OKAFOR, PRINCE CHARLES (2026) DESIGN AND IMPLEMENTATION OF AN AI-BASED TEXT-TO-IMAGE GENERATION SYSTEM. Other thesis, GODFREY OKOYE UNIVERSITY, ENUGU STATE..
Full text not available from this repository.Abstract
The present project aimed at designing and implementing an AI-powered text-to-image generation system called Artifex. Currently available text-to-image generators are limited by their opacity, incompleteness, and possible security issues. To tackle such issues, a multi-layered client-server architecture has been developed to encompass four levels – presentation (frontend in React), application (backend in Node.js), model (inference in Python with the use of pre-trained Stable Diffusion XL model), and data (database in PostgreSQL). A three-level content moderation approach including banned words detection, NSFW detection with CLIP algorithm, and embedding analysis was applied to provide for responsible usage of the system, as well as a rate-limiter policy of 10 generations per hour per user was implemented to ensure the equitable distribution of computing resources. Given that no model fine-tuning is used, the evaluation process is focused on integration-related indicators, including performance, efficiency, and user experience. The combined SDXL model provided the Fréchet Inception Distance (FID) of 31.8 and the CLIP score of 0.295. On average, the inference process for 512*512 images took 6.2 seconds while the system accuracy of content moderation was 97% and system availability reached 99.8%. The results of user testing were very positive, receiving on average a rating of 4.3 out of 5, as well as an 82 out of 100 SUS score, showing a highly usable system. All research goals have been met. Conclusions drawn from the study include the viability of a modular architecture for text-to-image generators, the appropriateness of using a pre-trained SDXL model for academic applications, achieving accuracy in automatic moderation, and a correlation between user satisfaction and usability. Suggestions for future research are expanding stylistic variation, better understanding abstract ideas, using cloud computing, and evaluating new generations like SD3/Flux.
| Item Type: | Thesis (Other) |
|---|---|
| Subjects: | Q Science > Q Science (General) |
| Divisions: | Faculty of Computing And Information Technology (FACIT) |
| Depositing User: | Nnenna Ayo |
| Date Deposited: | 11 Aug 2026 14:48 |
| Last Modified: | 11 Aug 2026 14:48 |
| URI: | http://eprints.gouni.edu.ng/id/eprint/6246 |
Actions (login required)
 |
View Item |
