Image generators have become very good at pictures. Text inside images is a different story. Ask a model for an infographic with a dozen small captions and the result often looks right from across the room and wrong up close. A year is off by a digit, a label drifts into the next column. For a mood board that is a nuisance. For a quarterly report, a board pack or an investor update, it is a governance problem.
A study published on October 1, 2026, puts numbers on that problem, model by model. It is narrow by design, which is part of what makes it useful.
What was measured
In September 2026, 29 participants received the same half-finished 1280 by 720 slide, a flat background with a title, plus a written brief: draw a horizontal timeline with exactly these lines, in this order, without touching the title. The 29 were 24 third-party image editors and five quality tiers of OpenAI’s GPT Image (version 2 at low, medium and high; version 2.5 at low and medium). Seven slides varied the number of blocks (3, 10 or 18), the script (Latin, Cyrillic or mixed) and the illustration style. Together the seven slides carry 71 lines of text.
The study produced 250 generations scored automatically with OCR, meaning software that reads text back out of an image. A person then checked 30 original images by eye against the reference text. All prompts, scores and images are published with the write-up, in this test of 29 image models on infographic slides with small text. The study was run by Wonderslide, a presentation software company whose own infographic feature uses GPT Image 2 at medium quality, which is worth knowing when you read the conclusions.
Three findings that matter for business use
The gap between models is large. In the human-reviewed first generation, GPT Image 2 medium and GPT Image 2.5 low each had 71 of 71 lines accepted. Nano Banana 2 Fast, the only third-party editor in that review, had 43 of 71. On automatic scoring across all participants, only the five GPT tiers cleared half the lines. Several well-known editors scored at or near zero on the automatic count, and two models, Qwen Image 3.0 and Qwen Image 3.0 Pro, refused every slide because the brief exceeded their prompt-length limit.
The gap also widened with density. At 18 blocks, every GPT tier kept at least 44% of lines on automatic scoring, and every other model kept at most 17%. A person who checked Nano Banana 2 Fast on that slide accepted 1 line of 18.
Paying more does not buy accuracy past a point. GPT Image 2 high reproduced 58 lines against 56 for medium on automatic scoring, at 3.7 times the price. Meanwhile GPT Image 2.5 low cost $0.245 across the seven slides against $0.770 for GPT Image 2 medium, roughly a third, and the study could not separate the two on text quality. At volume, cost and quality do not move together, which makes this a procurement decision.
Automated checking is unreliable too. This finding deserves the most attention. On 205 lines a person could judge, automatic OCR scoring disagreed with the human on 44, or 21.46%. In 41 of those 44 cases, the software flagged a line as wrong that the person accepted. The disagreement rate also varied by model, from 14.08% to 30.16%.
The practical meaning: if your plan was to let one AI generate the chart and another AI check it, both halves of that pipeline have known error rates. The checker is not neutral ground truth.
Why text is still hard
The study does not explain the mechanism, and it would be overreach to claim one from its data. The common hypothesis is that image models learn what text looks like rather than what it says, so small, dense, exact strings are where they slip. What the data does show is where they slip: dense layouts, non-Latin scripts for some models, and run-to-run variation. One GPT tier returned 8, 16 and 13 exact lines on the same 18-line slide across three identical runs on automatic scoring, though a person reviewing the first of those runs accepted all 18.
What this means for business reporting in 2026
- Model choice is a control, not a preference. If generated visuals carry numbers, choose the model as carefully as your spreadsheet tooling. On this test, 71 correct lines versus 43 is the difference between a usable chart and a correction notice.
- Human review stays in the loop. Automated OCR checks catch some errors and invent others. For anything external, a person compares every number with the source.
- Keep the numbers out of the pixels where you can. A native chart built from data cannot misspell a figure. Use image models for the illustration around the numbers, and let the presentation or reporting tool render the text itself.
- Re-test on your own templates. Results moved between identical runs, and the study used one template.
Limits worth stating
All seven slides were synthetic horizontal timelines. Every model was called through one API intermediary, on a catalog snapshot dated September 17, 2026, and prices come from that price list. Returned image sizes differed between models, and small text suffers on a small canvas. The human review covered the first generation of three models plus nine flagged images, and the reviewer saw the reference text, so the review was not blind. The results describe one kind of infographic, not every chart a business might need.
Bottom line
In autumn 2026, a handful of image models can render small infographic text reliably on a controlled template, and most cannot. The more important lesson for business teams is about verification. Neither the generator nor an automated checker is trustworthy on its own. Pick the model deliberately, keep critical numbers in native charts, and keep a person between the image and the board.
