Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DALLE-2 is Seeing Double: Flaws in Word-to-Concept Mapping in Text2Image Models

Royi Rassin, Shauli Ravfogel, Yoav Goldberg

TL;DR

The paper investigates how DALLE-2 maps individual prompt words to visual concepts, revealing violations of the single-role interpretation in language by displaying homonym duplication, cross-entity modification, and entity–modifier leakage. Using carefully constructed stimuli and control prompts, the authors quantify these effects across numerous images, showing that a word can instantiate multiple entities, modify several objects, or concurrently serve as both object and modifier. They also report second-order and cross-model observations, noting that model scale and architecture influence leakage. The findings highlight fundamental inductive biases in text-to-image models and motivate targeted mitigations and deeper examinations of encoding and decoding components. This work underscores important limitations for end users and provides a framework for systematic analysis of compositional language in multimodal generation.

Abstract

We study the way DALLE-2 maps symbols (words) in the prompt to their references (entities or properties of entities in the generated image). We show that in stark contrast to the way human process language, DALLE-2 does not follow the constraint that each word has a single role in the interpretation, and sometimes re-use the same symbol for different purposes. We collect a set of stimuli that reflect the phenomenon: we show that DALLE-2 depicts both senses of nouns with multiple senses at once; and that a given word can modify the properties of two distinct entities in the image, or can be depicted as one object and also modify the properties of another object, creating a semantic leakage of properties between entities. Taken together, our study highlights the differences between DALLE-2 and human language processing and opens an avenue for future study on the inductive biases of text-to-image models.

DALLE-2 is Seeing Double: Flaws in Word-to-Concept Mapping in Text2Image Models

TL;DR

The paper investigates how DALLE-2 maps individual prompt words to visual concepts, revealing violations of the single-role interpretation in language by displaying homonym duplication, cross-entity modification, and entity–modifier leakage. Using carefully constructed stimuli and control prompts, the authors quantify these effects across numerous images, showing that a word can instantiate multiple entities, modify several objects, or concurrently serve as both object and modifier. They also report second-order and cross-model observations, noting that model scale and architecture influence leakage. The findings highlight fundamental inductive biases in text-to-image models and motivate targeted mitigations and deeper examinations of encoding and decoding components. This work underscores important limitations for end users and provides a framework for systematic analysis of compositional language in multimodal generation.

Abstract

We study the way DALLE-2 maps symbols (words) in the prompt to their references (entities or properties of entities in the generated image). We show that in stark contrast to the way human process language, DALLE-2 does not follow the constraint that each word has a single role in the interpretation, and sometimes re-use the same symbol for different purposes. We collect a set of stimuli that reflect the phenomenon: we show that DALLE-2 depicts both senses of nouns with multiple senses at once; and that a given word can modify the properties of two distinct entities in the image, or can be depicted as one object and also modify the properties of another object, creating a semantic leakage of properties between entities. Taken together, our study highlights the differences between DALLE-2 and human language processing and opens an avenue for future study on the inductive biases of text-to-image models.
Paper Structure (16 sections, 4 figures, 4 tables)

This paper contains 16 sections, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Previous Work
  3. Methdos
  4. Stimuli
  5. Control Stimuli
  6. Results
  7. A single word realized as multiple entities
  8. A single word realized as a modifier of multiple entities
  9. A single word is realized as an entity and as a modifier of another entity
  10. Second-order stimuli
  11. Other text-to-image-models
  12. Conclusions
  13. Acknowledgements
  14. Limitations
  15. Appendix
  16. ...and 1 more sections

Figures (4)

  • Figure 1: The word bat is realized as two entities given the prompt a bat is flying over a baseball stadium.
  • Figure 2: The word seal is realized both as an object and as a modifier of another word (letter) in the prompt a seal is opening a letter, resulting in concept leakage. The prompt tends to generate sealed letters, while minimally-different prompts that do not contain the noun seal do not.
  • Figure :
  • Figure :