Pith. sign in

REVIEW 5 cited by

Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2403.05846 v2 pith:VWAN3P6K submitted 2024-03-09 cs.CV cs.CL

Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines

classification cs.CV cs.CL
keywords textdiffusionencoderlensmodelsrepresentationcomparedconcepts
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

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.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. When Do Diffusion Models learn to Generate Multiple Objects?

    cs.CV 2026-04 unverdicted novelty 8.0

    Diffusion models' multi-object generation is limited primarily by scene complexity and held-out combinations rather than imbalance, with counting difficult in low data and compositional generalization collapsing as mo...

  2. Verbalizable Representations Form a Global Workspace in Language Models

    cs.CL 2026-07 conditional novelty 7.0

    Language models represent their current reasoning in a small, readable set of verbalizable vectors (the J-space) that functions like a global workspace.

  3. Intermediate Text Representation Guided Text-to-Image Generation for Enhancing One-and-Only Alignment

    cs.CV 2026-06 unverdicted novelty 6.0

    IR-guided diffusion injects intermediate text representations into early denoising steps to improve alignment for one-and-only objects, reporting up to 19.1pp VQAScore gains on OAO-AttackBench and other benchmarks.

  4. When Do Diffusion Models learn to Generate Multiple Objects?

    cs.CV 2026-04 unverdicted novelty 6.0

    Using the mosaic controlled dataset framework, experiments show scene complexity dominates over concept imbalance in diffusion model failures for multi-object generation, with counting especially hard in low-data regi...

  5. SAEmnesia: Erasing Concepts in Diffusion Models with Supervised Sparse Autoencoders

    cs.CV 2025-09 conditional novelty 6.0

    A supervised sparse autoencoder binds each concept to a single neuron, letting Stable Diffusion erase a concept by steering one latent.