Developed a Creative Text Prompt-Based Relighting AI Model by Professor Baek Seung-ryeol's Team
Change Photo and Video Color Tones Without Distorting the Original Image...
Selected for AAAI 2025
An artificial intelligence that creates lighting effects for photos and videos using only text input has been developed.
Without using complex editing tools, it is now possible to easily adjust the color tones of photos or videos, capturing the emotional nuances of language such as "hot crispy chicken" or "cool blue light."
The AI modifies the color tones and other aspects of portrait photos based on creative text commands.
Professor Baek Seung-ryeol's team at the Graduate School of Artificial Intelligence, UNIST, developed an AI model called "Text2Relight" that changes lighting effects in portrait photos and videos based on creative text commands.
This research, conducted in collaboration with Adobe, was accepted by the Association for the Advancement of Artificial Intelligence (AAAI), one of the top three AI conferences, and will be presented at the 2025 annual conference held in Philadelphia, USA, starting on the 25th.
The developed AI model has the advantage of expressing various lighting characteristics such as color tone, brightness, and emotional atmosphere through creative natural language text.
It also adjusts the color tones of both the subject and the background simultaneously without distorting the original image. Existing text-based image editing AI models were not specialized for lighting data, resulting in distortion of the original image or limited lighting adjustments.
The research team created a large-scale synthetic dataset to enable the AI to learn the correlation between creative text and lighting. They generated lighting data using ChatGPT and text-based diffusion models, and built a large-scale synthetic dataset capable of learning various lighting conditions by applying the OLAT technique and Lighting Transfer.
Additional auxiliary training data such as shadow removal and lighting position adjustment were also trained to enhance visual consistency and lighting realism.
Professor Baek Seung-ryeol explained, "The Text2Relight technology has great potential in content fields by reducing work time in photo and video editing and enhancing immersion in virtual and augmented reality."
This research was led by researcher Cha Jun-wook of the Graduate School of Artificial Intelligence at UNIST as the first author.
The research was conducted with support from Adobe and the Ministry of Science and ICT.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
