Title: DetailMaster: Can Your Text-to-Image Model Handle Long Prompts?

Abstract:

Although recent Text-to-Image (T2I) models demonstrate remarkable proficiency in generating images from concise descriptions, they often falter when faced with the extensive, intricate prompts necessary for professional workflows. To address this gap, we introduce DetailMaster, a robust benchmark designed to assess T2I performance on lengthy, complex prompts, supported by an automated data creation pipeline and a structured evaluation framework. Our benchmark features prompts validated by experts, averaging 284.89 tokens in length, and evaluates models across four essential dimensions: Character Attributes, Structured Character Locations, Multi-Dimensional Scene Attributes, and Spatial/Interactive Relationships.

Our assessments of both general-purpose and long-prompt-optimized models highlight significant performance bottlenecks. Specifically, we find that weaker encoders fail to maintain syntactic dependencies within prompts, while diffusion models are prone to attribute leakage when handling detailed inputs. Furthermore, a controlled ablation study under varying constraints demonstrates that achieving high-fidelity generation necessitates a combined approach of increased prompt limits and specialized long-prompt training. To encourage advancements in long-prompt-driven T2I generation, we have open-sourced both our dataset and codebase.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC