DATA COLLECTION PLATFORM FOR MULTI-TURN TEXT-TO-IMAGE PREFERENCE LEARNING

Introduction

Human-in-the-loop generative AI systems increasingly rely on interaction data to improve prompting, alignment, and personalization. A platform that captures multi-turn user intent, revised prompts, generated outputs, and user preferences can support research on adaptive text-to-image assistance.

Objectives

·  Design and develop an online data-collection platform with for multi-turn text-to-image interactions.

·  Capture user interactions, prompt revisions, topic labels, and preference signals.

·  Prepare the collected data for future optimization of prompt-rewriting or personalization models.

Tasks

1.     Literature Review: Preference data collection, prompt optimization, human-in-the-loop generative AI.

2.     System Design: Define the session flow, user accounts, consent, logging schema, ML backend, and storage model.

3.     Implementation: Build a web-based platform where users submit prompts, inspect optimized prompts, view generated images, and continue or stop a session.

4.     Annotation Schema: Collect topic labels, preference judgments, and interaction outcomes.

5.     Pilot Study: Run a small user study and analyze data quality and usage patterns.

6.     Optional Extension: Add continual learning for adaptive user personalization.

Pre-requisites

Web development, Python, Databases, Interest in HCI/AI.

Work

20% Theory, 60% Programming/System Development, 20% Writing

Contact

Ali Sabzi Khoshraftar (a.sabzikhoshraftar@utwente.nl)