AI-Generated Slides Finally Look Real: This 9B Open Model Scores 7.94, a Hair Behind Google's Best

You know the tell. The slide your AI generated: the heading has a typo it invented, the bullet text overlaps the chart, and the whole thing has that glossy, slightly-off aesthetic that screams "made by a model." Presentation slides are one of the hardest things to generate well, because a single canvas demands five skills at once — accurate text rendering, following your instructions, grounding the visuals, decent design taste, and precise edits when you ask for changes. Most image models are good at maybe two of those.

A new paper from a 10-author team led by Jiahao Wang (submitted October 12, 2026) takes this problem head-on. It's called InfoVis, and it's a unified 9-billion-parameter open model that both generates and edits presentation slides. On their new benchmark, InfoVis beats every evaluated open-source model on all eleven metrics — and lands at 7.94 overall, within a hair of Google's proprietary Nano-Banana-2 at 7.99. The open slide-design race just got very interesting.

What the researchers did

The team's first observation was that nobody was training for this properly. Existing work covered only subsets of the slide skill set, and there was no large-scale slide-native training data or diagnostic evaluation. So they built both.

First, they formulated four slide-centric tasks — generation across three prompt settings (text prompt, document-to-slide, and layout-conditioned) plus instruction-guided editing — and built a scalable data engine to feed them: over 13 million continual-pretraining records and 1.9 million supervised fine-tuning records. That is slide-specific supervision at real scale.

Second, they introduced VisSlideBench, a benchmark of 1,527 manually screened instances covering all four tasks with multidimensional evaluation. Generation is scored on three dimensions — Content Faithfulness (does the slide say what you asked?), Instruction Following (did it obey your directions?), and Aesthetic Quality (does it look like a designer laid it out?). Editing gets two: Semantic Consistency and Presentation Quality. Crucially, the benchmark's automatic rankings agree with blinded human judgments — so the scores are not an artifact of a friendly evaluator.

Then they trained InfoVis on top of the open FLUX.2-klein-base-9B backbone. The results:

• Best open model, everywhere: InfoVis leads every evaluated open-source model on all eleven metrics, scoring 7.94 overall.
• Within 0.05 of Google's best: the proprietary Nano-Banana-2 scores 7.99 — a gap of just five hundredths.
• Slide training beats scale: slide-specific post-training alone added 1.6 points overall, and InfoVis outperforms even the 32B FLUX.2-dev. Specialized data beats a bigger model.
• Unification is nearly free: unlike existing unified open-source models, which trail task specialists by a wide margin, the unified InfoVis gives up almost nothing relative to single-task specialists.
• The honest gap: against proprietary models, the remaining gap is smallest in content faithfulness and largest in aesthetic quality — open models now get the content right; taste is the last frontier.

As a bonus, their document-to-slide workflow ranked second out of ten systems for Aesthetics on the external SlidesGen-Bench. The team says they will release the model, code, data, and benchmark.

Why it matters

Slides are the workhorse of knowledge work — pitches, lectures, client decks, conference talks. Today, generating a decent one with AI is a loop of disappointment: fix the typo, re-render, fix the layout, re-render, give up and open PowerPoint. InfoVis shows the fix isn't a bigger model or a better prompt — it's task-native training data. Thirteen million slide-specific training records did what 3x the parameters couldn't.

For developers and creators, this matters twice. First, it previews a near future where "make me a deck" actually works: generate from a document, then edit in place with plain instructions, all in one model. Second, the benchmark itself is a gift — VisSlideBench's five evaluation dimensions are the clearest checklist anyone has published for what "good" means in generated slides. You can grade any model, or your own prompting workflow, against them today.

And the open-source angle is real: an open 9B model sitting 0.05 points behind Google's flagship on a human-validated benchmark means independent developers will be able to self-host serious slide generation within months of the release.

How YOU can use it

InfoVis isn't released yet, so you can't download it today. But the paper's recipe translates directly into a workflow you can run now with the models you already have — and the researchers' own findings tell you exactly where to focus.

1. Steal their grading rubric

Before generating a single slide, adopt the VisSlideBench dimensions as your quality checklist:

• Content Faithfulness: every word on the slide must match your source. No invented stats, no paraphrased quotes that change meaning.
• Instruction Following: if you asked for three bullets, you get three bullets — not a paragraph.
• Aesthetic Quality: alignment, spacing, contrast, one idea per slide.

Grade every AI-generated slide on these three, 1–10. The paper's central finding is that faithfulness is now the solved part and aesthetics the gap — so spend your effort on the design dimension.

2. Use the three prompt settings deliberately

The paper trains generation across three settings, and they map neatly onto real workflows:

• Text prompt → slide: for single concept slides. Write the full slide content in the prompt — exact headings, exact bullet text — and let the model do layout.
• Document → slide: for turning a doc, paper, or report into a deck. Feed the source text and ask for one slide per key point, verbatim quotes preserved.
• Layout-conditioned: for brand consistency. Describe the layout ("title top-left, two-column body, chart placeholder right") and have the model fill it.

Pick the setting that matches the job instead of one generic "make a slide" prompt. This is the practical version of their data engine: the right framing per task.

3. Copy this prompt template

The biggest generation gains in the paper came in content faithfulness — the thing you can already force today by being explicit:

Create a single presentation slide, 16:9.

EXACT TEXT (render verbatim, no rewording, check spelling twice):
- Title: [your title]
- Bullets: [your exact bullet text]

LAYOUT: title top-left, bullets left column, simple chart visual right.
STYLE: dark background, high contrast, generous whitespace, minimal decoration.
DO NOT: invent text, overlap elements, or use more than 40 words total.

The "render verbatim" and "check spelling" instructions target the two failure modes the benchmark measures most harshly: hallucinated text and sloppy rendering.

4. Generate, then edit in place — don't regenerate

InfoVis's headline result is that one unified model does generation and editing nearly as well as specialists. Until it's released, replicate the pattern with a two-step loop: generate the slide once, then issue editing instructions ("move the chart right, enlarge the title, fix the typo in bullet two") rather than starting over. Editing preserves what already worked; regenerating re-rolls the dice on everything.

Try it tonight

Take one real slide you need this week. Generate it with your favorite image model using the prompt template above, then grade it on the three dimensions. If faithfulness or instruction-following scores below 8, the model isn't the problem — your prompt is. Tighten it and rerun. That's the InfoVis lesson in miniature: slide quality is a supervision problem, and your prompt is the supervision.

Key takeaways

• Unified beats specialized-at-scale: a single 9B model trained on slide-native data beats bigger general models and nearly matches Google's proprietary best.
• Data > parameters: slide-specific post-training added 1.6 points overall — more than tripling model size did.
• Faithfulness is solved; taste is next: the remaining gap to proprietary models is smallest on content accuracy, largest on aesthetics.
• Use the rubric today: Content Faithfulness + Instruction Following + Aesthetic Quality is a free, human-validated checklist for any slide workflow.
• Watch for the release: model, code, data, and benchmark are all promised open — this will be a self-hostable slide engine.

Read the original paper: "InfoVis: Unified Slide Image Generation and Editing" — Jiahao Wang, Xiyan Jiang, Chenyuan Wu, Haoqiang Zhang, Xin Luo, Yueze Wang, Defu Lian, Bo Zhang, Jiajun Zhang, Zheng Liu — arXiv:2610.16784 (cs.CV, submitted 12 Oct 2026).