Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines
This provides insights into the text encoder component of text-to-image pipelines, which is incremental as it analyzes existing models without proposing new methods.
The paper tackles the problem of interpreting how text encoders in text-to-image diffusion models process prompts, proposing the Diffusion Lens method to analyze intermediate representations by generating images from them. Results show that complex scenes are composed progressively and more slowly than simple ones, and uncommon concepts require more computation than common ones, with knowledge retrieval being gradual across layers.
Text-to-image diffusion models (T2I) use a latent representation of a text prompt to guide the image generation process. However, the process by which the encoder produces the text representation is unknown. We propose the Diffusion Lens, a method for analyzing the text encoder of T2I models by generating images from its intermediate representations. Using the Diffusion Lens, we perform an extensive analysis of two recent T2I models. Exploring compound prompts, we find that complex scenes describing multiple objects are composed progressively and more slowly compared to simple scenes; Exploring knowledge retrieval, we find that representation of uncommon concepts requires further computation compared to common concepts, and that knowledge retrieval is gradual across layers. Overall, our findings provide valuable insights into the text encoder component in T2I pipelines.