GPT Image 2 on PaperBanana — Finally, Legible Text in Your Diagrams

The Problem Every Researcher Knows

You draft a transformer-attention figure. The prompt describes the layers, the heads, the residual connections. The generated image looks beautiful — until you zoom in and the axis says "Atteniton Layerr," the legend says "Qury / Kye / Vauel," and the formula in the corner is just a blurry ink blot.

We've heard this complaint more times than we can count. It's the single most common reason PaperBanana users asked for refunds last quarter.

That problem is solved.

What Changed: GPT Image 2

OpenAI released GPT Image 2 on April 21, 2026 — the first image model with dedicated text-rendering improvements. Independent reviewers measured near-99% typography accuracy, up from 90–95% in the previous generation. It's currently ranked #1 across all Image Arena leaderboards with a +242 Elo lead on text-to-image — the widest margin any image model has held since the benchmark began.

As of today, GPT Image 2 is the first option in the Model dropdown on PaperBanana's generator.

What It Means for Academic Figures

GPT Image 2 handles the things that broke earlier models:

• Axis labels — x/y ticks render as actual numbers and units, not smudges
• Flowchart box text — short phrases ("Encoder", "Softmax", "Cross-Attention") stay legible at normal figure resolution
• Formula fragments — inline math like y = Wx + b or softmax(QK^T/√d) renders recognizably
• Multilingual labels — tested on English, Chinese, Japanese, Korean; mixed-language diagrams stay consistent

Pricing

Flat pricing. No plan-tier discount. Works with your existing subscription balance or one-time credit packs. See the Pricing page for credit bundles.

How to Use It

1. Open the generator
2. In the Model dropdown, pick GPT Image 2 (it's marked NEW and set as the default)
3. Choose your aspect ratio — auto is the safe default, or pick a specific ratio from 1:1 to 21:9
4. Pick Quality: Default for 5 credits, HD for 15 credits
5. Output format defaults to JPEG — switch to PNG or WebP if you need transparency or smaller file sizes
6. Write your prompt and generate

Prompting Tips for Text-Heavy Figures

The model renders text well, but prompt structure matters:

• Quote the exact text you want inside the image with double quotes, e.g. a flowchart with boxes labeled "Tokenizer", "Encoder", "Decoder"
• Keep box labels short — 1-5 words per element. 20-word paragraphs inside a figure still break legibility at normal resolution.
• Describe layout explicitly — "three stacked rectangles connected by downward arrows" gives the model a scaffold it can fill text onto
• Use image-to-image mode when you have a rough sketch — GPT Image 2 will keep your layout and clean up the text

Honest Limitations

We want you to spend credits, not request refunds. Here's what GPT Image 2 is not yet good at:

• Dense paragraph text inside an image (20+ words in a single block) — legibility still degrades
• Hand-drawn scientific notation on the level of LaTeX — it renders common math symbols, but complex notation (tensor indices, custom operators) is still unreliable
• Exact reproduction of copyrighted logos or journal templates — don't use it to fake IEEE/ACM formatting

If your figure has more than ~50 text elements, or requires publication-grade symbolic notation, use GPT Image 2 for the base layout and edit the final text in Adobe Illustrator or Inkscape.

A Note to Users Who Refunded Last Quarter

You told us the generated text was unusable for papers. You were right. The old models couldn't do it.

This model can.

If you refunded between January and March 2026 specifically because of garbled text, check your in-app notifications — we've sent a personal single-use discount code so you can verify the fix on your own figures at a reduced cost. No forms, no re-signup.

Try It Now

Open the generator and point the Model dropdown at GPT Image 2. One generation is enough to tell if it works for your figures.

If you have questions or want to see a specific diagram type benchmarked, reach out through the in-app feedback button — we read every message.