---
license: apache-2.0
task_categories:
- text-to-image
language:
- en
pretty_name: LPD-Bench
size_categories:
- n<1K
---
# LPD-Bench

The long-prompt benchmark of **PromptMoG: Prompt Embedding Mixture-of-Gaussian Sampling** (Ruan et al., BMVC 2026,
[arXiv:2511.20251](https://arxiv.org/abs/2511.20251)), copied unchanged from the official repository
[basiclab/PromptMoG](https://github.com/basiclab/PromptMoG) (commit `80e7c78`, `data/lpd_bench/prompt_XXX.json`,
Apache-2.0) into one `test` split for [ImgEval](https://github.com/Jialuo-Li/ImgEval).

1,000 prompts of 250-450 words in 25 themes (40 each). Each prompt comes with yes/no questions in three categories
(semantic, spatial, stylistic; two to four per category, three for most) and the descriptions they were written from.

| column | content |
|---|---|
| `prompt_id` | the official file index (`prompt_{prompt_id:03d}.json`) |
| `theme` | one of 25 themes |
| `prompt` | the long prompt, as in the file (the official generation and scoring code lower-case it) |
| `semantic_questions`, `spatial_questions`, `stylistic_questions` | the yes/no questions of each category |
| `semantic_description`, `spatial_description`, `stylistic_description` | the descriptions behind the questions |

Official scoring (`eval_utils/eval_lpd.py`): Qwen3-VL-30B-A3B-Instruct is asked each question followed by
" Please answer in Yes/No." with the image; a question scores the probability of the token "Yes" as the first answer
token, a category the mean over its questions, and the benchmark the mean over prompts per category.