Fatih Porikli tells Sam Charrington that realistic-looking generated images can still fail to follow a requested number of people, preserve distinct identities or achieve high resolution efficiently. He groups the remaining challenges around controllability, quality and deployment efficiency.
Porikli describes research using reinforcement learningReinforcement learning trains an AI system to choose actions using feedback about the results of its earlier choices. objectives to encourage distinct faces within an image and across runs while retaining count accuracy and image quality. He explains how a benchmarkA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions. and face comparisons help evaluate that diversity.
A second approach separates an architect that plans where subjects should appear from an artist model that renders the image. Porikli discusses evaluating the result for prompt alignment, composition and identity retention, and distinguishes offline training from inference.
The conversation also covers generating higher-resolution images within limited memory using upsamplingUpsampling increases the number of samples in a signal or image, such as producing an image with more pixels. and latent-space refinement. Porikli describes how guided noise and blending aim to limit visible seams, and discusses a related inpaintingInpainting fills or regenerates a selected region of an image while using the surrounding image as context. approach intended to reduce boundary artifacts.
Watch on YouTube
