T2I-CoReBench-Images: 1,080 Prompts for Reasoning and Composition

T2I-CoReBench-Images contains images generated from 1,080 challenging prompts designed to evaluate reasoning and composition in Text-to-Image models. Developed by researcher lioooox and documented in Arxiv 2509.03516, the collection includes between 10,000 and 100,000 total images generated by various model architectures. It specifically targets the ability of models to handle causal chains, object relations, and logical consistency.

Use Cases

• Benchmarking T2I model performance on causal chain reasoning prompts
• Analyzing object relation errors in images generated from the 1,080 core prompts
• Comparing compositionality across Diffusion and Autoregressive model architectures

Strengths

• 1,080 expert-designed prompts targeting reasoning failures
• Includes outputs from both closed-source and open-source models
• Apache 2.0 licensed for research and commercial use

Limitations

• Small prompt variety (1,080) relative to general-purpose T2I benchmarks
• Focus on 'challenging' prompts may not represent model performance on standard descriptive tasks

Provenance

Source

lioooox (Arxiv 2509.03516)

Collection Method

Synthetic generation using various T2I models based on a curated prompt set

Freshness

Last updated March 2026.