Datasets:
Tasks:
Text Generation
Formats:
parquet
Languages:
English
Size:
< 1K
Tags:
prompt-enhancement
speculative-decoding
text-to-video
image-to-video
text-to-image
image-to-image
License:
id stringlengths 10 44 | task stringclasses 1
value | system_prompt stringclasses 1
value | prompt stringlengths 12 143 | dataset stringclasses 4
values | base-model stringclasses 1
value | images images listlengths 1 1 |
|---|---|---|---|---|---|---|
gedit_bench_4b7e3f9099377e3823c1c3e0d924883a | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Make me look handsome. | gedit_bench | qwen-image-2.1 | |
gedit_bench_1db07f0d277222e32913bff2681faebb | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Make the image brighter. | gedit_bench | qwen-image-2.1 | |
gedit_bench_82713e857fa4a3972bd3bd560ad45d70 | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | change the weather to snow | gedit_bench | qwen-image-2.1 | |
gedit_bench_17f4fec8b5c096e4ee3e8f168f93d05d | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Delete the red headscarf. | gedit_bench | qwen-image-2.1 | |
gedit_bench_dd8355aceecda1bed1594a616b40cd11 | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Add stockings. | gedit_bench | qwen-image-2.1 | |
gedit_bench_6a498187c524c7adb7a739413c24f185 | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Add three persimmons with leaves in the bottom right corner of this painting. | gedit_bench | qwen-image-2.1 | |
gedit_bench_27dba5cccc5a6d4ca877b83eb2ca374e | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Based on this image, change the hidden "New York" text to "ALEX". | gedit_bench | qwen-image-2.1 | |
gedit_bench_37c16adc232e505fc6f0d6747d10e8f1 | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Change the color of the stockings. | gedit_bench | qwen-image-2.1 | |
gedit_bench_41d961b14b637889947080f1891f85ff | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Make him look less angry | gedit_bench | qwen-image-2.1 | |
gedit_bench_08ea4b4086ac3690f6aa0ab47d0da30f | image-to-image | # Edit Prompt Enhancer — General (v2, 精简版)
**FIRST — there are TWO separate language decisions. Do NOT conflate them.**
**(A) Language of the rewritten prompt's DESCRIPTIVE prose — every word OUTSIDE double quotes (the description you write for the diffusion model, NOT the text painted into the image). This decision ... | Swap the earring for one that looks less conspicuous. | gedit_bench | qwen-image-2.1 |
End of preview. Expand in Data Studio
Prompt Enhancement SpecBench
This repository is a collection of prompts for benchmarking speculative decoding models on prompt enhancement tasks for generative models.
It covers text-to-video and image-to-video prompt enhancement targeting minimax-h3, and text-to-image and image-to-image prompt enhancement targeting qwen-image-2.1.
Text-to-video sources
| Dataset | What we use from this dataset |
|---|---|
| VBench | Enhanced prompts |
| VBench 2.0 | Video prompts |
| FETV | Video prompts |
| VideoUFO | Detailed captions |
| World-R1 | Video prompts |
Image-to-video sources
| Dataset | What we use from this dataset |
|---|---|
| TI2V-Bench / MotiF | Image–prompt pairs |
| AIGCBench | Generated-image subset |
| Physics-IQ | Reference images and video descriptions |
| RBench | Image–prompt pairs |
| VideoUFO | Supplied middle frames paired with brief captions; used as opening references |
Text-to-image sources
| Dataset | What we use from this dataset |
|---|---|
| PartiPrompts | Image generation prompts |
| DiffusionDB | User-written image generation prompts |
| DPG-Bench | Detailed image generation prompts |
| GenEval | Compositional image generation prompts |
| T2I-CompBench / CompBench++ | Compositional image generation prompts |
Image-to-image sources
| Dataset | What we use from this dataset |
|---|---|
| GEdit-Bench | Source images and English editing instructions |
| ImgEdit | Source images and single-turn editing instructions |
| OmniEdit | Source images and editing instructions from the development split |
| UltraEdit | Source images and editing instructions from the free-form subset |
System prompts
Separate text-to-video and image-to-video instructions are included in each record. These original benchmark instructions reference MiniMaxAI's H3 Video Prompt Writing Guide; they are not official MiniMax prompts.
Qwen Image 2.1 uses its official text-to-image and image-to-image PE system prompts, copied unchanged from Qwen PE-T2I and Qwen PE-I2I. These system prompts carry the Qwen Research License, permitting research and evaluation use; commercial use requires a separate license.
Source material retains its original licenses. See licenses and attribution.
Citation
If you use this collection, please cite it and the relevant upstream datasets listed above.
@misc{fal2026promptenhancementspecbench,
author = {{fal}},
title = {{Prompt Enhancement SpecBench}},
year = {2026},
howpublished = {Hugging Face dataset},
url = {https://huggingface.co/datasets/fal/prompt-enhancement-specbench}
}
- Downloads last month
- 78