Google Gemini vs. OpenAI: AI product images for cosmetics & beauty in practical testing

1. What we tested

Three image types with different requirements

For the test, we used only real products from the cosmetics and beauty segment.

Input setup:

• Product photos taken with iPhone, natural light, neutral background
• Uniform prompt structure across all models

Tested image types:

Image type	Goal
Studio shot	Product cutout against a neutral background
Model shot	Product presented by a model
Mood shot	Staged in an atmospheric setting

Tested models

• Google Gemini Nano Banana Pro
• Google Gemini Nano Banana 2
• OpenAI GPT Image 1.5

All models received the same input and the same prompts.

2. Legal basis: Are you allowed to use AI-generated images commercially?

AI-generated images from tools like Google Gemini or OpenAI may generally be used commercially. Both providers waive their right to claim ownership of the generated content.

Legal classification from a copyright perspective

Under German law, AI-generated images often do not reach the necessary level of originality to be classified as copyrightable works. In practice, this means that there is no exclusive protection against third parties – comparable results could theoretically be achieved by other users as well.

IP protection via contracts (business and enterprise plans)

• Both Google and OpenAI contractually indemnify business customers against copyright claims.
• These exemptions apply under certain conditions: The output must not have been altered, safety functions must remain active, and there must be no obvious legal infringement.
• Explicitly excluded are claims relating to brands and trademarks.

Data protection and model training

In business plans and via the API, prompts and generated content are generally not used for further model development. This may vary for consumer products depending on individual settings.

EU AI Act: Labelling requirement from August 2026

• From August 2, 2026, providers must make synthetic content technically identifiable.
• Professional users are required to identify deepfakes as such.
• Anyone using AI-generated product images should plan a labeling concept early on.

The exact requirements in the e-commerce context are currently still being defined.

3. Google Nano Banana for Cosmetics & Beauty

First, we looked at Google Nano Banana Pro: In our tests, this model delivered convincing results with cosmetic products – as long as typography doesn't play a central role. Product shape and presentation are captured very well, but a structural limitation becomes apparent with delicate typography on curved surfaces.

Test examples (excerpt) including ratings:

Skin cream, tube (Studio: 5/10, Model: 7/10, Mood: 7/10)

Observations

The product body and shape are rendered very faithfully. The main problem: The small lettering on the side ("pH-skin neutral", "Fragrance-free") on the curved surface of the tube is glitchy and pixelated – a consistent result across all tests. Capital letters like "CARE+REPAIR" on the main surface, however, are reproduced well.

The model looks natural. The mood shot in its atmospheric setting (velvet blanket, clean beauty setting) is also convincing – however, the problem with the labeling remains the same across all results.

Nail polish (Studio: 9/10, Model: 8/10, Mood: 7/10)

Observations

Text on a flat 2D surface with a manageable amount of text: no problem. Google Gemini renders the lettering correctly and clearly. Only when the model holds the car at an angle does the legibility deteriorate due to the perspective.

The model looks natural, the mood shots work well.

What works exceptionally well:

• The product body and shape are depicted true to the original.
• The models appear natural – in most tests they are not immediately recognizable as AI.
• Text on flat 2D surfaces with little text: no problems
• Single-shot capability: First attempt delivers usable results
• The atmospheric mood shots are convincing.

Limits:

• Small print on curved surfaces (page labels, ingredients): glitchy and pixelated
• Perspective worsens text: Simply tilting the product with a model is enough to destroy readability.
• "Text all over" products: not a suitable use case

Nano Banana Pro vs. Nano Banana 2: Progress on the font problem

On February 26, 2026, Google introduced DeepMind Nano Banana 2. Google positions both models with a clear division of tasks: Nano Banana Pro remains the model for use cases with the highest quality and accuracy requirements, while Nano Banana 2 is designed for fast generation, precise prompt fidelity, and integrated image grounding.

We retested the same skin cream with Nano Banana 2 and directly compared the results:

	Nano Banana Pro	Nano Banana 2
Product body / shape	True to the original	True to the original
Capital letters (main area)	Easily reconstructible	Easily reconstructible
Small print on curved paper	Glitchy, pixelated	Clear text, even with fine details
speed	standard	Significantly faster
Improvement over predecessor	–	Significant progress

The result is clear: Nano Banana 2 can even process fine lettering on products with lettering on 3D surfaces.

We then conducted another test with a heavily printed shampoo bottle:

The result: Even with this product heavily printed with text, Nano Banana 2 delivers reliable, high-quality output and is therefore a viable solution for AI-generated cosmetic products with a lot of labeling.

Use Cases:

Cosmetic brands where design, color, and form take center stage. Mood images and category headers for collection pages.

Products where ingredients, claims or care information must be legible on the product image (using Nano Banana 2).

4. OpenAI for cosmetics

For cosmetic studio shots, OpenAI delivers sharp and legible text by default, even on curved packaging surfaces. However, model and mood shots fall significantly short.

Test examples (excerpt) including ratings:

Udder care tube (Studio: 8/10, Model: 4/10, Mood: 4/10)

Observations

OpenAI also renders the small side labels correctly – "pH-neutral" and "fragrance-free" remain legible, even on the curved surface. The product body and shape are rendered very faithfully. The studio shot is clearly PDP-compatible.

The model shots, on the other hand, are a letdown: The model looks plastic and artificial – significantly less realistic than Gemini. The mood shots rely too heavily on the original image and barely create any independent staging.

Nail polish (Studio: 9/10, Model: 6/10, Mood: 6/10)

Observations

Text on a flat 2D surface: completely flawless and true to life. Even when the model holds the paint at an angle, the lettering remains significantly more legible in OpenAI than in Gemini. The model shot is technically fine, but appears less authentic.

Shampoo with "Text all over" (Studio: 2/10)

Observations

For products with wraparound lettering, the text becomes distorted and illegible even with OpenAI – even if all text is fully typed into the prompt.

Body lotion (Studio: 7/10, Model: 3/10, Mood: 4/10)

Observations

Moderate amount of text, white lettering: the studio shot is appealing. The critical problem becomes apparent in the model shot – the body lotion bottle is depicted in unrealistic XXL proportions. Even after two correction prompts, the problem persists. This is a deal-breaker for use on a PDP.

What works well:

• Text on 3D surfaces remains legible – even with curvature and perspective.
• Products with approximately 60% text coverage on the main surface: very well suited for studio shots.
• Labels and care labels are sharp and true to the original.

Critical weaknesses:

• The model shots look artificial – significantly less realistic than Gemini.
• Proportion issues: Product size in relation to hand/body is sometimes unrealistic, even after a correction prompt.
• Mood shots are under-staged: often only changes the background instead of restaging.
• Multiple iterations are needed: No "single shot" – often 2-3 attempts are required to achieve a usable result.
• "Text all over" products: not a suitable use case

Recommended Use Cases:

Studio shots for products where label readability and text accuracy are critical. Not suitable for model shots, series with the same models, or products with wraparound lettering.

5. Our recommendation: When to use which model for cosmetic product images?

There is no general answer – the decision depends on what type of image you need and how much text plays a role on your packaging.

OpenAI as a reliable choice for studio shots with moderate text

If product labels are relevant to PDP (Product Data Processing) – meaning ingredients, claims, or product information must be legible in the image – OpenAI is a good choice for studio shots. Important: OpenAI does not automatically correct flawed source images. The input quality must be better than with Gemini.

We recommend OpenAI in these cases:

• Products with moderate font on the main area (headlines, claims, capital letters)
• Studio shots for PDPs where label readability is crucial.
• Brands that can deliver clean, straight source photos

Not suitable: Model shots, mood shots, products with small, wraparound lettering, "text all over" products

OpenAI – Qualitative Assessment

category	Score (1–5)	Reason
To use	4	A clear advantage for text rendering in studio shots. Limited usability for model and mood shots.
Time saving / efficiency	3	Multiple iterations are needed, higher input quality requirements. Not a single shot.
Implementation effort (5 = very easy)	3	Requires clean source photos – no optimization buffer like with Gemini.
Scalability	2	Scalable for studio shots, but inconsistency in models limits its suitability for mass production.

Gemini as the first choice for mood and model shots

For atmospheric scenes and model shots, we recommend Gemini. Its single-shot capability and natural model rendering clearly make Gemini the ideal choice for these types of images. With the latest version, Google Nano Banana 2, you can reliably render even printed 3D objects with a high proportion of text.

We recommend Gemini in these cases:

• Mood images and category headers where design and atmosphere take center stage.
• Model shots where product shape and staging are important
• Brands with simple, large main text (also okay for Nano Banana Pro) and small print (using Nano Banana 2)

Gemini – Qualitative Assessment

category	Score (1–5)	Reason
To use	4	Excellent for staging and model shots; text on curved surfaces is possible with Nano Banana 2.
Time saving / efficiency	4	Single-shot capability significantly reduces iteration effort.
Implementation effort (5 = very easy)	4	A structured approach is necessary – model profiles, settings, clear prompts.
Scalability	4	Easily scalable for mood and category images without critical font requirements.