Skywork/unipic_seedream_5images

---
license: apache-2.0
task_categories:
- image-to-image
- text-to-image
language:
- en
tags:
- image-composition
- multi-image
- image-fusion
- image-editing
- unipic
- 5-image-input
pretty_name: UniPic Nano 5Images
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: train
path: "*.jsonl"
---

# 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:

```bash
# 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:

```json
{
"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

```python
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

```python
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

```python
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

```python
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](https://arxiv.org/abs/2601.15664)
- **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:

```bibtex
@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](https://huggingface.co/datasets/Skywork/unipic_nano_5images).

---