLIP-independent metrics (DISTS and ImageReward) for rigorous instance-level stylization assessment. Extensive experiments against 15 baselines spanning TTO, diffusion, and VLM paradigms show that PRISM reduces background LPIPS by 64% relative to CLIPstyler, improves DISTS by 25%, and achieves the highest pairwise user preference (72.3% over CLIPstyler, p < 0.001 ) in a 50-participant study. Code and benchmark will be released upon publication.