ablmontazer/persian-ai-slide-benchmark

Persian AI Slide Benchmark

How well do AI presentation generators handle Persian? A reproducible benchmark of right-to-left slide generation.

Paper and archived release: doi.org/10.5281/zenodo.23976939 (Zenodo). Author: Abolfazl Montazer Lotfelahi, Perizent. Data collected 5 November 2026.

Eight Persian prompts from eight domains, each asking for an eight-slide deck, were run once with default settings on the free plans of Gamma and Canva and on Perizent, a Persian-first generator built by the author. Every deck is scored in two ways:

1. Structural metrics read from the PowerPoint XML: RTL paragraph flags, alignment, Persian digits, zero-width non-joiners, native tables and their RTL flag, embedded fonts, unfilled placeholders.
2. A vision-LLM judge that rates every slide image 1 to 5 on RTL layout, letter shaping, bidi order, no overflow and Persian language, on two views: the tool's own PDF export (own) and the downloaded PPTX rendered by LibreOffice 26.2 with a neutral font set (neutral).

Headline results

Judge criterion (own export, 1 to 5)	Perizent (n=8)	Gamma (n=8)	Canva (n=2)
RTL layout	4.86	4.53	3.56
Letter shaping	5.00	5.00	5.00
Bidi order	4.92	4.80	3.88
No overflow	5.00	4.92	4.94
Persian language	4.98	4.77	4.88
Slides with any score of 2 or lower	3%	11%	25%

Gamma right-aligned a larger share of Persian paragraphs than Perizent (98% vs 90%) but left 44 unfilled bracket placeholders in one deck, and none of its 5 native tables carried the PPTX table-RTL flag. Perizent produced no native PPTX tables and forces Persian digit glyphs onto Latin tokens such as model names. Canva's free plan produced only 2 of the 8 decks before asking for an upgrade. See the paper for the full tables and discussion.

Contents

Path	What
data/judge_scores.csv	288 rows, one per slide, tool and view, with the five judge scores, the judge's issue note and the slide image path
data/xml_metrics.csv	Per-deck structural metrics from the PPTX XML
data/prompts.csv, prompts.json, requirements.json	The eight prompts and their checkable requirements
decks/<tool>/	Generated decks: PPTX, the tool's PDF, LibreOffice PDF (.neutral.pdf), per-deck metadata (date, plan, settings, errors, retries)
`slides/<own	neutral>///s-N.jpg`
scores/	Raw judge responses, XML metrics, placeholder and digit-mixing counts, human spot check
tools/	Collection, conversion, scoring, analysis and paper-building scripts
paper/	The paper (PDF and HTML) and figures

Limitations and conflict of interest

Small, single-run sample (8 prompts, 2 Canva decks); differences of a few tenths should not be over-interpreted. The judge agreed exactly with a human spot check on 87% of ratings. LibreOffice is a proxy, not PowerPoint. Results describe free plans and default settings on 5 November 2026.

The author develops and commercially operates Perizent. Prompts were fixed before any competitor deck was generated, scoring is automated with the published code, the judge is a third-party model, and Perizent's own failures are reported. No funding was received.

License

Paper, data and scores: CC BY 4.0. Code in tools/: MIT. Decks produced by third-party tools are included only as research evidence and remain subject to those tools' terms.

Citation

@misc{montazer2026persianslides,
  author    = {Montazer Lotfelahi, Abolfazl},
  title     = {How well do AI presentation generators handle Persian? A reproducible benchmark of right-to-left slide generation},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23976939},
  url       = {https://doi.org/10.5281/zenodo.23976939}
}