DreamBench++ is a benchmark for personalized image generation evaluation, positioned as a "Human-Aligned" evaluation solution. According to public information, it focuses on personalized image generation tasks, collects diverse images and prompts, and uses GPT-4o for automated evaluation, making scoring results consistent with human preferences.
Main capabilities and approach: The information shows that its dataset is diverse, and provides a comparison with DreamBench's image distribution (t-SNE visualization) as well as the distribution of images and prompts. In the evaluation process, humans first write meta-prompting information, including task description, scoring criteria, scoring range, and format specifications; these instructions are then used to prompt GPT-4o for internal thinking, generating a self-aligned task summary and plan; finally, all prompts, reasoning outputs, and image samples are combined to output scores. The information also mentions that the evaluation results involve "Human Alignment Degree."
Usage: The information provides entry points to the project page and leaderboard (https://dreambenchplus.github.io/#leaderboard），但未提供具体的安装命令、代码示例或调用步骤，因此这些内容暂未提供。
Applicable scenarios: Suitable for research and development scenarios that need to evaluate personalized image generation models and want evaluation metrics close to human preferences.
Limitations: The information does not specify fees, pricing, supported platforms, specific model coverage, or runtime environment requirements; this information is not yet provided; its automated evaluation depends on GPT-4o, and related availability and costs need to be confirmed separately.
