Spotting and Reducing Hallucinations in AI Image Upscaling

​Generative super-resolution (GSR) currently sets the state of the art in terms of perceptual image quality, overcoming the “regression-to-the-mean” blur of prior non-generative models. However, a different class of artifacts, in which generated details fail to perceptually match the low-resolution image (LRI) or ground-truth image (GTI), is a critical but under-studied issue in GSR, limiting its practical deployment. In this talk, I will focus on measuring, analyzing, and mitigating these artifacts (i.e., “hallucinations”). More details here: https://arxiv.org/abs/2507.14367

​We’ll touch upon:
• Analyzing hallucinations in existing diffusion-based image super-resolution approaches
• Measuring hallucinations using multimodal large language models (MLLMs) such as GPT-4o, which closely align with human evaluations
• Mitigating hallucinations by leveraging certain deep feature distances based on DINO and CLIP, and aligning GSR models using such features as differentiable reward functions

About the speaker:
Raghav Goyal is a researcher at Samsung AI Center Toronto, working on camera-related technologies. He has been involved in AI and computer vision research for the past 10 years across both academia (PhD at UBC) and industry (research internships at big tech, video understanding at a startup)
LinkedIn: /raghav-goyal-1b312029
Webpage: https://www.cs.ubc.ca/~rgoyal14/