BloomBerry/figma-slide-benchmark

Figma Slide Editing Benchmark

Benchmark accompanying our EMNLP 2026 Industry Track paper "ACE: A Self-Correcting Agentic Canvas Editor for Multi-Slide Presentation Automation".

📄 Paper / archive URL: to be updated (xxx.archive — placeholder) 💻 Code: https://github.com/BloomBerry/agentic-canvas-editor

Overview

Each benchmark item is a slide-editing task defined as a pair of Figma Slides documents:

• *_TestA — the input deck the agent starts from.
• *_GroundTruthA — the reference deck representing a correct edit.

The benchmark contains 94 task pairs (188 decks) covering a wide range of real-world slide edits — e.g. converting tables to charts, fixing text placement/overflow, applying auto-layout, recoloring for colorblind palettes, translation, theme changes, adding agenda/bibliography slides, and more.

Directory layout

<CaseName>_TestA/            # input deck
<CaseName>_GroundTruthA/     # reference (edited) deck
  full_document.json         # full canvas/document structure (all slides)
  jsons/slide_XXX.json       # per-slide canvas structure
  frames/slide_XXX.png       # rendered PNG of each slide (1920x1080)
  index.html                 # simple viewer
  meta.json                  # file_key, slide_count, per-slide node ids & sizes

GroundTruth decks additionally include *_baseline variants (full_document_baseline.json, frames_baseline/) capturing the pre-edit state.

Usage

Clone the repository or stream files with huggingface_hub:

from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="BloomBerry/figma-slide-benchmark", repo_type="dataset")

Pair each *_TestA with its matching *_GroundTruthA by case name.

License & attribution

This dataset is released under CC BY 4.0.

The underlying slide designs originate from the Figma Community, whose free files are distributed under CC BY 4.0. When redistributing or building upon this data, you must give appropriate credit to the original Figma Community creators (see each case's meta.json file_key for provenance) in addition to citing this benchmark.

Citation

@inproceedings{ace-2026,
  title     = {ACE: A Self-Correcting Agentic Canvas Editor for Multi-Slide Presentation Automation},
  author    = {MIRIDIH},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track (EMNLP 2026)},
  year      = {2026},
  note      = {Archive URL to be updated},
  url       = {xxx.archive}
}