Why Image Generation Needs More Than Bigger Models with Fatih Porikli

In 2024, you can ask almost any AI system for a picture and get something that looks remarkably good. But "looks good" and "is correct" are not the same thing. Ask for several different people, and the model may generate variations of the same face. Ask for a specific number of subjects, and it may ignore that detail. Push toward higher resolution, and speed and memory become constraints. Fatih Porikli, Vice President of Technology at Qualcomm, argues that the next frontier in image generation is closing the gap between plausible images and precise, controllable results. His team presented more than twenty papers at CVPR 2024, several of which tackle this problem by rethinking objectives and architectures rather than just scaling up.

Diversity as an Explicit Training Objective

A core observation driving Porikli's work is that existing text-to-image models have not really learned to create truly distinct identities. When asked to generate a group of people, the model often duplicates the same face. As Porikli puts it, the problem is not image quality—pixel-wise, the output looks realistic—but a missing objective. "Existing training objectives focus heavily on realism and matching the user prompt, but they don't explicitly encourage diversity between people."