Datasets:

FireRedTeam
/

REDEdit-Bench

id stringlengths 1 4	a_to_b_instructions stringlengths 3 234	a_to_b_instructions_eng stringlengths 9 898	task stringclasses 15 values
1	在图片中绿色植物上增加一只七星瓢虫	Add a seven-spotted ladybug on the green plant in the picture	add
2	在咖啡杯里加一个白色心形拉花	Add a white heart-shaped latte art in the coffee cup	add
3	在马路上增加一个穿运动服跑步的男人	Add a man running in sportswear on the road	add
4	在熊的嘴里增加一条鲑鱼	Add a salmon in the bear's mouth	add
5	在鹿的旁边再增加一只鹿	Add another deer beside the deer	add
6	在湖上加条木船	Add a wooden boat on the lake	add
7	在粉色沙发上增加一只猫	Add a cat on the pink sofa	add
8	在画面左侧的桌子上增加几支鲜艳花朵	Add a few bright flowers on the table on the left side of the picture	add
9	加两个鸡蛋，在图片中动物的脚旁边	Add two eggs next to the animal's feet in the picture	add
10	在道路的中间增加一只大黑熊和一只小黑熊	Add a large black bear and a small black bear in the middle of the road	add
11	在图片中圆环瀑布里增加一道彩虹	Add a rainbow in the circular waterfall in the picture	add
12	在船旁边增加一条跃出水面的海豚	Add a dolphin leaping out of the water next to the boat	add
13	在树屋前面增加一只猫头鹰	Add an owl in front of the treehouse	add
14	在街道上增加熙熙攘攘的人群	Add a bustling crowd on the street	add
15	在蛋糕上面增加一个红色的樱桃	Add a red cherry on top of the cake	add
16	在白色椅子上增加一个很大的玩具熊玩偶	Add a very large teddy bear doll on the white chair	add
17	在红绿灯旁增加一个看报纸的人	Add a person reading a newspaper next to the traffic light	add
18	在海面上增加几只洁白的天鹅	Add a few white swans on the sea surface	add
19	在盘子前面增加一双筷子	Add a pair of chopsticks in front of the plate	add
20	在骑自行车的人后面增加一只跟着他的中华田园犬	Add a Chinese rural dog following behind the person riding the bicycle	add
21	增加一个骨头在图片中小狗雕像的嘴里，保持风格一致	Add a bone in the mouth of the puppy statue in the picture, keeping the style consistent	add
22	在足球场里靠近围栏的地方添加一个足球	Add a soccer ball near the fence in the soccer field	add
23	在图片中蘑菇的左侧添加一个带有剧毒标志的小木牌	Add a small wooden sign with a toxic symbol to the left of the mushroom in the picture	add
24	在草丛花朵之间增加飞舞的蝴蝶	Add fluttering butterflies among the flowers in the grass	add
25	在房子前面增加一只吃草的牛	Add a grazing cow in front of the house	add
26	增加一只企鹅	Add a penguin	add
27	在桌子上增加一对银质刀叉	Add a pair of silver knife and fork on the table	add
28	增加一个钓鱼的人	Add a person fishing	add
29	在图片中人物旁边增加一只翱翔的雄鹰	Add a soaring eagle next to the person in the picture	add
30	狼的旁边增加一只站立的土拨鼠	Add a standing marmot next to the wolf	add
31	在花瓶里增加一条金鱼	Add a goldfish in the vase	add
32	面包上增加一些彩色糖果	Add some colorful candies on the bread	add
33	再添加几只猫头鹰，动作保持一致	Add a few more owls, keeping the pose consistent	add
34	给图片中的两条腿上增加黑色丝袜	Add black stockings to the two legs in the picture	add
35	在空中增加一个热气球	Add a hot air balloon in the sky	add
36	道路中增加一辆汽车	Add a car in the road	add
37	在车辙旁边增加一列人类的脚印	Add a trail of human footprints next to the tire tracks	add
38	在泳池旁边放置几个游泳圈	Place a few swim rings next to the pool	add
39	在岸边增加一艘搁浅的木船	Add a stranded wooden boat on the shore	add
40	在画面中增加一辆收割机	Add a harvester in the scene	add
41	在飞机外面增加一个战斗机	Add a fighter jet outside the airplane	add
42	在鞋子旁边增加几个蚂蚁	Add a few ants next to the shoes	add
43	在道路中间增加一个墓碑	Add a tombstone in the middle of the road	add
44	给小鸟戴上圣诞帽	Put a Santa hat on the bird	add
45	给人物戴上一顶草帽	Put a straw hat on the person	add
46	在男孩的脖子上，衬衫领子下方，增加一条银色项链	Add a silver necklace around the boy's neck, below the shirt collar	add
47	在人物左侧的地板上增加几本散落的书	Add a few scattered books on the floor to the left of the person	add
48	在莲花叶之间增加几朵盛开的粉色莲花	Add several blooming pink lotus flowers between the lotus leaves	add
49	将天花板的玻璃穹顶变成彩色的花窗玻璃	Turn the glass dome on the ceiling into stained glass	add
50	在鞋底发光的部分增加火焰特效	Add flame effects to the glowing part of the soles	add
51	在前景中海鸥的嘴里增加一根薯条	Add a french fry in the mouth of the seagull in the foreground	add
52	在栏杆后面的草地上增加一个绿色的垃圾桶	Add a green trash can on the grass behind the railing	add
53	在远处的山脉上增加一座城堡	Add a castle on the distant mountains	add
54	在人物的左手手腕上系上一条与中式礼服风格匹配的红色丝带	Tie a red ribbon matching the style of the Chinese dress on the person's left wrist	add
55	在女人的脖子上增加一条项链，项链上有一个闪亮的红宝石吊坠	Add a necklace with a shiny ruby pendant around the woman's neck	add
56	在吧台下面增加两个吧台凳	Add two bar stools under the counter	add
57	给鸟戴上一顶草帽	Put a straw hat on the bird	add
58	将背景的干涸盐滩替换为海洋	Replace the dried salt flats in the background with an ocean	add
59	在远处的草地上增加几棵树	Add a few trees on the distant grass	add
60	在沙滩上，遮阳伞下面添加站着的人	Add standing people under the beach umbrella on the beach	add
61	在人物的耳朵上增加一个精致的金色耳坠	Add an exquisite gold earring to the person's ear	add
62	在图片中墙壁上增加几个红色的剪纸，增加画面喜庆的氛围	Add several red paper cuttings to the wall in the picture to increase the festive atmosphere	add
63	在桌子上添加一个绿色盆栽	Add a green potted plant on the table	add
64	给图片中的女人戴上一条红色围巾	Put a red scarf on the woman in the picture	add
65	让灯笼从内部发出温暖的红色光芒，照亮周围的积雪，营造出节日的氛围	Make the lantern glow with a warm red light from within, illuminating the surrounding snow to create a festive atmosphere	add
66	在猫的爪子前放一只老鼠	Place a mouse in front of the cat's paws	add
67	给少女戴上猫耳头带	Put a cat ear headband on the young girl	add
68	在鸭子下面增加几枚彩色的蛋	Add several colorful eggs under the duck	add
69	在亭子里增加几个正在观景的人	Add several people sightseeing in the pavilion	add
70	在白色雕塑底座上增加更多与现有风格一致的巴洛克式雕刻花纹	Add more Baroque carved patterns consistent with the existing style to the base of the white sculpture	add
71	给人物左手增加一个红色的手提包	Add a red handbag to the person's left hand	add
72	在草地上增加几头牦牛	Add a few yaks on the grassland	add
73	增加草地	Add grass	add
74	给南瓜灯戴上一顶巫师帽	Put a wizard hat on the jack-o'-lantern	add
75	在白色柜子的台面上增加一个笔记本电脑	Add a laptop on the counter of the blue cabinet	add
76	在人物右侧的石头上放一把二胡	Place an erhu on the stone to the right of the person	add
77	在浴缸旁边的地板上增加一个绿色的小边桌	Add a small green side table on the floor next to the bathtub	add
78	给图片中的女生穿上一件日系校服夹克	Put a Japanese school uniform jacket on the girl in the picture	add
79	在床上增加一个正在使用笔记本电脑的人	Add a person using a laptop on the bed	add
80	给小猪戴上一顶黄色的安全帽	Put a yellow hard hat on the piglet	add
81	在天花板上增加一个赛博朋克风格的霓虹灯牌	Add a cyberpunk-style neon sign on the ceiling	add
82	给图片中的人物穿上白色衬衫	Put a white shirt on the person in the picture	add
83	在踏板摩托车的座位上放一个白色的头盔	Place a white helmet on the seat of the scooter	add
84	在房子背后增加一座连绵的雪山	Add a continuous snowy mountain range behind the house	add
85	在人物的头上佩戴一顶皇冠，风格协调统一	Place a crown on the person's head, matching the overall style	add
86	在薄雾笼罩的湖面上增加一艘小木船	Add a small wooden boat on the mist-covered lake	add
87	在椅子上增加一只狗	Add a dog on the chair	add
88	在左侧的枯树枝上挂上几个小红灯笼	Hang a few small red lanterns on the withered branches on the left	add
89	在床上添加一个正在睡觉的人	Add a sleeping person on the bed	add
90	在图片中增加一个小矮人	Add a shiny gold coin to the dwarf's hand in the picture	add
91	添加一条从女性肩头垂下的精致丝巾，增强艺术和谐感	Add an exquisite silk scarf hanging from the woman's shoulder to enhance artistic harmony	add
92	在女孩的右脸颊上画一个红色的小爱心	Draw a small red heart on the girl's right cheek	add
93	在手指上增加一个戒指	Add a ring on the finger	add
94	天空中增加一只鸟	Add a bird in the sky	add
95	在人物后背增加一个红色书包	Add a red backpack to the person's back	add
96	在人物的头发上增加一个蝴蝶结发卡	Add a bow hairpin to the person's hair	add
97	给图片中的人物戴上一副黑色眼镜	Put a pair of black glasses on the person in the picture	add
98	在墙壁上增加一个简约风格的无框时钟	Add a minimalist frameless clock on the wall	add
99	给猫咪的脖子上加一个红色的小领结	Add a small red bow tie around the cat's neck	add
100	给人物戴上一个由白色小雏菊和白色康乃馨组成的花环	Put a wreath made of white daisies and white carnations on the person	add

End of preview. Expand in Data Studio

🚩 RedBench (REDEdit-Bench)

## 🔥 Introduction

RedBench (also known as REDEdit-Bench) is a comprehensive benchmark designed to evaluate the capabilities of current image editing models.

Our main goal is to build more diverse scenarios and editing instructions that better align with human language. We collected over 3,000 images from the internet, and after careful expert-designed selection, we constructed 1,673 bilingual (Chinese–English) editing pairs across 15 categories.

📢 Note on Dataset Size: The original benchmark described in the paper consists of 1,673 image pairs. However, due to strict redistribution licensing restrictions on certain commercial assets, the public release version has been curated to 1,542 pairs. This ensures full compliance with copyright laws while maintaining the diversity and quality of the benchmark.

✨ Key Features

🗣️ Human-Aligned Instructions: Diverse scenarios and editing instructions that closely mimic real-world human usage.
🌐 Bilingual Support: Full support for both Chinese and English editing instructions.
🛡️ Quality Assurance: Carefully curated by experts from a massive collection of source images.
🧩 Diverse Tasks: Covers 15 distinct categories including Object Addition, Removal, Replacement, Style Transfer, and more.

📂 Data Structure & Examples

The dataset is organized in JSONL format. Each entry contains the image source, bilingual instructions, and the specific task category.

Task Categories

The benchmark covers 15 different task categories:

Category	Count	Description
add	143	Object Addition
adjust	156	Attribute Adjustment
background	91	Background Modification
beauty	79	Beauty Enhancement
color	99	Color Modification
compose	100	Image Composition
extract	95	Element Extraction
lowlevel	47	Low-level Processing
motion	78	Motion Addition
portrait	102	Portrait Editing
remove	147	Object Removal
replace	140	Object Replacement
stylize	92	Style Transfer
text	123	Text Editing
viewpoint	50	Viewpoint Change
all	1542	All Tasks

Sample Data

{"id": "1", "source": "redbench/add/add-1.png", "a_to_b_instructions": "在图片中绿色植物上增加一只七星瓢虫", "a_to_b_instructions_eng": "Add a seven-spotted ladybug on the green plant in the picture", "task": "add"}
{"id": "2", "source": "redbench/add/add-2.png", "a_to_b_instructions": "在咖啡杯里加一个白色心形拉花", "a_to_b_instructions_eng": "Add a white heart-shaped latte art in the coffee cup", "task": "add"}
{"id": "3", "source": "redbench/add/add-3.png", "a_to_b_instructions": "在马路上增加一个穿运动服跑步的男人", "a_to_b_instructions_eng": "Add a man running in sportswear on the road", "task": "add"}

Generate Images

Before evaluating the model, you first need to use the provided JSONL file (which contains metadata information) along with the original image files to generate the corresponding edited images by editing model.

We provide the inference script redbench_infer.py for generating edited images. This script supports multi-GPU distributed inference using Accelerate.

Dependencies

Install required dependencies:

pip install accelerate diffusers transformers pillow tqdm

Then download our dataset REDEdit_Bench.tar. Please download the tar file and extract it.

Usage

To run the inference script, use the following command:

accelerate launch --num_processes <num_gpus> redbench_infer.py --model-path <path_to_model> --jsonl-path <path_to_redbench_jsonl> --save-path <path_to_save_results>

Arguments:

--model-path: Path to the model. Default is FireRedTeam/FireRed-Image-Edit-1.0.
--lora-name: Path to LoRA weights (optional).
--save-path: Directory to save the generated images (required).
--jsonl-path: Path to the JSONL file containing edit instructions (required).
--edit-task: Specific task to process (e.g., add, remove, stylize). Default is all.
--save-key: Key name for saving result path. Default is result.
--seed: Random seed. Default is 43.
--lang: Instruction language, cn or eng (default: cn).

Example:

# Generate all edited images using 8 GPUs
accelerate launch --num_processes 8 redbench_infer.py \
    --model-path FireRedTeam/FireRed-Image-Edit-1.1 \
    --jsonl-path ./redbench.jsonl \
    --save-path ./edited_images_cn \
    --edit-task all \
    --lang cn

Example Input/Output

Input

A JSONL file containing image edit instructions (redbench.jsonl):

{"id": "1", "source": "redbench/add/add-1.png", "a_to_b_instructions": "在图片中绿色植物上增加一只七星瓢虫", "a_to_b_instructions_eng": "Add a seven-spotted ladybug on the green plant in the picture", "task": "add"}
{"id": "2", "source": "redbench/add/add-2.png", "a_to_b_instructions": "在咖啡杯里加一个白色心形拉花", "a_to_b_instructions_eng": "Add a white heart-shaped latte art in the coffee cup", "task": "add"}
{"id": "3", "source": "redbench/adjust/adjust-144.png", "a_to_b_instructions": "将天空的颜色调成更深的蓝色", "a_to_b_instructions_eng": "Change the sky color to a deeper blue", "task": "adjust"}

A folder containing original images:

├── redbench                    
│   ├── add     
│   │   ├── add-1.png                 
│   │   ├── add-2.png                 
│   │   ├── ...                 
│   ├── adjust                             
│   │   ├── adjust-144.png
│   │   ├── ...
│   ├── ...

Output

A folder containing edited images:

# Without --multi-folder option:
├── edited_images                    
│   ├── 1.png                 
│   ├── 2.png            
│   ├── 3.png           
│   ...            
│   ├── result.jsonl

# With --multi-folder option:
├── edited_images                    
│   ├── add
│   │   ├── 1.png
│   │   ├── 2.png
│   │   ├── ...
│   ├── adjust
│   │   ├── 144.png
│   │   ├── ...
│   ...
│   ├── result.jsonl

Image Editing Evaluation using Gemini-3-Flash

This project evaluates image editing processes using the Gemini-3-Flash API. The system processes a set of original and edited images, comparing them according to a predefined set of criteria, such as instruction adherence, image-editing quality, and detail preservation.

We provide the evaluation script redbench_eval.py for automated evaluation using Gemini.

Overview

The goal of this project is to evaluate the quality of image editing processes using Gemini. The evaluation criteria include:

Instruction Adherence: The edit must match the specified editing instructions.
Image-editing Quality: The edit should appear seamless and natural.
Detail Preservation: Regions not specified for editing should remain unchanged.

Evaluation Criteria by Task Category

Different task categories use different evaluation metrics:

Task Category	Metrics
add, remove, replace, compose, extract	Prompt Compliance, Visual Seamlessness, Physical & Detail Fidelity
adjust, color, lowlevel	Prompt Compliance, Visual Seamlessness, Physical & Detail Fidelity
background, viewpoint	Prompt Compliance, Visual Seamlessness, Physical & Detail Fidelity
beauty, portrait	Prompt Compliance, Visual Seamlessness, Physical & Detail Fidelity
stylize	Style Fidelity, Content Preservation, Rendering Quality
motion	Prompt Compliance, Motion Realism, Visual Seamlessness
text	Text Fidelity, Visual Consistency, Background Preservation

Dependencies

pip install google-generativeai pillow tqdm

Setup

Gemini API Key: Set your Gemini API key as an environment variable:
```
export GEMINI_API_KEY="your-gemini-api-key"
```
Images and JSON File: You will need:
- A folder containing the edited images (--result_img_folder).
- A JSONL file containing edit instructions and metadata (--edit_json).
- A JSON file containing evaluation prompts for each task category (--prompts_json).

Usage

To run the evaluation script, use the following command:

python redbench_eval.py --result_img_folder <path_to_edited_images> --edit_json <path_to_redbench_jsonl> --prompts_json <path_to_prompts_json> --lang <language>

Arguments:

--result_img_folder: The directory containing the edited images (required).
--edit_json: Path to the JSONL file containing edit instructions and metadata (required).
--prompts_json: Path to the JSON file containing evaluation prompts for each task category (required).
--num_threads: Number of concurrent threads. Default is 50.
--lang: Instruction language, cn or eng (default: cn).

Example:

python redbench_eval.py \
    --result_img_folder ./edited_images \
    --edit_json ./redbench.jsonl \
    --prompts_json ./prompts.json \
    --num_threads 50 \
    --lang cn

Example Input/Output

Input

A JSONL file containing image edit instructions (redbench.jsonl):

{"id": "1", "source": "redbench/add/add-1.png", "a_to_b_instructions": "在图片中绿色植物上增加一只七星瓢虫", "a_to_b_instructions_eng": "Add a seven-spotted ladybug on the green plant in the picture", "task": "add"}
{"id": "2", "source": "redbench/add/add-2.png", "a_to_b_instructions": "在咖啡杯里加一个白色心形拉花", "a_to_b_instructions_eng": "Add a white heart-shaped latte art in the coffee cup", "task": "add"}
{"id": "3", "source": "redbench/adjust/adjust-144.png", "a_to_b_instructions": "将天空的颜色调成更深的蓝色", "a_to_b_instructions_eng": "Change the sky color to a deeper blue", "task": "adjust"}

A JSON file containing evaluation prompts for each task category (prompts_json):

{
  "add": "\nYou are a data rater specializing in grading object addition edits. You will be given two images ...",
  "remove": "\nYou are a data rater specializing in grading object removal edits. You will be given two images ...",
  "adjust": "\nYou are a data rater specializing in grading attribute alteration edits. You will be given two images ....",
  "stylize": "\nYou are a data rater specializing in grading style transfer edits. You will be given an input image, a reference style...",
  ...
}

A folder containing edited images (with --multi-folder option from inference):

├── edited_images                    
│   ├── add
│   │   ├── 1.png                 
│   │   ├── 2.png
│   │   ├── ...                
│   ├── adjust                             
│   │   ├── 144.png
│   │   ...
│   ...

Output

The script automatically computes and saves results in the result folder:

result.json - Detailed evaluation for each image:

{
    "0": "Brief reasoning: A seven-spotted ladybug was successfully added on the green plant with natural color and placement.\nPrompt Compliance: 5\nVisual Seamlessness: 4\nPhysical & Detail Fidelity: 5",
    "1": "Brief reasoning: A white heart-shaped latte art was added in the coffee cup with good blending.\nPrompt Compliance: 5\nVisual Seamlessness: 4\nPhysical & Detail Fidelity: 4",
    "2": "Brief reasoning: The sky color was changed to a deeper blue with smooth transition.\nPrompt Compliance: 5\nVisual Seamlessness: 4\nPhysical & Detail Fidelity: 5",
    ...
}

score.json - Final scores including per-category averages and overall score:

{
    "final_score": 4.3,
    "averaged_result": {
        "add": 4.5,
        "adjust": 4.2,
        "background": 3.8,
        ...
    },
    "averaged_data": {
        "0": 4.67,
        "1": 4.33,
        "2": 4.67,
        ...
    }
}

The redbench_eval.py script automatically computes:

Individual image scores (extracted from Gemini responses)
Per-category averages (averaged_result)
Overall final score (average of all category scores)

See the Output section above for the complete score.json structure.

🧩 License

REDEdit-Bench is released under the Creative Commons Attribution–NonCommercial–NoDerivatives (CC BY-NC-ND 4.0) license.

✅ Free for academic research purposes only
❌ Commercial use is prohibited

🖼️ Data Source: All images included in REDEdit-Bench were legally purchased and obtained through official channels to ensure copyright compliance.

By using this dataset, you agree to comply with the applicable license terms.

Downloads last month: 22

Size of downloaded dataset files:

5.57 GB

Size of the auto-converted Parquet files:

5.57 GB

Number of rows:

1,542

Paper for FireRedTeam/REDEdit-Bench

FireRed-Image-Edit-1.0 Techinical Report

Paper • 2602.13344 • Published 26 days ago • 4