Deep Dream Visualizations for Modern LLMs

matthen2 · x · 2026-07-20

The author explores generating "Deep Dream" style images using modern large models by optimizing the input image to maximize the probability of a specific prompt.

Comparisons reveal that Gemma 12B (which lacks a vision encoder and treats pixels as token embeddings) generates recognizable objects around the canvas. In contrast, E4B (equipped with a vision encoder) primarily exhibits texture drift.

Related event: Reviving Deep Dream with modern LLMs(2 posts)→

Original post →

More from Research

Research channel →