Image models in 2026: which to use for brand work

What changed between 2025 and 2026

Image models stopped being one-shot generators and became studio tools.

• Multiple references. Nano Banana 2 accepts up to 14 reference images (10 object, 4 character and 3 style); Nano Banana Pro takes 6 object and 5 character references. That is where product and character consistency across assets comes from.
• Instruction-based editing. You start from an approved image and describe the change in plain language. OpenAI positions GPT Image 2.5 Sunburst for precise editing and Flare for fast generation.
• Higher resolution. The Nano Banana family reaches 4K, and GPT Image 2.5 accepts up to 3,840 pixels per edge, with aspect ratios from 1:3 to 3:1.
• Text inside the image. Google says its models write legible text for menus, infographics and marketing; OpenAI warns that precise text placement can still fail.
• Retirements. On the Gemini API, Imagen 4 shut down on August 17, 2026 and the original Nano Banana shuts down on October 2. OpenAI removed DALL-E 2 and DALL-E 3 from its API on May 12, 2026 and will remove gpt-image-1-mini and gpt-image-1.5 on December 1.

The verified landscape as of September 2026

Only models whose versions, terms and prices we could confirm on primary sources.

Midjourney, Ideogram, Recraft, Seedream and Stability AI's image models are not in the table: we could not confirm their current versions, prices and terms on primary sources. If you use them, run them through the same test.

Which one for which asset

• Finals with product and fine retouching: Nano Banana Pro or GPT Image 2.5 Sunburst.
• Volume and format adaptations: Nano Banana 2.
• Exploration and live experiences that don't preserve a face: Nano Banana 2 Lite or GPT Image 2.5 Flare.
• Data that cannot leave your infrastructure: self-hosted Qwen-Image.

Brand-safe use: licenses, training data, logos and faces

Commercial terms and authorship. First check whether you can use the output in a paid deliverable; with open models, the license decides. Stability AI, for example, requires an Enterprise license above $1 million in annual revenue. And according to the US Copyright Office (January 2025), a prompt alone does not make an image protectable: document the human input in selection, arrangement and retouching.

Training data. Few vendors detail it. Ask for their copyright policy and the public summary the AI Act requires from general-purpose models; Google, OpenAI, Black Forest Labs and Bria AI signed the EU code for those models, which includes a copyright chapter. The route that closes the question is a model trained only on your own assets, such as Adobe Firefly Foundry. We cover this in rights and provenance of AI content.

Trademarks and logos. In Getty Images v Stability AI (High Court of England and Wales, November 4, 2025), Getty lost on copyright, but the judge found trademark infringements, historic and extremely limited, because older model versions generated iStock and Getty watermarks. A model can reproduce someone else's sign and draws yours badly: logos are composited in post.

Faces. A recognizable person requires written consent for that use. In Colombia, their photo is personal data that requires authorization (Law 1581 of 2012), and Law 2502 of 2025 aggravates impersonation done with AI; in France, law 2023-451 requires the mention Images virtuelles on influencer advertising with generated faces. L'Oréal, for its part, states that it does not use AI-generated life-like faces, hair or skin to enhance product benefits.

A production workflow in seven steps

The one we recommend at Slash for brand assets, from brief to delivery.

1. Brief: formats, markets, languages, what must be real (product, logo, legal lines) and where it will run, because that defines the labeling.
2. References: official packshots, product angles, palette, typefaces and signed consents if a real person appears.
3. Generation: cheap low-resolution exploration, then three to five directions chosen with the art director.
4. Editing: instruction-based fixes on the chosen image; logo, price and legal lines composited in post.
5. Human QA: product fidelity, hands, faces, text and accents, third-party marks; legal review when people or product claims appear.
6. Provenance and labeling: keep the vendor's C2PA metadata and watermark, log model, date, prompts and references, and agree the visible label with the client.
7. Delivery: masters, per-format versions and a provenance sheet archived with the project.

Vendor marks help, but they are not enough. Every Nano Banana image carries SynthID and OpenAI has been certified as a C2PA generator since January 2026, but NIST notes that metadata is often stripped when files are shared on social media, and a paper accepted at USENIX Security 2026 reports compromising SynthID-Image. Your internal log is the evidence that depends on no one.

Cost and speed: an illustrative calculation

Generation is cheap; human time weighs more and, live, so does latency.

Illustrative assumptions, not measurements from a project: a campaign of 20 final visuals (30 drafts, 10 edits and 2 final renders per visual) and an event photobooth with 3,000 participants and 2 images each.

A campaign's API bill is measured in tens of dollars; the real cost is the hours of art direction, retouching and review. Matching the model to the stage cuts that bill by more than three times. In live experiences, latency and retries matter more: OpenAI warns that a complex prompt can take up to 2 minutes, and Lite can't preserve the participant's face because it takes no character references. Measure timings in real conditions. More levers in the real cost of AI.

Limits and what to test yourself

No public ranking measures your product in your language. Before committing to a model, test it on your real references:

• Accents and text: ñ, accented vowels, ¿ and ¡, é, è, ç and œ, in short headlines set in the brand typeface. Count errors per image.
• Small logos: on a label or on a billboard in the background; if they get distorted, they go in post.
• Hands: product in hand, fingers, rings and watches.
• Product fidelity: shape, proportions, color and packaging text, compared with the official packshot at 100%.
• Consistency: the same character or product across ten scenes with the same references.
• Faces and skin tones: variety without bias or caricature; in a photobooth, a real likeness of the person.

It is the rule we apply to text models at Slash AI Lab: a fixed battery per client (30 prompts, in Spanish, English and French) run on five engines before every deployment. For images we recommend adding the client's references and a product-fidelity rubric. The method is in how we choose a model.

AI Act Article 50: what changes since August 2, 2026

Duties for whoever provides the model and whoever publishes the image.

Article 50 splits the work. Providers must mark generated images in a machine-readable format (paragraph 2); systems placed on the market before August 2, 2026 have until December 2, under the AI Omnibus. Deployers, meaning the brand or agency that publishes, must disclose deepfakes (paragraph 4) clearly and at the latest at first exposure (paragraph 5).

Mind the definition: a deepfake is content that resembles existing persons, objects, places, entities or events and would falsely appear authentic (Article 3, point 60). A photorealistic packshot of your real product on a real street can qualify. Evidently artistic or fictional works get a lighter regime, not an exemption.

The Code of Practice published on June 10, 2026 spells out the how: two machine-readable layers for providers (signed metadata plus an imperceptible watermark) and, for deployers, an icon visible at first exposure without user action, offline included. On September 24 the Commission published optional icons; the duty is not optional. About 190 organizations had signed the code by the end of July, including Bulgari, Iberdrola and Lufthansa.

For agencies in Colombia. The AI Act reaches providers and deployers outside the EU when the output is used in the Union (Article 2). If you produce for a European client, set out by contract who marks, who labels and who keeps the evidence. Fines reach €15 million or 3% of worldwide turnover, with lower caps for SMEs. More in our 2026 regulatory map.