Dataset Card for Crello

Dataset Summary

The Crello dataset is compiled for the study of vector graphic documents. The dataset contains document meta-data such as canvas size and pre-rendered elements such as images or text boxes. The original templates were collected from crello.com (now create.vista.com) and converted to a low-resolution format suitable for machine learning analysis.

Usage

import datasets

dataset = datasets.load_dataset("cyberagent/crello")

Old revision is available via revision option.

import datasets

dataset = datasets.load_dataset("cyberagent/crello", revision="3.1")

Supported Tasks and Leaderboards

CanvasVAE studies unsupervised document generation.

Languages

Almost all design templates use English.

Dataset Structure

Data Instances

Each instance has scalar attributes (canvas) and sequence attributes (elements). Categorical values are stored as integer values. Check ClassLabel features of the dataset for the list of categorical labels.

{'id': '592d6c2c95a7a863ddcda140',
 'length': 8,
 'group': 4,
 'format': 20,
 'canvas_width': 3,
 'canvas_height': 1,
 'category': 0,
 'title': 'Beauty Blog Ad Woman with Unusual Hairstyle',
 'type': [1, 3, 3, 3, 3, 4, 4, 4],
 'left': [0.0,
  -0.0009259259095415473,
  0.24444444477558136,
  0.5712962746620178,
  0.2657407522201538,
  0.369228333234787,
  0.2739444375038147,
  0.44776931405067444],
 'top': [0.0,
  -0.0009259259095415473,
  0.37037035822868347,
  0.41296297311782837,
  0.41296297311782837,
  0.8946287035942078,
  0.4549448788166046,
  0.40591198205947876],
 'width': [1.0,
  1.0018517971038818,
  0.510185182094574,
  0.16296295821666718,
  0.16296295821666718,
  0.30000001192092896,
  0.4990740716457367,
  0.11388888955116272],
 'height': [1.0,
  1.0018517971038818,
  0.25833332538604736,
  0.004629629664123058,
  0.004629629664123058,
  0.016611294820904732,
  0.12458471953868866,
  0.02657807245850563],
 'opacity': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 'text': ['', '', '', '', '', 'STAY WITH US', 'FOLLOW', 'PRESS'],
 'font': [0, 0, 0, 0, 0, 152, 172, 152],
 'font_size': [0.0, 0.0, 0.0, 0.0, 0.0, 18.0, 135.0, 30.0],
 'text_align': [0, 0, 0, 0, 0, 2, 2, 2],
 'angle': [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
 'capitalize': [0, 0, 0, 0, 0, 0, 0, 0],
 'line_height': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
 'letter_spacing': [0.0, 0.0, 0.0, 0.0, 0.0, 14.0, 12.55813980102539, 3.0],
 'suitability': [0],
 'keywords': ['beautiful',
  'beauty',
  'blog',
  'blogging',
  'caucasian',
  'cute',
  'elegance',
  'elegant',
  'fashion',
  'fashionable',
  'femininity',
  'glamour',
  'hairstyle',
  'luxury',
  'model',
  'stylish',
  'vogue',
  'website',
  'woman',
  'post',
  'instagram',
  'ig',
  'insta',
  'fashion',
  'purple'],
 'industries': [1, 8, 13],
 'color': [[153.0, 118.0, 96.0],
  [34.0, 23.0, 61.0],
  [34.0, 23.0, 61.0],
  [255.0, 255.0, 255.0],
  [255.0, 255.0, 255.0],
  [255.0, 255.0, 255.0],
  [255.0, 255.0, 255.0],
  [255.0, 255.0, 255.0]],
 'image': [<PIL.PngImagePlugin.PngImageFile image mode=RGBA size=256x256>,
  <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=256x256>,
  <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=256x256>,
  <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=256x256>,
  <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=256x256>,
  <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=256x256>,
  <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=256x256>,
  <PIL.PngImagePlugin.PngImageFile image mode=RGBA size=256x256>]}

To get a label for categorical values, use the int2str method:

data = dataset['train'] # obtain the train set
key = "font"
example = data[0] # obtain first sample in train set

data.features[key].feature.int2str(example[key]) # obtain the text equivalent of the encoded values

Data Fields

In the following, categorical fields are shown as categorical type, but the actual storage is int64.

Canvas attributes

Element attributes

Note that the color and pre-rendered images do not necessarily accurately reproduce the original design templates. The original template is accessible at the following URL if still available.

https://create.vista.com/artboard/?template=<template_id>

left and top can be negative because elements can be bigger than the canvas size.

Data Splits

The Crello dataset has 3 splits: train, validation, and test. The current split is generated based on appearance-based clustering.

Visualization

Each example can be visualized in the following approach using skia-python. Note the following does not guarantee a similar appearance to the original template. Currently, the quality of text rendering is far from perfect.

import io
from typing import Any, Dict

import numpy as np
import skia


def render(features: datasets.Features, example: Dict[str, Any], max_size: float=512.) -> bytes:
    """Render parsed sequence example onto an image and return as PNG bytes."""
    canvas_width = int(features["canvas_width"].int2str(example["canvas_width"]))
    canvas_height = int(features["canvas_height"].int2str(example["canvas_height"]))

    scale = min(1.0, max_size / canvas_width, max_size / canvas_height)

    surface = skia.Surface(int(scale * canvas_width), int(scale * canvas_height))
    with surface as canvas:
        canvas.scale(scale, scale)
        for index in range(example["length"]):
            pil_image = example["image"][index]
            image = skia.Image.frombytes(
                pil_image.convert('RGBA').tobytes(),
                pil_image.size,
                skia.kRGBA_8888_ColorType)
            left = example["left"][index] * canvas_width
            top = example["top"][index] * canvas_height
            width = example["width"][index] * canvas_width
            height = example["height"][index] * canvas_height
            rect = skia.Rect.MakeXYWH(left, top, width, height)
            paint = skia.Paint(Alphaf=example["opacity"][index], AntiAlias=True)

            angle = example["angle"][index]
            with skia.AutoCanvasRestore(canvas):
                if angle != 0:
                    degree = 180. * angle / np.pi
                    canvas.rotate(degree, left + width / 2., top + height / 2.)
                canvas.drawImageRect(image, rect, paint=paint)

    image = surface.makeImageSnapshot()
    with io.BytesIO() as f:
        image.save(f, skia.kPNG)
        return f.getvalue()

Dataset Creation

Curation Rationale

The Crello dataset is compiled for the general study of vector graphic documents, with the goal of producing a dataset that offers complete vector graphic information suitable for neural methodologies.

Source Data

Initial Data Collection and Normalization

The dataset is initially scraped from the former crello.com and pre-processed to the above format.

Who are the source language producers?

While create.vista.com owns those templates, the templates seem to be originally created by a specific group of design studios.

Personal and Sensitive Information

The dataset does not contain any personal information about the creator but may contain a picture of people in the design template.

Considerations for Using the Data

Social Impact of Dataset

This dataset was developed for advancing the general study of vector graphic documents, especially for generative systems of graphic design. Successful utilization might enable the automation of creative workflow that human designers get involved in.

Discussion of Biases

The templates contained in the dataset reflect the biases appearing in the source data, which could present gender biases in specific design categories.

Other Known Limitations

Due to the unknown data specification of the source data, the color and pre-rendered images do not necessarily accurately reproduce the original design templates. The original template is accessible at the following URL if still available.

https://create.vista.com/artboard/?template=<template_id>

Additional Information

Dataset Curators

The Crello dataset was developed by Kota Yamaguchi.

Licensing Information

The origin of the dataset is create.vista.com (formally, crello.com). The distributor ("We") do not own the copyrights of the original design templates. By using the Crello dataset, the user of this dataset ("You") must agree to the VistaCreate License Agreements.

The dataset is distributed under CDLA-Permissive-2.0 license.

Note

We do not re-distribute the original files as we are not allowed by terms.

Citation Information

@article{yamaguchi2021canvasvae,
  title={CanvasVAE: Learning to Generate Vector Graphic Documents},
  author={Yamaguchi, Kota},
  journal={ICCV},
  year={2021}
}

Releases

4.0.0: v4 release (Dec 5, 2023)

• Change the dataset split based on the template appearance to avoid near-duplicates: no compatibility with v3.
• Class labels have been reordered: no compabilitity with v3.
• Small improvement to font rendering.

3.1: bugfix release (Feb 16, 2023)

• Fix a bug that ignores newline characters in some of the texts.

3.0: v3 release (Feb 13, 2023)

• Migrate to Hugging Face Hub.
• Fix various text rendering bugs.
• Change split generation criteria for avoiding near-duplicates: no compatibility with v2 splits.
• Incorporate a motion picture thumbnail in templates.
• Add title, keywords, suitability, and industries canvas attributes.
• Add capitalize, line_height, and letter_spacing element attributes.

2.0: v2 release (May 26, 2022)

• Add text, font, font_size, text_align, and angle element attributes.
• Include rendered text element in image_bytes.

1.0: v1 release (Aug 24, 2021)

Contributions

Thanks to @kyamagu for adding this dataset.