A Reproducible Test for Readable Text and Multi-Reference Consistency in AI Images

AI image models are easy to judge with one attractive sample and surprisingly hard to evaluate as production tools. A useful test needs repeatable inputs, clear pass/fail criteria, and more than one attempt.

This post describes a small benchmark for three common jobs:

1. rendering readable text inside an image,
2. combining several references without losing identity, and
3. making a local edit without rebuilding the composition.

For these tests I used the Nano Banana Pro workflow in PixMind. The same test structure can be applied to other image models, so it is useful for model selection rather than just prompt experimentation.

Why a reproducible test matters

A prompt such as “make a beautiful product poster” hides most failure modes. The model can choose an easy layout, avoid small text, change the product, or resolve ambiguity in a way that looks plausible but is wrong.

A better evaluation fixes the important variables:

• the exact copy that must appear,
• the references that must remain recognizable,
• the position of key objects,
• the requested edit boundary,
• aspect ratio and output size, and
• the number of attempts allowed.

I run each case three times with the same prompt. I do not silently repair spelling or composite the result in another editor before scoring it. This makes failures visible and helps distinguish a reliable workflow from a lucky sample.

Test 1: readable text in a structured poster

The first case is a fictional launch poster. It contains a headline, a subtitle, three feature labels, and a call-to-action. The point is not award-winning typography. The point is whether the model can preserve hierarchy and reproduce supplied copy.

Create a vertical 4:5 launch poster for a fictional productivity app.
Use a clean editorial grid with generous spacing.

Render this copy exactly:
Headline: TURN IDEAS INTO ACTION
Subtitle: A focused workspace for creative teams
Feature labels: PLAN / CREATE / REVIEW
CTA: START YOUR FIRST PROJECT

Use a deep navy background, warm white type, and one lime accent.
Do not add any other words, logos, badges, or watermarks.
Keep every line fully visible and leave safe margins around the text.

I score the output on five checks:

• every required string is present,
• spelling and punctuation match,
• visual hierarchy follows the requested order,
• no invented text appears, and
• all copy remains legible at normal viewing size.

The “do not add any other words” instruction matters. Many polished generations fail because they invent labels, decorative microcopy, or logo-like letterforms.

For multilingual testing, I repeat the prompt with one non-English subtitle instead of translating the entire poster. Mixing scripts in one composition is a more demanding and realistic test for campaign assets.

Test 2: multi-reference composition

The second case uses three references with different responsibilities:

• reference A defines the product,
• reference B defines the person or character,
• reference C defines composition and lighting.

The prompt should explicitly assign each reference a role. Saying only “use these references” forces the model to guess which properties matter.

Build a 16:9 product hero image.

Reference A is the exact product reference. Preserve its silhouette,
materials, color, button placement, and visible branding.
Reference B is the exact person reference. Preserve facial identity,
hair, age, and wardrobe colors.
Reference C is a layout reference only. Follow its camera angle,
negative space, and soft side lighting, but do not copy its objects.

Place the person on the left holding the product naturally.
Keep the right third uncluttered for headline text.
Use a restrained studio background with no extra accessories.

I compare identity separately from style. A result can look coherent while still changing a product’s geometry or a character’s face. My checklist is:

• product geometry and controls,
• brand color and material,
• face and hair consistency,
• pose plausibility,
• composition requested from reference C, and
• absence of objects borrowed accidentally from the layout reference.

This also reveals whether adding references improves control or increases prompt conflict. If the result degrades, I simplify the task: lock the product and person first, then request the layout in a second step.

Test 3: a bounded local edit

The third case starts from the strongest poster result and requests one narrow change.

Edit only the call-to-action button.
Change its fill from lime to coral and replace its text with:
TRY THE WORKFLOW

Preserve every other element exactly: headline, subtitle, feature labels,
font style, object positions, background, lighting, crop, and aspect ratio.
Do not regenerate or reinterpret unchanged regions.

The button edit is easy. The real test is whether unrelated regions remain stable. I compare before and after images at full size and note any drift in:

• face or product shape,
• text outside the edit region,
• spacing and alignment,
• color grading,
• crop and camera position.

A local edit that creates a nicer image but changes the headline is still a failed edit. In production, predictability is usually more valuable than an unsolicited redesign.

Record results without fooling yourself

A simple table is enough:

Case	Attempt	Required text	Reference fidelity	Layout	Unwanted changes	Pass
Poster	1	5/5	n/a	correct	none	yes
Poster	2	4/5	n/a	correct	added label	no
References	1	n/a	product drift	correct	none	no
Local edit	1	5/5	stable	stable	subtitle moved	no

Keep failed attempts. Deleting them creates survivorship bias and makes every model appear more reliable than it is.

I also record the full prompt, reference order, model/version, ratio, resolution, and date. Model behavior can change, and a prompt that worked months ago may not reproduce the same result later.

Prompt design lessons

Assign responsibilities

Tell the model what each reference controls. “Reference A controls product identity” is more actionable than “inspired by A.”

Separate invariants from preferences

Invariant details must survive every attempt: exact copy, product geometry, person identity, and edit boundary. Style, mood, and decoration are preferences. Put invariants first and state them concretely.

Make negative constraints specific

“Do not add any other words” is better than “keep it clean.” “Do not borrow objects from the layout reference” is better than “preserve accuracy.”

Reduce simultaneous goals

When one prompt asks for identity, typography, layout, storytelling, and a complex edit, failure becomes difficult to diagnose. Split the workflow into generation, selection, and bounded refinement.

Judge at delivery size

Text that looks correct in a thumbnail may contain broken characters at full size. Conversely, tiny decorative imperfections may not matter in a social post. Score the result in the context where it will be used.

Choose a model by failure cost

The best model is task-dependent. For a mood board, variety may matter more than exact text. For a product launch image, one changed button or invented label can invalidate the asset.

Choose based on the most expensive likely failure:

• use text-focused evaluation for posters, menus, infographics, and UI-like graphics;
• prioritize reference fidelity for catalog images and campaign variants;
• prioritize edit stability when an approved composition must be refined;
• prioritize iteration speed when exploring directions before production.

This benchmark does not reduce image quality to a single score. It makes trade-offs visible. A model may be excellent at structured typography but less predictable with multiple references, or strong at identity consistency but prone to global changes during editing.

A practical workflow

My production sequence is:

1. define the deliverable and its non-negotiable details,
2. upload references in a deliberate order,
3. assign each reference one clear role,
4. generate three controlled attempts,
5. score all attempts with the same checklist,
6. choose one result for a bounded edit,
7. compare changed and unchanged regions, and
8. save the prompt and settings with the final asset.

For a longer walkthrough with more prompt patterns, see the extended Nano Banana Pro image guide.

The core idea is simple: evaluate image models like production systems, not portfolios. A repeatable test reveals where a model saves time, where it needs tighter prompting, and where a different tool is the safer choice.