README.md

---
pretty_name: Poplar-9K
license: apache-2.0
language:
- en
task_categories:
- text-to-image
tags:
- human-centric
- synthetic-data
- image-text-pairs
- smartphone-photography
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: Poplar-9K.jsonl
---

# Poplar-9K

<div align="center">
  <img src="assets/teaser.jpg" alt="Examples from Poplar-9K">
</div>

<div align="center">
  <img src="https://img.shields.io/badge/arXiv-coming_soon-b31b1b?style=for-the-badge&logo=arxiv" alt="arXiv coming soon"> &nbsp;
  <a href="https://choucisan.github.io/publications/poplar"><img src="https://img.shields.io/badge/Website-Project_Page-blue?style=for-the-badge" alt="Website"></a> &nbsp;
  <a href="https://github.com/choucisan/poplar"><img src="https://img.shields.io/badge/GitHub-Poplar-181717?style=for-the-badge&logo=github" alt="GitHub"></a> &nbsp;
  <a href="https://huggingface.co/collections/choucsan/poplar"><img src="https://img.shields.io/badge/Hugging_Face-Collection-yellow?style=for-the-badge" alt="Hugging Face"></a> &nbsp;
  <a href="https://www.modelscope.cn/collections/choucisan/Poplar"><img src="https://img.shields.io/badge/ModelScope-Collection-5E4AF5?style=for-the-badge" alt="ModelScope"></a> &nbsp;
  <a href="https://choosealicense.com/licenses/apache-2.0"><img src="https://img.shields.io/badge/License-Apache_2.0-green?style=for-the-badge" alt="License"></a>
</div>

**Poplar-9K** is a curated synthetic dataset of 9,401 human-centric images and
their original English generation prompts. It focuses on the visual character
of everyday personal photography: selfies, mirror photos, casual portraits,
small group photos, and ordinary moments recorded by a phone.

The collection contains 10,877 depicted people according to its structured
synthesis specifications. It covers varied subjects, clothing, activities,
scenes, camera viewpoints, lighting conditions, and photographic compositions.


## Dataset Summary

| Property | Value |
|---|---:|
| Curated images | 9,401 |
| Depicted people | 10,877 |
| Average people per image | 1.157 |
| Single-person images | 8,243 (87.68%) |
| Multi-person images | 1,158 (12.32%) |
| Prompt language | English |
| Image format | PNG |
| Aspect ratios | 1:1, 4:3, 3:4, 3:2, 2:3 |

## What Poplar-9K Contains

Poplar-9K is designed around human-centered, personally recorded photographs
rather than studio portraits or editorial imagery. Its scenes include domestic
interiors, neighborhood streets, workplaces, public venues, social events,
fitness and leisure settings, and travel environments.

The prompts describe both semantic content and photographic capture conditions:

- subject count, adult age range, gender presentation, and cultural-background descriptor;
- hairstyle, facial characteristics, expression, pose, clothing, color, and accessories;
- scene, activity, and relationships between people;
- selfie, mirror, timer, friend-taken, or candid capture method;
- framing, lighting, phone-camera characteristics, and composition.

Every released image retains the exact prompt used for generation. Prompts were
not captioned, rewritten, or repaired after image review.

## Construction and Curation

Poplar-9K was produced through three stages:

1. **Specify:** structured attributes were sampled under compatibility rules and
   expressed as photography-oriented prompts.
2. **Render:** each specification was rendered at one composition-appropriate
   aspect ratio, followed by conservative checks for grayscale, severe blur,
   low detail, and invalid image files.
3. **Inspect:** every image-prompt pair was reviewed for realistic appearance,
   image integrity, visible faces, physical plausibility, and material prompt
   consistency.

The construction run submitted 11,765 candidates for final review. A total of
9,401 images were retained and 2,364 were rejected, giving a 79.9% acceptance
rate. Rejection reasons included conspicuous synthetic appearance, collage or
duplication artifacts, cropped or missing faces, anatomy and object-interaction
errors, implausible content, monochrome output, and critical or major
prompt-image mismatches.

<div align="center">
  <img src="assets/pipeline.jpg" alt="Poplar dataset construction process">
</div>

## Dataset Composition

### Age

Younger adults form the largest portion of the synthesis specification. People
in their early or late 20s account for 66.91% of all depicted people. The
dataset also includes adults specified in their 30s through 70s.

| Age specification | People | Share |
|---|---:|---:|
| Early 20s | 4,266 | 39.22% |
| Late 20s | 3,012 | 27.69% |
| 30s | 1,646 | 15.13% |
| 40s | 1,045 | 9.61% |
| 50s | 620 | 5.70% |
| 60s | 211 | 1.94% |
| 70s | 77 | 0.71% |

### Clothing and Scenes

The most frequent clothing groups are casual clothing (18.76%), trendy feminine
styles (15.50%), homewear (12.28%), sportswear (10.45%), and streetwear
(10.29%). Smaller groups cover workwear, formal events, weather layers, and
contemporary or classic cultural clothing.

Scene groups include home (35.01%), travel and outdoor settings (23.15%),
fitness and leisure (16.99%), work and study (12.78%), and social events
(12.09%). The scene distribution has a long tail: the most frequent individual
scene accounts for only 3.82% of the collection.

### Aspect Ratios

| Aspect ratio | Images | Share |
|---|---:|---:|
| 4:3 | 6,628 | 70.50% |
| 3:4 | 1,849 | 19.67% |
| 1:1 | 840 | 8.94% |
| 3:2 | 81 | 0.86% |
| 2:3 | 3 | 0.03% |

Additional tables and publication-quality plots are available in
[`statistics/`](statistics/).

## Data Format

```text
Poplar-9K/
├── Poplar-9K.jsonl
├── images/
│   ├── Poplar-9K_000001.png
│   ├── Poplar-9K_000002.png
│   └── ...
├── assets/
└── statistics/
```

Each line of `Poplar-9K.jsonl` describes one image-text pair:

| Field | Description |
|---|---|
| `id` | Contiguous public sample identifier |
| `image` | Relative path to the PNG image |
| `source_id` | Identifier from the original construction run |
| `prompt` | Immutable English generation prompt |
| `prompt_hash` | SHA-256 hash of the normalized prompt |
| `taxonomy_version` | Version of the synthesis taxonomy |
| `metadata` | Structured subjects, clothing, scene, capture, and lighting attributes |
| `render` | Canvas, seed, model, adapter, and preliminary-filter provenance |
| `review` | Final keep decision and inspector provenance |

The `render` and `review` fields support auditing and reproducibility. Users
interested only in image-text training can use the `image` and `prompt` fields.

## Intended Uses

Poplar-9K may be useful for:

- human-centric text-to-image research;
- image-text representation learning and retrieval;
- prompt-following and compositional generation studies;
- synthetic-data filtering and quality-control research;
- analysis of everyday photographic descriptions and capture conditions.

The dataset is not intended for face recognition, identity matching,
demographic measurement, surveillance, or estimating real-world population
statistics.

## Limitations

- All people are synthetically generated and do not represent verified real identities.
- Demographic-style fields are requested synthesis attributes, not labels inferred from images.
- The configured sampling distribution is deliberately curated and should not be treated as a natural population distribution.
- Automated review can miss subtle artifacts or reject plausible uncommon content.
- Retained images may still contain minor generation artifacts or minor prompt differences.
- The dataset is relatively small and emphasizes everyday phone-style photography rather than the full space of human imagery.

## Citation

```bibtex
@misc{zou2026poplar,
  title  = {Poplar: A Scalable Pipeline for Human-Centric Image Dataset Synthesis},
  author = {Zhishan Zou},
  year   = {2026},
  note   = {arXiv preprint}
}
```

## Acknowledgements

The dataset was constructed with
[Qwen3.5-27B-FP8](https://huggingface.co/Qwen/Qwen3.5-27B-FP8),
[Krea 2 Turbo](https://huggingface.co/krea/Krea-2-Turbo), and
[Krea2-realism-V2](https://huggingface.co/RudySen/Krea2-realism-V2).

For questions or corrections, contact
[choucisan@gmail.com](mailto:choucisan@gmail.com).