AI product images for fashion: Google Gemini vs. OpenAI in a practical test

1. What we tested: 3 image types, 2 models

Three image types with different requirements

We tested three image types with two AI models: studio shots, model shots, and mood shots . Both models received identical mobile phone photos as input and identical prompts.

Studio shots (Ghost Mannequin):
Product images for product detail pages – clothing floating freely on a neutral background without a visible model. The "invisible mannequin look" focuses purely on the product.

Model shots:
Models wear the clothing in a neutral studio environment. Model profiles in the prompts define age, body type, and specific features (tattoos, vitiligo, freckles).

Mood shots:
Atmospheric staging in defined settings (greenhouse, industrial loft, café).

Two AI models in direct comparison

We tested Google Gemini (Nano Banana Pro) and OpenAI GPT Image 1.5 in parallel. Both models used identical prompts to make the results comparable. Simple photos of the products from multiple perspectives (e.g., on hangers, laid flat) served as input – deliberately not optimized, professional studio shots, in order to reflect realistic conditions for retailers.

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

AI-generated images from Google Gemini and OpenAI can be used commercially . Neither provider claims ownership rights to the generated output.

Copyright classification:

Under current German law, AI-generated images are often not protected as independent copyrighted works if they lack sufficient human originality. Therefore, exclusivity against third parties is not guaranteed – similar outputs can also be created by other users.

Contractual IP exemptions (Business/Enterprise plans):

• Google and OpenAI offer contractual indemnifications for copyright claims.
• Valid under the following conditions: unmodified output, no bypassing of safety features, no knowledge of obvious violation.
• Exclusions for brand/trademark claims in commercial transactions

What should fashion brands pay particular attention to when it comes to AI-generated images?

Naturally, AI-generated images should also convey an authentic impression of the product. In the fashion industry, for example, it's common practice to photograph sample pieces that may differ slightly from the final production model. These industry-standard tolerances apply analogously to AI-generated product images – minor visual discrepancies between the AI ​​image and the actual product are within the usual range for fashion e-commerce. Nevertheless, it should be clear that customers rely on authentic product photographs, and significant deviations can lead to returns.

Not for training use:

Prompts and outputs are typically not used for model improvement in business/API offerings . Consumer apps can contribute to model improvement depending on their configuration.

Labelling requirement from 2 August 2026 (EU AI Act):

• Providers must make AI-generated content machine-readable and identifiable.
  
• The labeling concept for realistic AI-generated product images should be planned early on.
• Specific implementation details for e-commerce are currently being finalized.

3. Google Gemini for Fashion: Authentic models, consistent results

In our tests, Google Gemini delivered usable fashion images. Models appeared natural, features remained consistent across multiple images, and faulty inputs were automatically corrected.

Test examples (excerpt):

Observations

The cardigan was hanging crooked on the hanger in the input photo. Google Gemini corrects this automatically – the stripes are straight and neat in the generated image. The fabric drapes realistically, and the stripe pattern is captured precisely.

The drawback becomes apparent in the model rendering: the older model clearly appears to be AI-generated. The mood shots also exhibit a high degree of smoothness, which diminishes their authenticity.

Floral sweater

Observations

Here too, the original image wasn't ideal – Gemini automatically corrects the floral pattern and straightens it. The androgynous model looks surprisingly natural, and freckles are rendered in great detail. In this test, the research team wouldn't have immediately suspected that the model was AI-generated.

The café setting in the mood shot, however, stands out due to its excessive perfection – too polished, too perfectly staged. Less suitable for authentic shop images.

Shirt with print

Observations

The shirt print is reproduced exactly 1:1, and the fabric drapes realistically. The crucial test: The model's tattoos remain identical across multiple tests – same position, same design. This allows, for example, the creation of a reusable model library.

The model looks natural, not immediately recognizable as AI. The industrial loft setting works well and emphasizes the edgy look of the product.

Striped shirt

Observations

The model with vitiligo appears natural and authentic – here too, the research team didn't immediately suspect AI generation. The vitiligo patches remain consistent across all images.

The greenhouse setting was rated as creating the strongest mood – the atmosphere with deep green plants, rusty metal struts, and diffused light works exceptionally well. The product lighting remains clear, and the fashion focus is not lost.

What works exceptionally well:

• The models appear natural and realistic – in most cases not immediately recognizable as AI-generated.
• Single-shot capability: The first attempt delivered usable results in every test.
• Model features remain identical across multiple images (tattoos, vitiligo, freckles)
• Automatic optimization: Skewed or uneven products from source images are straightened and optimized in the output, and unfavorable perspectives are corrected.
• Format understanding: Strictly adheres to prompt specifications (e.g., instruction: product images on Shopify must be square)
• Mood shots: Atmospheric staging, such as a greenhouse setting, is convincing.

Limits:

• Labels with small print or typography on curved surfaces quickly become pixelated.
• Very smooth surfaces sometimes appear "too perfect" (clearly AI-generated).
• Some moody settings also appear too plastic and artificial, such as the café setting.

Recommended Use Cases:

Fashion brands that require corrective editing of the original images and value consistent models with appropriate staging. This applies when the presentation of text, such as information on the label, is not important.

4. OpenAI for Fashion: Better details, but inconsistent models

OpenAI delivers improved text readability and more realistic skin textures in the models. However, the high inconsistency of the outputs and the copying of input errors result in increased user effort.

Test examples (excerpt):

Observations

OpenAI reproduces the input error exactly – the stripes remain crooked, even after explicit correction prompts. Label rendering works better than with Gemini; the model appears more relaxed and naturally positioned. However, a fundamental problem becomes apparent with the mood shots: instead of a re-staging, the studio pose is retained, and only the background is replaced. Additionally, the top of the head is cut off.

Floral sweater

Observations

The floral pattern remains skewed and is not automatically optimized – even an explicit prompt stating "pattern should be straight" changes nothing. Additionally, OpenAI ignores format specifications: instead of being square as desired, the model image is output in portrait format.

The models' proportions appear unnatural (long neck, small head). The café setting also suffers from excessive smoothing – both models produce noticeably AI-like results in this setting.

Shirt with print

Observations

The shirt print is also captured well, but printed crookedly (similar to the input). OpenAI shows more realism in skin rendering – pores and skin texture are more visible than with Gemini.

The critical problem becomes apparent in the mood shots: The first two attempts exactly copy the studio pose, despite requests for a new take. Attempt three finally brings a different pose – but: The piercing switches sides, tattoos change position and design, and the model looks different. For campaigns with multiple products, this is a deal-breaker.

Striped shirt

Observations

The stripes are crooked again (input error is copied). OpenAI also added a white border that wasn't requested in the prompt – when asked, the model denies its existence. The vitiligo patches are inconsistent – ​​the hand pattern changes between images.

What works well:

• Labels and care labels remain more legible than with Gemini.
• Skin texture appears more realistic in most tests (pores, details more visible).

Critical weaknesses:

• Copying input errors 1:1: Crooked products (e.g., on the hanger) remain crooked (no automatic correction)
• Partially ignores format specifications (portrait format instead of square despite specific instructions in the prompt)
• It typically requires 2-3 iterations instead of one to achieve the ideal result.
• Partially unnatural model portrayal (Hit or Miss)
• Mood shots: Often only change the background without restaging the model or pose.
• Models are inconsistent across multiple images : e.g., piercings change side, tattoos change position, slightly different look.

Recommended Use Cases:

Individual product images with a focus on label readability . According to our tests, this is not particularly suitable for series with consistent models . Often, several iterations are necessary for an ideal result.

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

Based on our tests, we recommend Google Gemini (Nano Banana Pro) as the main solution for fashion brands, and OpenAI for specific use cases with a text focus.

Google Gemini as the main solution:

In tests, Google Gemini delivered consistent product renderings across multiple items. Its single-shot capability reduces iteration time – usable results were generated for all examples on the first run. Optimizing the source material corrects faulty inputs : if a garment hangs crookedly in the photo, it will be displayed straight in the generated image.

We especially recommend Google Gemini (Nano Banana Pro) in these cases:

• Seasonal collections with multiple products
• Category header for Collection Pages
• Mood images for atmospheric campaigns (however, case-specific differences: greenhouse and industrial settings yielded satisfactory results, while café settings did not appear authentic)
• Fashion brands with limited budgets for photoshoots
• Campaigns that require consistent models across multiple products (when model images are used as input, the same model is displayed consistently across multiple iterations)

Not suitable: Products with mandatory legible care labels or labels containing legally relevant information.

OpenAI for text-focused products:

In general, our tests in other industries also show that the model processes text better and more reliably than Google Gemini. However, the iteration effort required to achieve a satisfactory output was higher in our tests (an average of 2–3 attempts). Furthermore, the lack of optimization or high fidelity to the original photo necessitates better input quality.

• Brands with a focus on label presentation
• Studio shots where text readability is critical