---
license: cc-by-4.0
language:
- en
task_categories:
- image-to-image
- image-text-to-image
pretty_name: PaintBench
size_categories:
- 1K<n<10K
tags:
- benchmark
- visual-editing
- image-editing
- pixel-space-generation
- diffusion-models
- evaluation
dataset_info:
- config_name: PaintBench
  features:
  - name: category
    dtype: string
  - name: task
    dtype: string
  - name: mode
    dtype: string
  - name: visual_condition
    dtype: string
  - name: problem_id
    dtype: int32
  - name: instruction
    dtype: string
  - name: input_image
    dtype: image
  - name: answer_image
    dtype: image
  - name: metadata
    dtype: string
  splits:
  - name: test
    num_bytes: 42756475
    num_examples: 2016
  - name: dev
    num_bytes: 5801478
    num_examples: 280
  download_size: 46011937
  dataset_size: 48557953
- config_name: TinyGrafixBench
  features:
  - name: category
    dtype: string
  - name: task
    dtype: string
  - name: mode
    dtype: string
  - name: visual_condition
    dtype: string
  - name: problem_id
    dtype: int32
  - name: instruction
    dtype: string
  - name: input_image
    dtype: image
  - name: answer_image
    dtype: image
  - name: metadata
    dtype: string
  splits:
  - name: test
    num_bytes: 55530852
    num_examples: 600
  download_size: 55366623
  dataset_size: 55530852
configs:
- config_name: PaintBench
  data_files:
  - split: test
    path: PaintBench/test-*
  - split: dev
    path: PaintBench/dev-*
- config_name: TinyGrafixBench
  data_files:
  - split: test
    path: TinyGrafixBench/test-*
---


# PaintBench

<b><em>Deterministic Evaluation of Precise Visual Editing</em></b>

*Anonymized release for double-blind review.*

---

**A precise, deterministic visual editing benchmark.** Evaluates whether native pixel-space image generation models can execute "MS-Paint-style" edits — geometric transforms, color changes, structural manipulation, and symbolic-reasoning edits — with pixel-level correctness. Every problem is a `(input_image, instruction, answer_image)` triplet generated programmatically, so the answer is pixel-exact and the answer distribution is known by construction.

This dataset ships two configurations that share the same schema and evaluation pipeline:

| Config | Splits | Problems | Scope |
|---|---|---|---|
| `PaintBench` | `test`, `dev` | 2,016 + 280 | Main benchmark — 20 tasks × 8 visual conditions × 12 problems-per-cell |
| `TinyGrafixBench` | `test` | 600 | Chart-edit micro-benchmark — 5 chart types × 4 subtasks × 30 problems |

## Quick start

```python
from datasets import load_dataset

# Full PaintBench test set (1,920 scored problems + 96 preservation diagnostic)
ds = load_dataset("PaintBenchICLR2027/PaintBench", "PaintBench", split="test")
print(ds[0]["instruction"])
ds[0]["input_image"].show()
ds[0]["answer_image"].show()

# Stratified 280-problem dev split for fast iteration
dev = load_dataset("PaintBenchICLR2027/PaintBench", "PaintBench", split="dev")

# Filter by visual condition (one of 8 perturbation axes)
n_xhigh = ds.filter(lambda r: r["visual_condition"] == "n_xhigh")

# Filter to scored problems only (exclude the preservation diagnostic)
scored = ds.filter(lambda r: r["task"] != "preservation")

# Chart-edit subset
charts = load_dataset("PaintBenchICLR2027/PaintBench", "TinyGrafixBench", split="test")
```

## Dataset structure

### Schema (both configs)

| Column | Type | Description |
|---|---|---|
| `category` | string | Top-level grouping. PaintBench: one of 4 task categories; TGF: one of 5 chart types. |
| `task` | string | Canonical task name (e.g. `translation`, `bar_chart_add_bar`). |
| `mode` | string | Subtask / variant within a task, or `"default"` for tasks with no mode dimension (7 single-mode PaintBench tasks + all TGF tasks). |
| `visual_condition` | string | One of 8 PaintBench-side perturbation axes, or `""` for TGF rows (no visual-condition axis). |
| `problem_id` | int32 | Index within the `(category, task, mode)` cell. |
| `instruction` | string | Natural-language editing instruction. |
| `input_image` | image | The source image to edit. |
| `answer_image` | image | The pixel-exact expected output. |
| `metadata` | string | JSON dump of the per-problem context (seed, scene shapes / colors, condition parameters, etc.). |

### PaintBench task categories

| Category | Tasks |
|---|---|
| `geometric_transformation` | `translation`, `rotation`, `reflection`, `scaling`, `shearing` |
| `structural_manipulation` | `construction`, `removal`, `copying`, `border`, `cropping` |
| `color_change` | `recolor`, `flood_fill`, `blending`, `gradient`, `point_operations` |
| `symbolic_reasoning` | `comparison`, `ordering`, `pattern`, `counting`, `legend` |

The `preservation` task (96 problems, exposed as `task=="preservation"`) is a diagnostic that copies the input as the expected output and is excluded from aggregate scoring.

### Visual conditions

Each PaintBench problem is rendered under one of 8 conditions; each varies exactly one axis from baseline:

| `visual_condition` | What varies | Detail |
|---|---|---|
| `baseline` | (nothing) | 1024 × 1024 canvas, default palette, default density |
| `horizontal` | canvas aspect ratio | 1024 × 576 |
| `vertical` | canvas aspect ratio | 576 × 1024 |
| `nonstandard` | palette | Non-standard color palette |
| `striped` | background | Striped (vs. solid) background |
| `n_med` | scene density | Medium-density scene |
| `n_high` | scene density | High-density scene |
| `n_xhigh` | scene density | Extreme-density scene |

Each `(visual_condition, task, mode)` cell contains 12 / num_modes problems (12 for single-mode tasks, 6 each for two-mode, 4 each for three-mode). Across all 20 scored tasks this is 1,920 problems (240 per visual condition); the `dev` split picks the `slot=0` problem per cell for a 280-problem stratified subsample.

### TinyGrafixBench subtasks

| Chart type | Subtasks |
|---|---|
| `bar_chart` | `add_bar`, `sort_bars`, `remove_bar`, `recolor_bar` |
| `heatmap` | `add_cell`, `shift_heatmap`, `mask_cells`, `change_colormap` |
| `line_chart` | `draw_segments`, `normalize_series`, `filter_series`, `shade_interval` |
| `network` | `add_node`, `swap_nodes`, `remove_node`, `recolor_node` |
| `scatter_plot` | `draw_best_fit_line`, `swap_axes`, `remove_outlier`, `recolor_class` |

## The `dev` split

`PaintBench` config ships a 280-problem `dev` split alongside the full 2,016-problem `test` split. Construction: **one problem per `(visual_condition, task, mode)` cell** across the 20 scored tasks (35 task-modes × 8 visual conditions = 280; preservation excluded). The dev split is a strict subset of the test split, generated deterministically (`slot=0` of each cell), and is intended for fast model iteration — not as a standalone benchmark. Reproducibility: the same row's `(task, mode, visual_condition, problem_id)` quadruple uniquely identifies it in both splits.

## Evaluation

Per-problem scoring is pixel-comparison-based: for each pixel, compute the CIE76 ΔE between the model's output and the expected `answer_image`, threshold at multiple ΔE levels (0, 1, ..., 10), and compute IoU and edit / preservation accuracy. Aggregate scores use macro-averaging at the task / category / visual_condition / benchmark levels with task-bootstrap 95% CIs to match the sampling unit of the displayed mean.

## License

Released under [Creative Commons Attribution 4.0 (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/). All problems and images are generated programmatically — no third-party content.