Dreamspace

Dreamspace is an infinite canvas for visually exploring outputs from large language models. Users can place prompt nodes on the canvas, run them with different models (text and image), and compare the resulting generations side by side. Outputs appear as nodes that can be inspected to view details or linked to show relationships. The tool supports chaining by continuing from any message in a conversation thread or forking prompts to use different text or models. It allows using previous outputs as context for new prompts, making iteration and experimentation straightforward. The interface is designed for direct manipulation: clicking places nodes, arrow keys move them, and delete removes them. The demo showcases image generation with models like DALL-E 3 and laion-ai/ongo, showing execution times and visual outputs.

Key Benefits

• Visual canvas for comparing prompt outputs side by side
• Supports chaining and forking of prompts for iterative experimentation
• Integrates multiple models including DALL-E 3 and open-source models
• Direct manipulation interface for quick iteration