Reviving Deep Dream with modern LLMs
An experimenter revived Deep Dream by optimizing images to maximize target prompt probability in modern LLMs, finding that Gemma 12B—lacking a vision encoder—treats pixels like tokens, which sparked discussion of SAE feature maximization and internal representations.
2026-07-20 ~ 2026-07-20 · 2 related posts
- Deep Dream Visualizations for Modern LLMs — matthen2 · 2026-07-20
- Deep Dream and SAE Feature Maximization in Modern LLMs — torchcompiled · 2026-07-20