AI research

Visual Prompt Engineering

Two peer-reviewed papers on pairing LLMs with diffusion models to speed up creative writing — 39–52% fewer prompt iterations to hit the same story quality.

Paper: Journal of Student Research (2024) Paper: National High School Journal of Science (2025)

The same duck subject regenerated across three prompt iterations: a flat circular vignette, then a monochrome city street, then a full-colour street scene with passers-by — One subject across successive prompt iterations. Played at 4× speed.

Tech

GPT-4DALL·EPrompt engineeringEvaluation rubricHuman graders

Highlights

Compared prompt iterations
Evaluated stories across multiple dimensions
Published in two peer-reviewed journals

Overview

Independent research (Jun–Nov 2024) on whether visual prompts can improve creative story writing and how stories change across prompt iterations — published as two peer-reviewed papers in 2025.

Method

I designed controlled experiments comparing text-only prompt refinement against visual prompt engineering, combining LLM story generation with diffusion-generated visuals — GPT-4 and DALL·E. Stories generated under each condition were evaluated with a rubric across multiple dimensions, with human graders scoring the outputs.

Findings

Follow-up experiments found 39–52% fewer prompt iterations were needed to reach matched story-quality targets. Full methodology and results are in the papers linked above.

Publications

“Efficient Visual Prompt Engineering for Creative Story Writing” — Journal of Student Research 13(4) (2024), independent research.

“Enhancing Narrative Efficiency in LLMs via Prompt Engineering” — The National High School Journal of Science (2025), independent research.