K2

K² · Artificial intelligence

Ask an AI for eight images, get eight versions of the same picture

The authors introduce a method for controlled diversity that enables 'Semantic Browsing', allowing users to navigate structured image galleries through systematic traversal of meaningful, interpretable axes of variation.

SD
MV

Sara Dorfman, Maya Vishnevsky, Omer Dahary et al.

5 authors · cs.CV, cs.AI, cs.GR, cs.LG

arXiv preprintArtificial intelligenceJun 2026 · ~65s read

The 30-second scan

Ask a modern image generator for a "cozy coffee shop" eight times, and you get eight near-identical coffee shops.

  1. The paper identifies that modern text-to-image models sacrifice diversity for prompt adherence, causing generated samples to collapse into a single visual interpretation.
  2. The authors argue that existing methods to improve diversity produce outputs driven by incidental variations rather than meaningful design choices.
  3. The method induces diversity directly at the text level instead of relying on stochastic variation within the text-to-image model.
Read the originalPDF