DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generative Transformers

By Jaemin Cho, Abhaysinh Zala, Mohit Bansal. 2022

133 citations recorded by the source corpus.

Subfields: Generative Models, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Generative Models

In Safety, Robustness & Interpretability

Citations in the corpus

Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding (2022, methodology): “Recent work by [10] proposed a new evaluation set called PaintSkills to systematically evaluate visual reasoning skills and social biases beyond COCO.”
Video Diffusion Models (2022, neutral): “In particular, our text-to-video models inherit the challenges faced by prior work on text-to-image models, and our future work will involve auditing for forms of social bias, similar to [6, 7, 50, 12] for image-to-text and image labeling models.”
TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering (2023, methodology): “We include 161 from DrawBench used in Imagen [47] (texts that are categorized as “misspellings” and “rare words” are removed); 1420 from PartiPrompt used in Parti [61] (texts in category “abstract” are removed); and 500 texts from PaintSkill used in DALL-Eval [6].”
Underspecification in Scene Description-to-Depiction Tasks (2022, background): “Ethical questions concerning bias, safety and misinfor-mation are increasingly recognized (Saharia et al., 2022; Cho et al., 2022); nevertheless, understanding which system behaviors are desirable requires a vocabulary and framework for understanding the diverse and quickly expanding capabilities…”
GenEval: An Object-Focused Framework for Evaluating Text-to-Image Alignment (2023, methodology): “Spatial understanding has been found by prior works to be challenging for T2I models [14, 9]; our findings show this is still true.”
Stable Bias: Evaluating Societal Representations in Diffusion Models (2023, unknown): “…understand their inner workings, such as the functioning of their safety filters [51], the structure of their semantic space [14], the extent to which they generate stereotypical representations of given demographics [10, 19, 44] and memorize and regenerate specific training examples [16, 62].”
When and why vision-language models behave like bags-of-words, and what to do about it? (2022, background): “…uneven distribution across people can assist in advocating to reform, avoid, or reject the use of these models, and this connects to a broader body of literature directly exposing many biases and stereotypes perpetuated by these models (Cho et al., 2022; Bansal et al., 2022; Wolfe et al., 2022).”
EvalCrafter: Benchmarking and Evaluating Large Video Generation Models (2023, methodology): “Since we have already collected a very large scale of prompts from real-world users and there are also available text-to-image evaluation prompts, e.g ., DALL-Eval [12] and Draw-Bench [45], we also integrate these prompts to our benchmark list.”
Text-to-image Diffusion Models in Generative AI: A Survey (2023, methodology): “However, PaintSKills [80] only focuses on unseen objectcolor and object-shape scenario [63].”
LayoutGPT: Compositional Visual Planning and Generation with Large Language Models (2023, methodology): “For spatial reasoning, each prompt falls into one of the four types of relations ({ left, right, top, below }) and we use the bounding box center for evaluation following PaintSkills [7].”
T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation (2023, unknown): “DALL-EVAL [39] proposes PaintSkills to evaluate visual reasoning skills, image-text alignment, image quality, and social bias by 7,330 prompts.”
Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale (2022, background): “, 2022); and recent research has already begun extending this critical analysis to these image-generation models (Cho et al., 2022; Bansal et al., 2022).”