Can an AI Image Generator Draw Your Pitch Deck Infographic?

Most pitch decks have at least one slide that is really an infographic. The traction timeline. The product roadmap. The go-to-market plan drawn as five boxes with arrows. Founders often build these at midnight, and in 2026 the temptation is obvious: describe the slide to an image generator and let it draw. The question an investor will not ask out loud, but will notice, is whether “Q3 2025” on your timeline actually says Q3 2025.

A published test from September 2026 answers that question better than most demos do, because it measured exactly this kind of slide.

The test, in one paragraph

Twenty-nine image models and model tiers were given the same task: add a horizontal timeline with specific captions to a slide that already had a background and a title, without changing anything else. Seven versions of the slide carried 3, 10 or 18 blocks, in English, Russian or a mix, with 71 captions in total. That is close to what a traction or roadmap slide looks like. The results were scored by text-recognition software across 250 images, and a person checked a sample of original images by eye. The data, prompts and images are public, and the write-up is here: which image models can draw an infographic with small text, tested on 29 models.

Results translated into pitch-deck terms

Only a few models are usable for text-heavy slides. When a person reviewed the first attempt on all seven slides, GPT Image 2 at medium quality and GPT Image 2.5 at low quality had every one of the 71 captions accepted. Nano Banana 2 Fast had 43 accepted, plus 19 readable but wrong, 8 unreadable and 1 missing. Many other popular editors barely registered on the automatic count. For a deck, 43 of 71 means roughly two captions in five need fixing, and you will not always spot which.

Your roadmap slide is the hard case. Roadmaps get crowded. The densest slide in the test had 18 blocks. There, the automatic count put every GPT tier at 44% of captions or better and every other model at 17% or worse. The human reviewer found only 1 correct caption of 18 in the Nano Banana 2 Fast image. If your roadmap has more than a handful of milestones, the choice of model matters more than the wording of the prompt.

Long briefs can be refused outright. Founders like detailed prompts. Some models do not. In the first round, 36 of 203 attempts were refusals because the brief was longer than the model’s prompt limit. Two models refused every slide. A refusal costs you time, not accuracy, but it is worth knowing before the night of the deadline.

Pitching in a second language. If you raise in more than one market, script matters for some models. On the same ten events, a person accepted 10 of 10 English captions from Nano Banana 2 Fast but only 2 of 10 in Russian. The two GPT tiers that were reviewed had all ten accepted in both languages. The test covered Russian only, so do not read it as a verdict on every alphabet.

The cost per slide is small. List prices per generation, through the API intermediary used in the test as of September 17, 2026: GPT Image 2.5 low $0.035, Nano Banana 2 Fast $0.045, GPT Image 2 medium $0.110, GPT Image 2 high $0.410. The top tier did not measurably beat medium on text. For a startup, even the expensive option is pocket change per slide, so the real cost is the review time when a model gets captions wrong.

Run it more than once

The same prompt does not always produce the same slide. Across three identical runs, GPT Image 2.5 medium never moved by more than one line on any slide by automatic count. On the 18-line slide, GPT Image 2 medium swung between 8 and 16 exact lines, although a person checking the first of those images accepted all 18, so part of that swing came from the scoring software. Either way, generate two or three versions and keep the best, rather than shipping the first one.

A workflow for founders

1. Write the captions first. Put the exact text of every milestone in a document. That document is your checklist later.
2. Keep captions short. A year and a few words per block. Fewer characters, fewer chances to fail.
3. Pick a model that held up on text. On this test, that meant one of the GPT Image tiers.
4. Generate a few versions. Choose the cleanest.
5. Proofread against your document, not against memory. Read every number aloud. Automatic checking is not a safe substitute: in the study it disagreed with the human reviewer on about one judged line in five.
6. Keep the real numbers native. Revenue, user counts and unit economics belong in an actual chart built from data, which cannot misspell a figure.

Presentation tools are starting to build this into their own workflow. Wonderslide, which ran the test, uses GPT Image 2 medium for its infographic mode, and the test was partly about whether a cheaper tier would do as well. Whatever you use, the proofreading step stays yours.

What the test cannot tell you

• It used one template: synthetic horizontal timelines. Funnel diagrams, org charts and maps were not tested.
• Models returned images of different sizes, so each count holds only at the size that model returned.
• A person reviewed only the first attempt of three models plus nine images already flagged as failures. Everything else is automatic scoring.
• Models update often. A result from September 2026 is a starting point for your own check, not a permanent ranking.

Bottom line

Yes, an image generator can draw the infographic slide in your pitch deck, if you choose one of the few models that handle small text and you proofread every caption. Investors forgive a plain slide. They remember a wrong date on a traction chart.