---
license: cc-by-4.0
language:
- en
tags:
- art
- design
- agent
- tool
configs:
- config_name: default
  data_files:
  - split: replication
    path: "replication_gt/*.jsonl"
  - split: modification
    path:
    - "modification_gt/task-1/*.jsonl"
    - "modification_gt/task-2/*.jsonl"
    - "modification_gt/task-3/*.jsonl"

---
<style>
.figure-tag {
  display: inline-flex;
  align-items: center;
  justify-content: center;
  width: 20px;
  height: 20px;
  margin-right: 6px;
  margin-left: 2px;
  background: #000;
  color: #fff;
  font-weight: 700;
  font-size: 14px;
  line-height: 1;
  border-radius: 1px;
  vertical-align: middle;
}
</style>

# CANVAS
This repository contains the dataset accompanying the paper [CANVAS: A Benchmark for Vision-Language Models on Tool-Based UI Design (AAAI 2026)](https://canvas.kixlab.org/).
CANVAS designed to evaluate a VLM's capability to generate a UI design with tool invocations in two tasks: <span class="figure-tag">A</span>**Design Replication** and <span class="figure-tag">B</span>**Design Modification**.
![canvas-huggingface-infro](https://cdn-uploads.huggingface.co/production/uploads/63ba74e0d90985e7acd6a52a/FK4EvzHbZD4O9rUVzAM_Z.png)

**A. Design Replication**. Single Tasks involving restoring (implementing) a given UI image as is.

**B. Design Modification**. Multiple Tasks involving modifying or transforming existing UIs.
![canvas-modification-task](https://cdn-uploads.huggingface.co/production/uploads/63ba74e0d90985e7acd6a52a/B3sZwOCvDtEm9EY8FmlAC.png)


## Dataset Statistics
CANVAS contains 598 design tasks from 3,327 mobile UI designs across 30 categories. Each design includes an image and a JSON file representing the Figma node structure. This file organizes components (e.g., vector, text, rectangle) in a hierarchical tree with attributes like position, size, and color.
![image](https://cdn-uploads.huggingface.co/production/uploads/63ba74e0d90985e7acd6a52a/mkysMwugYSH_QoqmJHjHR.png)
<span class="figure-tag">A</span>Frequency of the five most frequent UI design types, with other categories grouped. <span class="figure-tag">B</span>The distribution of node tree depth is similar across the replication and modification sets with a Gaussian-like pattern. <span class="figure-tag">C</span>The node count distribution is also similar across both sets. <span class="figure-tag">D</span>The skewed frequency of node types per design indicates common patterns in component usage.

---

## Directory Structure

```
canvas
├── modification_gt/
│   ├── task-1/       # Attribute Modification
│   ├── task-2/       # Component Insertion
│   └── task-3/       # Mode Change
└── replication_gt/   # UI Replication
```

---

## Modification Ground Truth (`modification_gt`)

### Task 1: Attribute Modification

Consists of 5 sub-tasks:
- Type 1: Fill Color Adjustment (20 examples) - `color_adjustment`
- Type 2: Size Adjustment (20 examples) - `size_adjustment`
- Type 3: Text Content Change (20 examples) - `text_content_change`
- Type 4: Corner Radius Change (20 examples) - `corner_radius_change`
- Type 5: Position Adjustment (20 examples) - `position_adjustment`

#### File Structure

Each task scenario consists of a "base" and "target" file pair:
- **base**: UI design before modification
- **target**: UI design after modification

Each scenario comprises 8 files (4 for base, 4 for target):

```
{sub_task_type}-gid{id}-{target|base}.json       // Full Figma data
{sub_task_type}-gid{id}-{target|base}-meta.json  // Metadata
{sub_task_type}-gid{id}-{target|base}.svg        // SVG
{sub_task_type}-gid{id}-{target|base}.png        // PNG
```

Example: `color_adjustment-gid1-base.json`

#### Metadata Structure

Each metadata file (`*-meta.json`) includes:
- `url`: Source Figma file URL
- `license`: License information (typically CC BY 4.0)
- `node_id`: Figma node ID
- `sub_task_type`: Sub-task type (e.g., "color_adjustment")
- `instruction`: Text instruction describing the required modification

### Task 2: Component Insertion

Adding a component to an existing UI.

#### File Structure

Each task scenario consists of a triplet of "base", "target", and "component" files:
- **base**: UI design before modification
- **target**: UI design after modification
- **component**: The UI component being added

Each scenario comprises 12 files (4 for base, 4 for target, 4 for component):

```
gid{id}-target.json      // Full Figma data
gid{id}-target-meta.json // Metadata
gid{id}-target.svg       // SVG
gid{id}-target.png       // PNG
gid{id}-base.json      // Full Figma data
gid{id}-base-meta.json // Metadata
gid{id}-base.svg       // SVG
gid{id}-base.png       // PNG
gid{id}-component.json      // Full Figma data
gid{id}-component-meta.json // Metadata
gid{id}-component.svg       // SVG
gid{id}-component.png       // PNG
```

Example: `gid1-base.json`

#### Metadata Structure

Each metadata file (`*-meta.json`) includes:
- `url`: Source Figma file URL
- `license`: License information (typically CC BY 4.0)
- `node_id`: Figma node ID
- `instruction`: Text instruction describing how to add the component

### Task 3: Mode Change

Day & Night Screen State Change.
Total of 100 examples (10 image groups x 10 pairs).

#### File Structure

Each task scenario consists of a "base" and "target" file pair:
- **base**: UI design before modification
- **target**: UI design after modification

Each scenario comprises 8 files (4 for base, 4 for target):

```
gid{group_id}-{subindex}-target.json      // Full Figma data
gid{group_id}-{subindex}-target-meta.json // Metadata
gid{group_id}-{subindex}-target.svg       // SVG
gid{group_id}-{subindex}-target.png       // PNG
gid{group_id}-{subindex}-base.json      // Full Figma data
gid{group_id}-{subindex}-base-meta.json // Metadata
gid{group_id}-{subindex}-base.svg       // SVG
gid{group_id}-{subindex}-base.png       // PNG
```

Example: `gid10-1-base.json`

#### Metadata Structure

Each metadata file (`*-meta.json`) includes:
- `url`: Source Figma file URL
- `license`: License information (typically CC BY 4.0)
- `node_id`: Figma node ID
- `gid`: Image group ID
- `subindex`: Pair index within the image group (1-10)
- `instruction`: Text instruction describing the required modification

---

## Replication Ground Truth (`replication_gt`)

Tasks involving restoring (implementing) a given UI image as is.
Total of 298 examples.

#### File Structure

In this dataset, there is no distinction between "base" and "target"; each UI design is an independent data point.

Each data point consists of 4 files:

```
gid{group_id}-{index}.json      // Full Figma data
gid{group_id}-{index}-meta.json // Metadata
gid{group_id}-{index}.svg       // SVG
gid{group_id}-{index}.png       // PNG
```

Example: `gid1-5.json`

#### Metadata Structure

Each metadata file (`*-meta.json`) includes:
- `url`: Source Figma file URL
- `license`: License information (typically CC BY 4.0)
- `node_id`: Figma node ID
- `type_id`: UI type ID
- `type_name`: UI category name (e.g., "Maps & Location")
- `json_node_depth`: Depth of the Figma node tree
- `json_node_count`: Total number of nodes
- `json_node_type_count`: Count by node type (FRAME, GROUP, VECTOR, TEXT, etc.)
- `difficulty`: Difficulty level (e.g., "standard")




## BibTex
If you use this dataset, please cite:

```
@article{jeong2025canvas,
      title={CANVAS: A Benchmark for Vision-Language Models on Tool-Based User Interface Design}, 
      author={Daeheon Jeong and Seoyeon Byun and Kihoon Son and Dae Hyun Kim and Juho Kim},
      year={2025},
      eprint={2511.20737},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.20737}, 
}
```