Image editing often requires making numerous small yet significant changes, a process that can be tedious and time-consuming for artists or even frustrating for consumers who may lack experience with image editing tools. The advent of region- and maskbased diffusion models offers a promising solution to this challenge by enabling efficient, intuitive, and user-friendly image modifications. This research investigates state-of-theart inpainting models, focusing on their evaluation under specific input conditions and contexts. To be able to functionally evaluate these models across various scenarios and contexts, a tool was developed to streamline the process. It is possible to define multiple regions on an image, assign unique prompts to each region, and process these inputs through different diffusion and inpainting technologies. The resulting edits are merged into copies of the original image for side-by-side evaluation. Evaluation is performed using several proposed techniques from different papers to be able to statistically determine the overall best image quality given the context and model used. In order to calculate accurate metrics for evaluation, the COCO dataset from Microsoft is used, as it provides a variety of annotated images that include masks, the category images correspond to and several image captions.
Object Replacement: Evaluation of State-Of-The-Art Diffusion and Inpainting Models
Serban, R. M. (Author). 2025
Student thesis: Master's Thesis