deepfakesMSU/TIQA_crops_and_images

TIQA: Text-in-Image Quality Assessment

Datasets from the paper "TIQA: Human-Aligned Perceptual Text Quality Assessment in Generated Images" (Kirill Koltsov, Aleksandr Gushchin, Dmitriy Vatolin, Anastasia Antsiferova).

Text-to-image models produce globally realistic images, but rendered text often has malformed glyphs, broken strokes and irregular spacing. TIQA is a no-reference task: predict a human-aligned perceptual quality score for text regions in generated images. The score measures how the text looks, not whether it says the right thing.

The repository has two parts:

Archive	Content	Labels
crops.tar.gz — TIQA-Crops	120,197 horizontal text crops from AI-generated images, 16 generators	~10k MOS from subjective study + ~110k proxy labels
images.zip — TIQA-Images	1,440 text-heavy images from 12 recent T2I models, including proprietary ones	overall-quality and text-quality MOS

Baseline model: ANTIQA — code and checkpoint. It reaches PLCC/SROCC 0.942/0.935 on TIQA-Crops and 0.842/0.837 on TIQA-Images text-quality MOS (unseen generators).

Samples in the Dataset Viewer

The viewer above shows small random samples, each as an image | score table:

Subset	Folder	Examples	Score
crops	TIQA_Crops_samples/	50	crop quality score (rating, 0–5)
images_overall	TIQA_Images_samples/tiqa_images_overall/	up to 25	overall-quality MOS (score_overall)
images_text_only	TIQA_Images_samples/tiqa_images_text_only/	up to 25	text-quality MOS (score_text), shown on *_whiten_.png images

Each sample folder has a metadata.csv with columns file_name and score.

---

TIQA-Crops

Text crops detected with PP-OCRv5 on images from open-source generators. Every crop has a quality score on a 0–5 scale (higher is better).

• 9,978 crops have a MOS from a crowdsourced subjective study (about 50 ratings per crop).
• 110,219 crops have proxy labels: PP-OCRv5 recognition confidence mapped to the MOS scale with a 5-parameter logistic (5PL) fit. Use them for pretraining.

Generators: CogView4, DeepFloyd IF, FLUX.1-dev, Kandinsky 2, OmniGen, PixArt-α, PixArt-Σ, Qwen-Image, SD 2.1, SD 3 Medium, SD 3.5 Medium, SD 3.5 Large, SD 3.5 Large Turbo, SDXL 1.0, SDXL Turbo, SDXL Lightning.

Folders with the real_world_ prefix use prompts from a different distribution (real-world scenes with text) than folders without the prefix.

Structure

tiqa_crops/
├── dataset.csv
├── info.txt
└── crops/
    └── <generator>/<image_id>/
        ├── <image_id>_crop_<k>.png
        └── ocr_texts.txt

dataset.csv

Column	Description
path	Crop path relative to crops/
ocr_score	PP-OCRv5 recognition confidence
rating_real	True — score is a MOS from the subjective study; False — proxy label from the 5PL mapping of ocr_score
rating	Crop quality score, 0–5
for_test	True if the crop is in the test split
ocr_text	Text recognized by PP-OCRv5

import pandas as pd

df = pd.read_csv("tiqa_crops/dataset.csv")
mos   = df[df.rating_real]            # human-labelled crops
proxy = df[~df.rating_real]           # proxy-labelled crops for pretraining
test  = df[df.for_test]

---

TIQA-Images

Text-heavy prompts rendered by recent text-to-image models, including proprietary APIs. Each image has two subjective scores: overall quality and text quality.

For most models there are 30 prompts × 5 seeds (150 images); this set supports the best-of-5 selection experiment in the paper. Three models have one image per prompt (30 images).

Model	Images
FLUX 1.1 Pro, FLUX.2 Max, Ideogram v3 Turbo, Imagen 4 Fast, Nano Banana, Qwen-Image, SDXL, Seedream 4.5, Z-Image-Turbo	150 each
GPT Image (ChatGPT), FLUX.1 Kontext Pro, Nano Banana Pro	30 each

Structure

tiqa_images/
├── TIQA_Images.csv
├── tiqa_images_overall/
│   └── <model>/p<NN>/<seed>.png              # NN = prompt id 01–30, seed = 0–4
└── tiqa_images_text_only/
    └── <model>/p<NN>/
        ├── <seed>_whiten_.png
        └── <seed>/<seed>_crop_<k>.png

TIQA_Images.csv

One row per image. Main columns:

Column	Description
image	Image name, <model>-p<NN>-<seed>.png
generator_name, prompt_num, seed_num	Model, prompt id and seed
result_x, len_x	Individual text-quality ratings and their number
score_text	Text-quality MOS (mean of result_x)
result_y, len_y	Individual overall-quality ratings and their number
score_overall	Overall-quality MOS (mean of result_y)

The remaining columns (ocr_confidence, *_score) hold OCR confidence and predictions of automatic quality metrics.

---

Citation

@article{koltsov2026tiqa,
  title   = {TIQA: Human-Aligned Perceptual Text Quality Assessment in Generated Images},
  author  = {Koltsov, Kirill and Gushchin, Aleksandr and Vatolin, Dmitriy and Antsiferova, Anastasia},
  journal = {arXiv preprint arXiv:2603.07119},
  year    = {2026}
}