AI Video and Image Generation in Marketing - Tools, Costs and Copyright in 2026

In March 2026 OpenAI announced it was shutting down Sora. The app stopped working on 26 April, and the API was switched off on 24 September - less than a year after the high-profile launch of Sora 2. According to reports, running the service cost around a million dollars a day, and the user count dropped quickly after its peak. For a company planning AI-assisted production, that's an important lesson: video and image generation tools change faster than marketing campaigns do, so build your process around use cases, not around a single tool brand.

OpenAI shut down Sora less than a year after the Sora 2 launch, and Veo 3.1 now costs $0.05 to $0.40 per second of video. Which AI video and image tools work in 2026, what an accepted asset really costs, who owns the copyright, how to label content since August 2026 and how to build a process with human review.

The second market signal is less technical. Coca-Cola drew a wave of criticism for an AI-generated holiday ad for the second year running, and an IAB study showed the share of consumers negative about AI-made ads is 12 percentage points higher than in 2024. AI lowers the cost of producing assets, but used badly it can lower the perceived value of the brand too. In this post I show which tools work as of September 2026 and what they cost, what they're good for in a small or mid-sized company, where copyright and likeness stand, which labeling obligations have applied since August 2026, and how to build a production process with human review.

Tools and prices in 2026

The market splits into four groups: video generators, image generators, avatar and voiceover platforms, and editing tools. The prices below come from official API price lists and plans as of September 2026. They change often, so check current rates before you commit.

Two notes change the math. First, a list price is the cost of one generation, not of a finished asset - in practice each accepted clip or graphic takes several attempts, so multiply the real cost by 3-5. Second, licence terms depend on the plan: Midjourney requires companies with more than a million dollars in annual revenue to be on the Pro or Mega plan to use outputs commercially. How to keep API costs down at scale I covered in OpenAI API cost optimization - the routing and batch-processing principles apply equally to images and video.

What AI is good for in company marketing

It delivers the best results where the number of variants and speed matter more than the uniqueness of a single shot.

• Ad variants. The same product in five settings, three formats and two seasons - without a photo shoot. You test more creatives for the same budget.
• Backgrounds and product photo editing. Swapping the background, lighting or surroundings of a real product photo is safer than generating the product from scratch, because the product stays true to reality.
• Thumbnails and content graphics. YouTube thumbnails, post graphics and infographics - how such graphics work for visibility I cover in the posts on video in AI visibility strategy and image SEO.
• Short video and cutaways. A few seconds of footage for ads and reels, complementing material shot on camera.
• Training materials and language versions. An avatar with a voiceover lets you quickly produce a tutorial in several languages and update it without re-recording.

There are also places where AI hurts more often than it helps. Don't generate the product the customer will later receive - every difference between the image and reality means a complaint and lost trust. Be careful with image campaigns for brands whose value lies in craft and authenticity - the criticism of AI in ads from Coca-Cola, McDonald's and Gucci shows audiences read synthetic imagery as cutting corners on quality. And never create the likeness of real people without their consent.

Copyright in AI images and video

The key principle is simple: a work protected by copyright has to result from a human's creative effort. In Poland this follows from Article 1(1) of the Copyright Act, which protects "any manifestation of creative activity of an individual nature" - and only a human can be an author. In the US the matter is even clearer: in March 2026 the Supreme Court declined to hear Thaler v. Perlmutter, leaving in place the appeals court ruling that a work created solely by AI isn't protected.

In practice, a graphic generated from a single prompt may belong to no one - a competitor can use it too. The more genuine human contribution there is (composition, your own photos as a base, editing, compositing), the better the chance the final result is protected. Key brand assets such as a logo, a mascot or the visual identity are therefore worth designing with a designer, treating AI as a supporting tool.

A separate issue is training data. In November 2025 a UK court largely dismissed Getty Images' claim against Stability AI, finding that model weights aren't copies of the works. In the US, the case brought by Disney, Universal and Warner Bros. against Midjourney is in discovery. For a company using these tools the takeaway is practical: choose providers with clear commercial terms, and for large campaigns check whether the provider offers protection against third-party claims - Adobe does for Firefly and Google Cloud for selected generative services, among others.

Likeness and voice are a separate risk area. Distributing a person's image requires their permission (in Poland, Article 81 of the Copyright Act), and cloning an employee's or voice actor's voice should rest on written consent that defines the scope and duration of use. A generated avatar resembling a well-known person is a fast route to a dispute - and to the issues I covered in the post on deepfakes and AI fraud.

Labeling AI content since August 2026

Since 2 August 2026, Article 50 of the AI Act applies, splitting obligations between tool providers and tool users.

• Tool providers must mark generated images, video and audio in a machine-readable way - hence watermarks like SynthID in Google's tools and C2PA metadata.
• Companies publishing content must clearly label deepfakes, meaning material that resembles real people, places or events and could be taken as authentic. Where the material is evidently artistic, satirical or fictional, disclosure in a way that doesn't spoil the work is enough.

On 10 June 2026 the European Commission published a code of practice on marking and labeling AI-generated content, including an EU icon for labels. The code is voluntary, but it will become a reference point when compliance is assessed. On top come platform rules: YouTube requires disclosure of realistic content generated or altered with AI, and Meta adds an "AI info" label to ads created or significantly edited with AI tools, including third-party ones, detecting them through industry-standard signals in the files. A practical rule: don't strip C2PA metadata when processing files, and keep a record of how each creative was made.

How to keep the brand consistent

Generators' biggest problem isn't the quality of a single image; it's repeatability. A series of graphics where the product, colors and style shift between shots looks amateurish. A few rules that limit this:

• Reference images instead of descriptions. Newer models, including Nano Banana, accept several photos as a model - product, character, style - and stick to them far better than to a text description alone.
• A fixed style description in every prompt: brand color palette, lighting type, framing, what to avoid.
• A banned list: other companies' logos, public figures, symbols that could be misread, text in languages you can't check.
• Quality control before publishing: hands and faces, text in the image, product proportions and details, the logo, compliance with brand guidelines.

The same rules apply to AI-assisted text - how to keep quality up and avoid the mass-production trap I covered in the posts on AI in content marketing and Google's policy on AI content.

A production process with human review

A proven process has six steps and two human checkpoints.

1. 1.Brief. The asset's goal, channel, format, audience and brand reference images.
2. 2.Generating variants. Several to a dozen versions from the same brief, so there's something to choose from.
3. 3.Selection and edits. A human picks the best variants and refines them in an editor - this is where genuine creative contribution happens.
4. 4.Legal and brand check. Likeness and consents, trademarks, the labeling obligation, compliance with guidelines.
5. 5.Publishing with metadata. Files with provenance information preserved and a note on how they were made.
6. 6.Measurement. Comparing the results of AI creatives with traditional material on the same metrics.

The check in step 4 isn't a formality - it's the only place you'll catch a likeness or trademark problem before a customer sees it. Write it down as a short checklist and attach the result to each campaign's documentation.

A step-by-step implementation plan

1. 1.Pick one use case with a high number of variants, e.g. product photo backgrounds or ad variants.
2. 2.Prepare a brand pack: reference images, style description, palette, banned list.
3. 3.Test two or three tools on the same brief and calculate the cost per accepted asset, not per generation.
4. 4.Check the licence terms of your chosen tool for your company size and type of use.
5. 5.Write a checklist for quality, likeness, trademarks and labeling.
6. 6.Set labeling rules in line with Article 50 of the AI Act and the rules of the platforms you publish on.
7. 7.Collect consents for using likeness and voice if you use avatars or voice cloning.
8. 8.Compare results of AI and traditional creatives on the same campaign metrics.
9. 9.Review the tool market quarterly - Sora's shutdown showed a provider can disappear within months.

---

I design AI-assisted production processes - from choosing tools and building a brand pack, through automating variant generation, to a legal checklist and labeling rules. I do this as part of AI automation and SEO content marketing. Get in touch - I'll start with one use case and calculate what an accepted asset really costs you.

Worth reading next:

• YouTube and Video in AI Visibility Strategy - The Strongest Signal You're Not Using
• Image SEO in the AI Era - Google Lens, Visual Search and Graphics Models Cite
• AI in Content Marketing - How to Produce 10x More Content Without 10x the Budget
• Deepfake Fraud - How Companies Lose Money to Fake CEOs, and How to Defend Against It