The evaluation of content quality in newly generated images is conducted through the construction of a test instruction set, developed via two primary approaches. First, creative prompts are solicited from user groups with relevant experience through professional research platforms, thereby collecting a substantial volume of text-to-image generation requests that closely mirror real-world application scenarios. Second, frequently used prompts from leading global image generation platforms are referenced and subsequently refined and optimized in accordance with assessment requirements to serve as a supplementary component of the core instruction set. In the formulation of these instructions, comprehensive coverage of thematic content and stylistic diversity is ensured, encompassing not only fundamental categories such as natural landscapes and biological subjects but also a wide range of artistic expressions and commercial design needs.

In the domain of safety and responsibility, the evaluation framework is constructed based on a range of internationally recognized safety assessment datasets. The associated test instruction set addresses several critical dimensions, including but not limited to: discrimination and bias (e.g., racial or gender-based discrimination), illegal activities (e.g., terrorist acts, unlawful surveillance), harmful or explicit content (e.g., depictions of violence or pornography), ethical considerations (e.g., animal cruelty, vandalism), copyright infringement, and violations of privacy or portrait rights. These prompts undergo rigorous professional screening and optimization to ensure that they comprehensively represent the spectrum of potential risk scenarios associated with AI image generation.

In designing the test content, particular attention is paid to the delineation of safety boundaries across diverse cultural contexts and social norms, encompassing both explicit violations and more subtle manifestations of value-laden biases.

The construction methodology for the test instruction set pertaining to image revision tasks aligns with that employed for new image generation. A portion of the prompts is sourced from online users, while others are adapted by research assistants based on pre-existing image-prompt pairs. The instructions span three primary categories: object-specific revisions, scene-level modifications, and stylistic transformations.