TextRefinementBench

English | 简体中文

TextRefinementBench is a benchmark for glyph-conditioned local text refinement. Every item is a full-resolution image paired with the text regions that should be re-rendered: each target carries the intended text, the polygon of that region, and its axis-aligned box in original-image pixels. A model is given the image and the targets and has to regenerate those regions in place, so the benchmark measures both whether the intended text comes out correct and complete and whether everything outside the edited regions survives untouched.

The test split holds 960 images and 2,357 target regions.

Dataset configurations

images

The test split contains 960 rows:

• image_id: image identifier, also the sample name used across the annotation rows.
• image: embedded RGB PNG image, byte-identical to the source file.
• image_path: canonical relative path, images/<image_id>.png.
• width, height: image dimensions.

annotations

The test split contains 2,357 rows, one per target region:

• sample_id: unique target-level identifier, <image_id>-<item_index>.
• image_id: foreign key into the images configuration.
• image_path: canonical relative image path.
• item_index: index of the target within its image, in annotation order.
• text: the text that region should contain.
• box: four polygon corner points [x, y] in original-image pixels.
• rect: axis-aligned [x0, y0, x1, y1] crop rectangle in original-image pixels.
• annotation_status: how the targets of that image were produced, either existing_human_annotation (167 images) or reviewed_miner_proposals (793 images).

Join the two configurations on image_id.

Loading

from datasets import load_dataset

images = load_dataset(
    "parquet",
    data_files={"test": "data/images/test-00000-of-00001.parquet"},
    split="test",
)
annotations = load_dataset(
    "parquet",
    data_files={"test": "data/annotations/test-00000-of-00001.parquet"},
    split="test",
)

Group the annotation rows by image_id to recover the ordered target list of each image.

Running a refinement model

Each annotation row is one edit request. With the PixText inference interface, the targets of a single image become a list of edits:

from pixtext import PixTextPipeline, TextEdit

pipeline = PixTextPipeline.from_pretrained("models/PixText-MeanFlow")
edits = [TextEdit(bbox=tuple(row["rect"]), text=row["text"]) for row in image_rows]
result = pipeline.edit("images/sample_000001.png", edits=edits, seed=42)

Boxes are half-open intervals in original-image pixel coordinates and targets are processed in annotation order. Pixels outside the edited footprints are expected to be preserved exactly, so preservation is part of the measurement.

License

The benchmark images and annotations are released under the Apache License 2.0. See LICENSE.