EvitFan/SmolTextRender

Text-in-Image OCR Dataset

Built for Project 12 — Efficient Image Generation, as part of the ENSTA course CSC_5IA21

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 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

{
  "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

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:

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.

---

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.