EvitFan/SmolTextRender

---
license: mit
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: val
path: data/val-*
- split: test
path: data/test-*
dataset_info:
features:
- name: id
dtype: int64
- name: image_id
dtype: string
- name: image
dtype: image
- name: text
dtype: string
- name: caption
dtype: string
- name: prompt
dtype: string
- name: split
dtype: string
- name: ocr_confidence
dtype: float64
- name: ocr_backend
dtype: string
- name: caption_model
dtype: string
- name: source
dtype: string
- name: sharpness
dtype: float64
- name: brightness
dtype: float64
- name: contrast
dtype: float64
- name: resolution_w
dtype: int64
- name: resolution_h
dtype: int64
- name: text_length
dtype: int64
- name: word_count
dtype: int64
- name: phrase_reconstructed
dtype: bool
splits:
- name: train
num_bytes: 58573006
num_examples: 800
- name: val
num_bytes: 6821157
num_examples: 100
- name: test
num_bytes: 6848431
num_examples: 100
download_size: 72132017
dataset_size: 72242594
task_categories:
- image-to-text
- text-to-image
language:
- en
tags:
- ocr
- image-captioning
- text-rendering
- synthetic
- blip2
- easyocr
- flux
size_categories:
- 1K<n<10K
source_datasets:
- stzhao/AnyWord-3M
---

# Text-in-Image OCR Dataset

*Built for **Project 12 — Efficient Image Generation**, as part of the ENSTA course [CSC_5IA21](https://giannifranchi.github.io/CSC_5IA21.html)*

**Team:** Adam Gassem · Asma Walha · Achraf Chaouch · Takoua Ben Aissa · Amaury Lorin
**Tutors:** Arturo Mendoza Quispe · Nacim Belkhir

---

## Dataset Summary

A curated text-in-image dataset designed for fine-tuning text-to-image generative models (e.g. FLUX, Stable Diffusion, ControlNet) on accurate **text rendering**. Each sample pairs a real-world image containing readable text with:

- a verified OCR transcription (EasyOCR),
- a visual caption (BLIP-2),
- and a training prompt that embeds the OCR text verbatim.

Images are sourced from [AnyWord-3M](https://huggingface.co/datasets/stzhao/AnyWord-3M) and pass a rigorous multi-step quality pipeline before inclusion.

---

## Dataset Structure

| Split | Size |
|-------|------|
| train | 800 samples |
| val | 100 samples |
| test | 100 samples |

### Fields

| Field | Type | Description |
|-------|------|-------------|
| `image` | Image | The filtered image (512 px, JPEG) |
| `text` | string | Verified OCR text found in the image |
| `caption` | string | General visual description generated by BLIP-2 |
| `prompt` | string | Training prompt embedding the OCR text verbatim |
| `ocr_confidence` | float | EasyOCR confidence score (0–100) |
| `ocr_backend` | string | OCR engine used (`easyocr`) |
| `caption_model` | string | Captioning model used (`blip2` or `blip`) |
| `source` | string | AnyWord-3M subset of origin |
| `sharpness` | float | Laplacian variance of the image |
| `brightness` | float | Mean pixel brightness |
| `contrast` | float | Pixel standard deviation |
| `resolution_w` / `resolution_h` | int | Image dimensions in pixels |
| `text_length` | int | Character count of the OCR text |
| `word_count` | int | Word count of the OCR text |
| `phrase_reconstructed` | bool | Whether the full phrase was expanded beyond the bounding box |

### Sample record

```json
{
"image": "<PIL.Image>",
"text": "OPEN",
"caption": "A storefront with a neon sign above the door.",
"prompt": "A storefront with a neon sign above the door, with the text \"OPEN\" clearly visible",
"ocr_confidence": 87.5,
"source": "AnyWord-3M/laion",
"sharpness": 142.3,
"resolution_w": 512,
"resolution_h": 384
}
```

---

## Usage

```python
from datasets import load_dataset

ds = load_dataset("your-org/your-dataset-name")

# Access a training sample
sample = ds["train"][0]
print(sample["prompt"])
sample["image"].show()
```

For fine-tuning with the prompt field:

```python
for sample in ds["train"]:
image = sample["image"] # PIL image
prompt = sample["prompt"] # text-conditioned training caption
text = sample["text"] # ground-truth OCR string
```

---

## Creation Pipeline

Images are drawn from AnyWord-3M (streamed) and pass through the following stages:

```
AnyWord-3M stream
│
▼
1. Annotation filtering → valid, short, English text regions only
│
▼
2. Image quality gate → resolution ≥ 256 px, sharpness ≥ 80,
brightness 30–230, contrast ≥ 20
│
▼
3. EasyOCR verify → confirm annotated text is readable (conf ≥ 0.40)
│
▼
4. EasyOCR reconstruct → expand to the full visible phrase
│
▼
5. BLIP-2 caption → general visual description
│
▼
6. Prompt construction → natural sentence with OCR text in quotes
│
▼
7. Split & save → 80 % train / 10 % val / 10 % test
```

---

## Source Subsets

| Subset | Description |
|--------|-------------|
| `laion` | Web-crawled natural images |
| `OCR_COCO_Text` | COCO scene text |
| `OCR_mlt2019` | Multi-language (English filtered) |
| `OCR_Art` | Artistic / designed text |

---

## Citation & Project

This dataset was produced as part of the **Efficient Image Generation** project at ENSTA Paris.
Full methodology, training experiments, and inference benchmarks are documented in the [project report](https://drive.google.com/file/d/1ay4-cBOSt4LbLhwgQ0gBykda1Bu0HUXY/view?usp=drive_link).

---

## License

Released under the **MIT License** — free to use, modify, and distribute without restriction. Note that the AnyWord-3M source dataset and BLIP-2 model are subject to their own respective licenses on HuggingFace.