Skywork/unipic_seedream_5images

UniPic-Nano-5Images: A Multi-Image Composition Dataset

⚡ Quick Start

The image archive is split into multiple parts for easier downloading. To reconstruct and extract:

# Step 1: Concatenate split files into a single zip
cat nano-banana.part_* > nano-banana-5images.zip

# Step 2: Extract the images
unzip nano-banana-5images.zip

📖 Overview

UniPic-Nano-5Images is a high-quality multi-image composition dataset containing 47,461 samples designed for training advanced image fusion and composition models. Each sample consists of 5 input images and 1 output image, where elements from all five input images are seamlessly combined based on natural language instructions. This dataset is part of the UniPic series and has been used in UniPic3 for training multi-image composition models with complex, multi-element fusion capabilities.

🎯 Key Features

• 5-Image Input: Each sample uses exactly 5 input images for complex composition
• Multi-Element Fusion: Combines person + 4 objects/elements in sophisticated ways
• Diverse Composition Patterns: Covers extensive composition scenarios with multiple simultaneous actions
• High Quality: 47,461 carefully curated samples with detailed natural language instructions
• Production Ready: Used in UniPic3 for real-world multi-image composition applications
• Simple Format: Clean JSON format with straightforward input/output structure

📊 Dataset Statistics

Action Distribution

Action	Count	Percentage	Description
Holding	40,589	85.5%	Person holding objects
Wearing	31,374	66.1%	Person wearing accessories/clothing
Standing	28,294	59.6%	Person standing in scene
Sitting	11,401	24.0%	Person sitting on furniture
Playing	6,242	13.2%	Person playing instruments
Resting	3,543	7.5%	Object resting in scene
Using	2,982	6.3%	Person using devices
Leaning	2,764	5.8%	Person/object leaning
Carrying	2,001	4.2%	Person carrying items
Other Actions	~3,500	7.4%	Cleaning, cooking, driving, riding, etc.

Action Combination Distribution

Action Combination	Count	Percentage	Description
Holding + Standing + Wearing	15,754	33.2%	Person standing, wearing items, and holding objects
Holding + Standing	6,297	13.3%	Person standing and holding objects
Holding + Sitting + Wearing	4,708	9.9%	Person sitting, wearing items, and holding objects
Holding + Wearing	3,866	8.1%	Person wearing and holding items
Holding + Sitting	2,538	5.3%	Person sitting and holding objects
Standing + Wearing	1,709	3.6%	Person standing and wearing items
Sitting + Wearing	1,325	2.8%	Person sitting and wearing items
Holding + Playing + Wearing	948	2.0%	Person wearing, playing, and holding
Holding + Playing + Standing	871	1.8%	Person standing, playing instrument, holding objects
Holding + Sitting + Standing	790	1.7%	Complex pose combinations
Other Combinations	~8,655	18.2%	Various other action combinations

Element Type Distribution

Element Type	Count	Percentage	Description
Objects	43,425	91.5%	Handheld items (cups, bottles, books, cameras, etc.)
Wearables	33,330	70.2%	Accessories (glasses, hats, watches, jewelry, etc.)
Furniture	15,179	32.0%	Seating and surfaces (sofas, chairs, beds, etc.)
Vehicles	13,719	28.9%	Transportation (cars, motorcycles, bicycles, etc.)
Instruments	12,625	26.6%	Musical instruments (piano, guitar, drums, etc.)
Appliances	12,103	25.5%	Home devices (refrigerators, lamps, TVs, etc.)
Scenes/Backgrounds	6,307	13.3%	Environmental elements (trees, buildings, etc.)

Top Object Categories

Object Category	Count	Object Category	Count
Plate	2,685	Surfboard	2,007
Cup	2,564	Couch	1,958
Wine Glass	2,531	Drum	1,957
Kettle	2,433	Saxophone	1,906
Pot	2,394	Towel	1,840
Bottle	2,392	Handbag	1,837
Bucket	2,377	Candle	1,833
Canned	2,375	Tea Pot	1,831
Bowl	2,310	Baseball Bat	1,787
Guitar	2,310	Stool	1,762
Tennis Racket	2,282	Bed	1,715
Vase	2,206	Flute	1,687
Piano	2,139	Backpack	1,683
Skateboard	2,109
Fishing Rod	2,105
Golf Club	2,090
Chair	2,042

📁 Dataset Structure

Data Format

Each sample in the dataset is a JSON object with the following structure:

{
  "input_images": [
    "path/to/0.png",
    "path/to/1.png", 
    "path/to/2.png",
    "path/to/3.png",
    "path/to/4.png"
  ],
  "instruction": "A woman in a denim jacket and black skirt from Image1 is elegantly holding a plate with the word \"CONTCOIVER\" from Image2, while wearing a bracelet from Image4 and carrying a brown leather handbag from Image5, with a bottle of Kiehl's Calendula Deep Cleansing Foaming Face Wash from Image3 placed nearby, creating a stylish and sophisticated ensemble.",
  "output_image": "path/to/fusion_result.png",
  "id": 1
}

Field Descriptions

• input_images: List of exactly 5 input image paths

  • Image1: Contains the main subject (person/people)
  • Image2-5: Contains various objects/elements to be composed (objects, accessories, furniture, instruments, vehicles, scenes, etc.)

• instruction: Natural language description of how to combine all five images, typically following patterns like:

  • Subject description from Image1
  • Multiple actions with elements from Image2-5
  • Scene/atmosphere description

• output_image: Path to the composed output image

• id: Unique identifier for the sample

Composition Pattern

The dataset follows a consistent 5-element composition pattern:

[Subject from Image1] + [Elements from Image2-5] → [Fused Output]

Example instructions:

• "A woman in a denim jacket and black skirt from Image1 is elegantly holding a plate from Image2, while wearing a bracelet from Image4 and carrying a brown leather handbag from Image5, with a bottle of cleanser from Image3 placed nearby."
• "A woman from Image1 is elegantly wearing a necklace with a flower pendant from Image5, holding a golf club from Image2, standing in a stylish bedroom with a green bed from Image3, and holding a silver plate from Image4."
• "A man from Image1 is wearing a navy blue riding helmet from Image2, holding a violin from Image3, standing next to a pair of front-loading washing machines from Image4, and cooking a pizza in a bright pink oven from Image5."

🚀 Usage

Loading the Dataset

Using Hugging Face Datasets

from datasets import load_dataset

# Load the dataset from Hugging Face
dataset = load_dataset("Skywork/unipic_nano_5images", split="train")

# Access a sample
sample = dataset[0]
print(f"Input images: {sample['input_images']}")  # 5 images
print(f"Instruction: {sample['instruction']}")
print(f"Output image: {sample['output_image']}")

Direct JSON Loading

import json

# Load from local JSONL file
samples = []
with open("seedream_5imgs_all.jsonl", "r", encoding="utf-8") as f:
    for line in f:
        sample = json.loads(line.strip())
        samples.append(sample)

print(f"Total samples: {len(samples)}")  # 47,461

Using PyTorch DataLoader

from torch.utils.data import Dataset, DataLoader
from PIL import Image
import json

class UniPicNano5ImagesDataset(Dataset):
    def __init__(self, jsonl_path, image_root):
        self.samples = []
        with open(jsonl_path, "r", encoding="utf-8") as f:
            for line in f:
                self.samples.append(json.loads(line.strip()))
        self.image_root = image_root
    
    def __len__(self):
        return len(self.samples)
    
    def __getitem__(self, idx):
        sample = self.samples[idx]
        # Load all 5 input images
        input_imgs = [
            Image.open(f"{self.image_root}/{sample['input_images'][i]}")
            for i in range(5)
        ]
        # Load output image
        output = Image.open(f"{self.image_root}/{sample['output_image']}")
        
        return {
            "input_images": input_imgs,
            "instruction": sample["instruction"],
            "output_image": output,
            "id": sample["id"]
        }

dataset = UniPicNano5ImagesDataset("seedream_5imgs_all.jsonl", "images/")
dataloader = DataLoader(dataset, batch_size=4, shuffle=True)

Filtering by Composition Pattern

import json

def categorize_composition(instruction):
    """Categorize a sample based on its composition pattern."""
    instruction = instruction.lower()
    
    has_wearing = 'wearing' in instruction
    has_holding = 'holding' in instruction
    has_sitting = 'sitting' in instruction
    has_standing = 'standing' in instruction
    has_playing = 'playing' in instruction
    
    actions = []
    if has_wearing: actions.append('Wearing')
    if has_holding: actions.append('Holding')
    if has_sitting: actions.append('Sitting')
    if has_standing: actions.append('Standing')
    if has_playing: actions.append('Playing')
    
    return ' + '.join(actions) if actions else 'Other'

# Filter samples by composition pattern
with open("seedream_5imgs_all.jsonl", "r") as f:
    samples = [json.loads(line) for line in f]

standing_wearing_holding = [
    s for s in samples 
    if categorize_composition(s['instruction']) == 'Wearing + Holding + Standing'
]
print(f"Standing + Wearing + Holding samples: {len(standing_wearing_holding)}")  # ~15,754

🔬 Task Categories

1. Multi-Object Compositions (91.5%)

Person from Image1 interacting with multiple objects from Image2-5:

• Multi-Handheld: Holding multiple objects simultaneously
• Object + Scene: Objects placed in specific scenes/backgrounds
• Object + Furniture: Objects positioned on/near furniture

2. Wearable + Object Compositions (70.2%)

Person from Image1 wearing items and holding objects:

• Accessory + Handheld: Wearing glasses/watches while holding items
• Clothing + Handheld: Wearing specific clothing while carrying objects
• Full Ensemble: Complete outfit with multiple accessories and held items

3. Furniture + Activity Compositions (32.0%)

Person from Image1 on furniture with various activities:

• Sitting + Playing: Sitting on furniture while playing instruments
• Sitting + Holding: Seated with multiple held objects
• Standing Near: Standing near furniture with objects

4. Complex Multi-Element Compositions (28.9%+)

Person from Image1 in complex scenes with vehicles, appliances, instruments:

• Vehicle + Objects: Person in/on vehicle with multiple objects
• Instrument + Accessories: Playing instrument while wearing items
• Appliance + Scene: Using appliances in specific settings

🎓 Applications

This dataset is designed for training and evaluating:

• Advanced Multi-Image Composition Models: Learn to combine 5+ images seamlessly
• Complex Scene Understanding: Models that understand spatial relationships between many elements
• Instruction-Following Vision Models: Models that follow complex, multi-part composition instructions
• Multi-Element Fusion: Sophisticated blending of person + multiple objects/scenes/accessories

🔗 Related Work

This dataset is part of the UniPic dataset series:

• UniPic3: A unified multi-image composition framework. For more details, see Skywork UniPic 3.0: Unified Multi-Image Composition via Sequence Modeling
• UniPic-Nano-2Images: The 2-image version with simpler compositions
• UniPic-Nano-3Images: The 3-image version with moderate complexity

📝 Citation

If you use this dataset in your research, please cite:

@misc{wei2026skyworkunipic30unified,
  title={Skywork UniPic 3.0: Unified Multi-Image Composition via Sequence Modeling}, 
  author={Hongyang Wei and Hongbo Liu and Zidong Wang and Yi Peng and Baixin Xu and Size Wu and Xuying Zhang and Xianglong He and Zexiang Liu and Peiyu Wang and Xuchen Song and Yangguang Li and Yang Liu and Yahui Zhou},
  year={2026},
  eprint={2601.15664},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2601.15664}, 
}

📄 License

Please refer to the license terms on the Hugging Face dataset page.

---