T2I Diversity Evalprompts: 4,340 English Prompts for Text-to-Image Model Evaluation

T2I-Diversity Evaluation Prompt Set is a collection of 4,340 English prompts for evaluating text-to-image models. The dataset was created by AIML-TUDA and last updated on June 23, 2025. It contains 1,085 base prompts sourced from DrawBench and Parti-Prompts, each expanded into four variants of increasing descriptive density using GPT-4o.

Use Cases

• Benchmarking text-to-image model performance based on prompt descriptive density.
• Evaluating model robustness to prompt length and detail based on the four prompt variants per concept.
• Analyzing the relationship between prompt complexity and image generation quality.
• Training or fine-tuning text-to-image models using a structured prompt set.

Strengths

• Contains 4,340 prompts, providing a substantial test set.
• Offers four structured variants per concept (minimal, short, medium, long), enabling controlled comparisons.
• Base prompts are sourced from established benchmarks like DrawBench and Parti-Prompts.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the source benchmarks and the GPT-4o generation process.

Provenance

Source

AIML-TUDA, with base prompts from DrawBench and Parti-Prompts.

Collection Method

1,085 base prompts were expanded into four descriptive variants using GPT-4o.

Time Range

null

Freshness

Last updated 2025-06 23 18:03:39.

Geography

null