Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 40 of the 40 references it cites were found in a registry or an index; 3 of the 5 sources its claims cite were checked for support.

Design Guidelines for Prompt Engineering Text-to-Image Generative Models

Open access
Liu, Vivian, Chilton, Lydia B · 2022 · CHI Conference on Human Factors in Computing Systems

Abstract

Text-to-image generative models are a new and powerful way to generate visual artwork. However, the open-ended nature of text as interaction is double-edged; while users can input anything and have access to an infinite range of generations, they also must engage in brute-force trial and error with the text prompt when the result quality is poor. We conduct a study exploring what prompt keywords and model hyperparameters can help produce coherent outputs. In particular, we study prompts structured to include subject and style keywords and investigate success and failure modes of these prompts. Our evaluation of 5493 generations over the course of five experiments spans 51 abstract and concrete subjects as well as 51 abstract and figurative styles. From this evaluation, we present design guidelines that can help people produce better outcomes from text-to-image generative models.

Overview

CoChat Summary

TL;DR

This CHI 2022 paper studies prompt engineering for text-to-image generation using VQGAN+CLIP, evaluating 5,493 generations across five experiments on prompt permutations, random seeds, optimization length, style breadth, and subject–style interaction. It derives design guidelines for users, including focusing on subject/style keywords, trying multiple seeds, using shorter optimization lengths, and warning about misinterpretation and offensive outputs.

Key findings

  • Prompt rephrasings with the same subject and style keywords did not significantly differ: Experiment 1 tested 1,296 images across 9 permutations and found no significant difference (Chi-square statistic 0.354, p=0.55), despite 71.3% annotator agreement and a low Cohen’s kappa of 0.0013.
  • Random seeds significantly affected generation quality: Experiment 2 tested 1,296 images across 9 seeds and found outlier generations significant at p<0.01, with inter-rater reliability 0.13; the authors recommend generating 3 to 9 seeds.
  • Optimization length did not correlate simply with preference: Experiment 3 evaluated 1,000 iterations every 100 steps and found significant differences (p=0.01), with 200, 100, and 500 iterations most preferred; the authors suggest 300 iterations as a default and 100–500 as sufficient for fast iteration.
  • Style and subject categories mattered: Experiment 4 found abstract styles averaged 2.63 (SE 0.06) versus figurative 3.16 (SE 0.06), p<0.01; Western 2.92 (SE 0.07) versus non-Western 2.95 (SE 0.06), p=0.377; digital 2.41, modern 2.83, premodern 3.11, p<0.001. Experiment 5 found concrete subjects generally rated higher, with top ten subjects averaging concreteness 4.47 and a Pearson correlation of 0.62 between concreteness and generation quality.

Methods

  • Five experiments generated 5,493 images with VQGAN+CLIP pretrained on ImageNet with a 16,384 codebook; images were 256×256 and typically run for 300 steps on a local NVIDIA GeForce RTX 3080 GPU, except Experiment 3 varied iterations up to 1,000.
  • Experiments manipulated prompt permutations (12 subjects × 12 styles × 9 permutations), random seeds (12 subjects × 12 styles × 9 seeds), iteration length (6 subjects × 12 styles, evaluated every 100 iterations), style breadth (51 styles), and subject–style interaction (51 subjects × 31 styles).
  • Annotators with media arts/art practice backgrounds rated generations using grids or rubrics; analyses included Chi-square, Fisher’s exact, Mann-Whitney, Kruskal-Wallis, two-way ANOVA, Pearson correlation, and inter-rater reliability.

Limitations

  • The paper states that most experiments used only one prompt, “SUBJECT in the STYLE of,” with VQGAN+CLIP; future work could test other prompts, models, image conditioning, intermediate iteration controls, and modifiers such as “4k” or “2048px.”
  • The paper states that qualitative analysis of styles and subjects could be extended to understand cultural nuance and whether generations move beyond flat reproductions.
  • Evident from the text: annotation was highly subjective, with low inter-rater reliability in Experiment 1 (Cohen’s kappa 0.0013) and slight agreement in Experiment 2 (0.13); only two annotators rated most tasks, one combination was removed for inappropriate content, and the subject-only pilot was abandoned because generations were too poor to evaluate.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3491102.3501825

Cites

40 of 40 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

3 Valid · 6 Not valid · 2 Not checked · 149 Own results and interpretations

100 more claims

Create a free account to see all 160 claims and every source verdict.

See all claims — free account
  • The quote only partly states this claim. 1
  • It may state two separate claims as one. 30
  • Probably not a substantial claim. 32

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

4 of 4

Citations

Cites

Show all 28 references

Research works printed but not linked: 1

  1. wikiart.org. 2022. Visual Art Encyclopedia. https://www.wikiart.org/ wikiart.org. 2022. Visual Art Encyclopedia. https://www.wikiart.org/

Cited by

See all 639

Metadata

Authors
Liu, V., & Chilton, L. B.
Affiliations
Columbia University (US)
Venue
CHI Conference on Human Factors in Computing Systems
Publisher
ACM
Published
2022-04-29
Pages
1-23
OpenAlex
W3199400376
Cited by
639
References
28
Citations by year
2026: 112 · 2025: 212 · 2024: 220 · 2023: 91 · 2022: 4
FWCI
47.78
Type
conference paper
Language
en
Open access
Locations
  • CHI Conference on Human Factors in Computing Systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
NSF (National Science Foundation) (DGE - 1644869)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
  • Physical Sciences › Computer Science › Computer Graphics and Computer-Aided Design › Digital Media and Visual Art
  • Social Sciences › Arts and Humanities › Visual Arts and Performing Arts › Art, Technology, and Culture
Keywords
text-to-image generation, prompt engineering, design guidelines

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-13, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-13 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Design Guidelines for Prompt Engineering Text-to-Image Generative Models. https://app.cochat.ai/papers/10.1145/3491102.3501825
The paper
Liu, V., & Chilton, L. B. (2022). Design Guidelines for Prompt Engineering Text-to-Image Generative Models. CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3491102.3501825

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.