Decoupling Global Structure and Local Refinement: Blueprint-Guided Scroll Generation with Direct Preference Optimization

ICLR 2026 Conference Submission1917 Authors

Keywords: Long Scroll Generation, Preference-Optimization, Text-to-Image Generation

Abstract: Existing methods for generating long scroll images, often fail to maintain global structural and stylistic consistency, resulting in artifacts like content repetition. To address this, we propose the Dual-Resolution Scroll Generation with Preference Optimization (DRSPO) framework. Our approach decouples global composition from local refinement by first generating a low-resolution (LR) blueprint to establish a coherent overall structure. This LR blueprint then guides a high-resolution (HR) feature to render fine-grained details. We further enhance generation quality by incorporating Direct Preference Optimization (DPO) at both stages, and we introduce a novel theoretical adaptation to apply preference tuning directly to the region-based generation process. Experimental results demonstrate that our method produces high-quality long scroll images with reasonable global structure and fine-grained details.

Primary Area: generative models

Submission Number: 1917