Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Surrealistic-like Image Generation with Vision-Language Models

Elif Ayten, Shuai Wang, Hjalmar Snoep

TL;DR

This study systematically compares three vision-language image generators—DALL-E 2, DreamStudio, and Deep Dream Generator—on their ability to produce surrealistic-like imagery from realist base images. It combines text prompts derived from YOLO-labeled scene objects with ChatGPT-generated descriptions, explores varying prompt lengths (15 vs 50 words), and tests base-image edits (downscaling, blurring). A three-part experiment plus a survey with 18 artists reveals that DALL-E 2 paired with 50-word ChatGPT prompts yields the most convincing surrealistic outputs, while language-model prompts broadly improve results across models; base-image edits offer limited gains. The work highlights the value of integrating large language models into prompt design to enhance creativity in AI-assisted surrealism and discusses potential biases from base-image selection and terminology. Future directions include expanding artist-name prompts and exploring additional stylistic guidance to push the surrealistic capabilities of these systems.

Abstract

Recent advances in generative AI make it convenient to create different types of content, including text, images, and code. In this paper, we explore the generation of images in the style of paintings in the surrealism movement using vision-language generative models, including DALL-E, Deep Dream Generator, and DreamStudio. Our investigation starts with the generation of images under various image generation settings and different models. The primary objective is to identify the most suitable model and settings for producing such images. Additionally, we aim to understand the impact of using edited base images on the generated resulting images. Through these experiments, we evaluate the performance of selected models and gain valuable insights into their capabilities in generating such images. Our analysis shows that Dall-E 2 performs the best when using the generated prompt by ChatGPT.

Surrealistic-like Image Generation with Vision-Language Models

TL;DR

This study systematically compares three vision-language image generators—DALL-E 2, DreamStudio, and Deep Dream Generator—on their ability to produce surrealistic-like imagery from realist base images. It combines text prompts derived from YOLO-labeled scene objects with ChatGPT-generated descriptions, explores varying prompt lengths (15 vs 50 words), and tests base-image edits (downscaling, blurring). A three-part experiment plus a survey with 18 artists reveals that DALL-E 2 paired with 50-word ChatGPT prompts yields the most convincing surrealistic outputs, while language-model prompts broadly improve results across models; base-image edits offer limited gains. The work highlights the value of integrating large language models into prompt design to enhance creativity in AI-assisted surrealism and discusses potential biases from base-image selection and terminology. Future directions include expanding artist-name prompts and exploring additional stylistic guidance to push the surrealistic capabilities of these systems.

Abstract

Recent advances in generative AI make it convenient to create different types of content, including text, images, and code. In this paper, we explore the generation of images in the style of paintings in the surrealism movement using vision-language generative models, including DALL-E, Deep Dream Generator, and DreamStudio. Our investigation starts with the generation of images under various image generation settings and different models. The primary objective is to identify the most suitable model and settings for producing such images. Additionally, we aim to understand the impact of using edited base images on the generated resulting images. Through these experiments, we evaluate the performance of selected models and gain valuable insights into their capabilities in generating such images. Our analysis shows that Dall-E 2 performs the best when using the generated prompt by ChatGPT.

Paper Structure

This paper contains 18 sections, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Artistic Image Generation
  4. Methodology
  5. Models
  6. Base Images
  7. Experimental Setup
  8. Image Generation
  9. Part 1: Comparative analysis of multiple generative models
  10. Part 2: Comparing various prompt settings within each generative model
  11. Part 3: Base image editing comparison for Deep Dream Generator and DreamStudio
  12. Survey
  13. Evaluation
  14. Aesthetics Analysis
  15. Generated Images
  16. ...and 3 more sections

Figures (6)

  • Figure 1: The original painting and some generated images in diverse styles of image generation using different models and corresponding text. Each model's distinct approach yields varied styles of output for image captions.
  • Figure 2: YOLO labeled images
  • Figure 3: Survey Part 1 Results
  • Figure 4: Survey Part 2 and 3 Results
  • Figure 5: Examples for the comparison of aesthetics of generated images
  • ...and 1 more figures