| --- |
| language: |
| - en |
| tags: |
| - Image_Generation |
| pretty_name: Object State Bench |
| size_categories: |
| - n<1K |
| --- |
| |
| # Dataset Card: Object State Bench & GenAI‑Object‑State |
|
|
| ## Dataset Details |
|
|
| ### Dataset Description |
|
|
| This repository provides two English‐only prompt collections designed to **evaluate** text‑to‑image models on their ability to depict **object absence** (empty or negated states): |
|
|
| 1. **Object State Bench** |
| - **Size:** 200 prompts |
| - **Origin:** 100 prompts from a fully‑automatic synthetic pipeline; 100 prompts authored by human annotators. |
| - **Focus:** Common objects explicitly in empty or absent states (e.g., “A table without any bottle on it”). |
|
|
| 2. **GenAI‑Object‑State** |
| - **Size:** 214 prompts |
| - **Origin:** Manually filtered subset of GenAI‑Bench’s “negation” category, retaining only prompts requiring object absence. |
| - **Focus:** Negation examples (e.g., “A vase with water, but no flowers to nourish”). |
|
|
| Both datasets are **prompt‑only** (no images or labels) and intended **exclusively for evaluation**. |
|
|
| - **Curated by:** Tianle Chen, Chaitanya Chakka, Deepti Ghadiyaram et al. |
| - **Based on paper:** “Improving Physical Object State Representation in Text‑to‑Image Generative Systems” |
|
|
| ## Uses |
|
|
| ### Direct Use |
|
|
| - **Evaluation of text‑to‑image models**: compute metrics such as CLIP‑Score, VQA‑Score, or GPT‑based correctness on generated images for each prompt. |
|
|
| ### Out‑of‑Scope Use |
|
|
| - **Training**: these lists lack paired ground‑truth images. |
| - **Non‑English prompts**: prompts are English only. |
|
|
| ## Dataset Structure |
|
|
| All files are flat JSON lists of strings, with no splits. |
|
|
