Authors:

Puruo Wang

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

As part of the CVGM (Computer Vision and Generative Models) group's broader effort to develop a comprehensive evaluation framework for generative models, our sub-team focuses on measuring how well generated images align with their intended prompts. In this preliminary study, we implement a CLIP-based Cross-Modal Maximum Mean Discrepancy (CMMD) metric to assess image-text consistency. The pipeline computes feature embeddings for both prompts and generated images using CLIP, then calculates CMMD against real image references from the COCO-30K dataset. We benchmark four diffusion models-Stable Diffusion v1.4, Stable Diffusion 2.1, SDXL-Turbo, and aMUSEd-across six categories of prompts, including simple objects, complex scenes, and multi-object scenarios. This allows us to evaluate prompt fidelity and visual coherence in a standardized, reproducible manner. Our sub-team's work contributes one component of the larger CVGM initiative: building a reliable, modular evaluation system for generative model research. The methods and metrics established here will support other teams in the group and help guide future model development. Keywords: Generative Models; Evaluation Metrics; Model Benchmarking; Computer Vision

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Puruo Wang