AI video for a brand film in Dubai what the tool companies publish about their own limits
A brand gets shown an AI video option for a Dubai brand film and nobody in the room can check what was claimed. This article reports what the companies building these models publish about their own products: clip length and what extension switches off, four different answers to whether a real person can be generated, where each vendor says its own model is weak, what a per-second price leaves out, and who downstream has published what they will refuse. All of it was read from those companies' own documentation on 30 August 2026.
The one line to hold onto
Every capability sentence here names a company, a model, a page and the date it was read, because that is the only form in which any of it stays checkable.
Three published facts that decide an AI video brief in Dubai
A generation is seconds long, and both flagship vendors publish a route past that. Google publishes the Veo 3.1 clip length as 4, 6 or 8 seconds, and an extension whose output combines the input video and the generated extension for up to 148 seconds at 720p. OpenAI publishes that each extension adds up to 20 seconds, six times over, to a 120-second maximum.
What is switched off at that ceiling is the more useful finding. OpenAI publishes that extensions accept only a source video and a prompt, and support neither characters nor image references. Those are the two features that hold a subject looking like itself, documented as unavailable on the route that makes a piece long enough to be a film.
And one vendor ranks misrepresenting the customer's own product as an anticipated failure of one of its own modes. Amazon's AI Service Card for Nova Reel gives the error types for that model's multi-shot mode, ranked by estimated negative impact, as harmful or otherwise inappropriate content, and misrepresenting the input product.
Nothing here says what AI can or cannot do in general, because four companies publish four different answers to that question, and all four documents were open on the same day.
What the specification says, in the vendors' own numbers
Watch out
We have established nothing about how any of these models handles Arabic on screen or in speech, because none of the vendor documents read on that date addresses it, and nothing was tested. What a second language costs and changes on a filmed job is in the Journal's article on Arabic video production.
Two things the table cannot say. Google publishes that generated videos are stored two days, with only a video from that window eligible for extension, and that sound is always on for Veo 3 and 3.1. And the Sora 2 duration differs by whose page you open: OpenAI publishes 16 and 20 seconds, Microsoft 1 to 20 for Azure, unresolved on any page read.
Every vendor document retrieved that day that publishes a frame rate for a generated clip publishes 24 FPS, Runway 24 or 25, and the only higher figure in any of them is Adobe's 30 FPS for extending footage that already exists.
| What a brief asks | Veo 3.1, from Google | Sora 2, from OpenAI | Where the wider set differs |
|---|---|---|---|
| Single generation | 4, 6 or 8 seconds | 16 or 20 seconds | Runway publishes 30 seconds on seedance2_5, 10 on Gen-4.5 |
| Extension ceiling | 148 seconds, at 720p | 120 seconds over six extensions | Kling publishes extension as unavailable on all six of its models |
| Aspect ratios | Two: 9:16 and 16:9 | 1920x1080 or 1080x1920 at 1080p | Runway publishes 21:9 on hailuo3, 24 ratios on seedance2 |
| Frame rate | 24 FPS | Not on OpenAI's video generation developer guide | Runway publishes 24 or 25 on Gen-4.5 |
| Resolutions | 720p, 1080p, 4K, the last two eight-second-only | 1080p requires sora-2-pro | Luma publishes 360p to 1080p on Ray 3.2 |
| Prompt language | English, others not evaluated | English or another Latin script, per Microsoft's Azure page | Amazon publishes English prompts only for Nova Reel, while noting short non-English strings may still render |
Whether a real person can be generated, and why four companies answer it differently
Watch out
The sentence "AI cannot generate a real person" was wrong in all four directions above on the day they were read. The answerable question is which company, which model, which tier, which region, and it is settled in a contract before anyone writes a prompt.
A Dubai brief hits this first, because the subject is often a founder, a chef or a general manager. Four companies answer it from four different architectures.
Google adds a fifth item of a different shape, publishing that in EU, UK, CH and MENA locations the only permitted personGeneration value for Veo 3 and 3.1 is allow_adult. That is Google restricting its own model in a region including the UAE, not a UAE rule.
The one party publishing a consent test a producer can apply is a buyer. Netflix publishes a Digital Replica consent definition to its production partners, with three exceptions: output substantially as scripted or recorded, activities no human could safely perform, and a performer rendered unrecognisable. One company's requirement of its own suppliers, not law.
OpenAI: a block with an eligibility route
Real people, including public figures, cannot be generated. Human-likeness character uploads are blocked by default, input images with faces are rejected, and likeness access runs through an account manager.
Runway: a per-request dial
A contentModeration.publicFigureThreshold parameter, auto by default, with low documented as making the system less strict about preventing generations that include recognisable public figures. The buyer carries the likeness obligation.
Amazon: deterrence, nothing to configure
Safeguards to deter the generation of famous or public figures, in that word, and a published statement that the safety filters cannot be configured or turned off.
Adobe: an enterprise route around a real likeness
A custom-model use case named photoshoot of a person, for photorealistic images of an existing human likeness. Images rather than video, enterprise-gated, and not permission to use a face.
What the vendors publish about their own weaknesses
A production house cannot credibly write this section on its own account. Five companies published it instead.
Amazon publishes that Nova Reel is designed to create novel videos through its own expression, not to imitate any particular existing works. On composition, it publishes that absent clear instruction the model may crop elements, giving a close-up of a wheel or a door handle for the prompt "a red sports car", and may vary camera angles, lighting and object pose. A brand film exists to show a particular existing thing, which is that design intent in reverse. The Journal's article on brand film production covers why its subject is a real, named person.
Runway publishes three named failure modes of its own current flagship: causal reasoning, where effects precede causes; object permanence, where objects disappear or appear across frames; and success bias, where actions disproportionately succeed.
Adobe publishes three known issues for Firefly's Generate Video: human and animal forms distort as they move, unwanted objects are added, and there is unintended subject or object motion.
Google's Veo 3 model card publishes that complete consistency through complex scenes or complex motion remains a challenge, and that during testing for unfair bias Veo 3 appeared to skew towards lighter skin tones when race is not specified in the prompt. That is a governance point before a craft one.
Microsoft publishes, for the Azure deployment of Sora 2, that the model may have difficulty with complex physics, causal relationships such as bite marks on a cookie, spatial reasoning such as left from right, and precise time-based event sequencing. Read against a brief: bite marks is food, left from right is screen direction, sequencing is the grammar of a cut.
What a benchmark can and cannot tell you
Watch out
No benchmark retrieved had tested Veo 3.1, Sora 2, Gen-4.5 or Kling 2.5 or above. We have established no ranking between the models, only what each vendor publishes about its own limits.
Published evaluations measure the systems they tested, and those are not the models a buyer is sold today.
VBench-2.0, a 2025 benchmark, tested HunyuanVideo, CogVideoX-1.5, Sora at 480p and Kling 1.6, and for those four reports dynamic spatial relationship at 19 to 21 per cent and motion order at 15 to 29 per cent. Translated once: screen direction, and the order events happen in.
A 2025 human evaluation of on-screen text across ten named systems current in May 2025 reports that all show noticeable failures with textual content, best average 0.37, alongside Amazon's own line that users should not expect Nova Reel to generate coherent text. A 2025 physics benchmark adds that all models tested score below 0.60 on average in each law category, and that law-specific prompt hints fail to remedy physics violations.
The price, the arithmetic, and the number that inverts it
Watch out
We have established no figure for how many generations a usable shot takes. Eight named vendor documents were read in full and none states a success, acceptance, retry or failure rate.
The published per-second rates are real and low. Google prices Veo 3.1 Standard with audio at USD 0.40 a second at 720p and 1080p and 0.60 at 4K, with Fast and Lite below that. OpenAI prices sora-2 at USD 0.10 a second and sora-2-pro at 0.30, 0.50 and 0.70 by resolution. Runway sells credits at USD 0.01 each, and its reseller rate for veo3.1 with audio comes to USD 0.40 a second, the figure Google publishes itself.
Then the number that inverts the picture, off one page. Adobe publishes Generate Video at 100 generative credits per second at 1080p24, publishes monthly allocations of 2,000, 4,000, 10,000 and 50,000 credits across its four Firefly plans, and publishes that unused credits expire each month and cannot be carried forward.
That row is arithmetic on two published figures from one page, not a market finding. It is also the number a marketing team meets, because a team buys a subscription rather than an API call, and a rate that reads as almost free becomes minutes a month at 1080p.
None of those published rates covers the brief, the script, the edit, the sound, the grade, the versioning, the approvals, or the generations that were not kept. The Journal's guide to video production cost in Dubai carries the cost drivers. Refused generations are not free everywhere either: Google charges only for a successful generation, Luma refunds failed ones, and Runway publishes that a moderated generation costs the same as a successful one and that an account making too many will be suspended. On what trips a filter, Luma publishes that named public figures and reference images depicting real people are among the common causes of a false positive.
| Monthly credit allocation | 2,000 | 4,000 | 10,000 | 50,000 |
|---|---|---|---|---|
| Seconds of 1080p24 video a month | 20 | 40 | 100 | 500 |
Who has to accept the film, and what they publish they will refuse
Four parties have published what they require, and not one of the four is a UAE rule.
Netflix, as a buyer, publishes to its production partners that where output includes talent likeness, personal data or third-party IP, written approval is required before proceeding, and that one of its five principles is that the tools used do not store, reuse or train on production data. Hold that against Runway's Terms of Use, which grant Runway a perpetual, worldwide, sublicensable licence to use inputs and outputs for training, in the same clause that says Runway claims no ownership and does not restrict commercial use. The product photographs, the unreleased packaging, the founder's portrait: on those consumer terms, uploading them sits inside that grant. Whether an enterprise agreement changes it was not established.
In the United Kingdom a regulator has answered a version of this. The ASA has stated that even though an ad was produced using Google's automated tools, the advertiser was still primarily responsible for compliance with the CAP Code. That is the UK's CAP Code, not evidence about the UAE.
On indemnity, AWS publishes a defence obligation naming Amazon Nova Reel, then expressly excludes any claim that output infringes a third party's trademark, which is close to the definition of a brand film. Google lists Veo among its indemnified services. Microsoft's Customer Copyright Commitment requires a customer tendering a claim to demonstrate compliance, including a retained testing and evaluation report from guided red teaming or systematic measurement. Adobe's indemnity names its text-to-video and image-to-video features among the eligible ones and excludes capabilities the interface denotes as powered by non-Adobe trained models. Adobe publishes separately that Luma Ray3, Runway Gen-4.5 and Veo run inside Firefly's own Generate video feature, each with its own settings panel.
YouTube, the only platform document retrieved, requires disclosure where realistic content has been altered or synthetically generated, labels photorealistic AI content in the player, and may apply that label automatically to content carrying C2PA metadata. Four vendors publish that they write that provenance in, and Adobe publishes that a user must not remove, alter or disable Content Credentials.
Every indemnity we read runs to the model provider's own direct customer, and we have established nothing about whether any of it reaches a brand at the end of an agency chain. The Journal's article on usage rights and ownership covers what a client owns on a normal production.
The question is not whether it looks real. It is whether the brand that owns the product will sign it off.
Where generation already earns its place
The vendors built for the brand job by name.
Runway's API exposes fifteen video-generation model identifiers and only four of them are Runway's own, the rest third-party models resold under Runway's billing, so a supplier saying it works in Runway has not yet said which model made the film. Runway publishes seven Recipes endpoints aimed at exactly this work, including product_ad, product_swap, product_ugc and ad_localization. Its product-ad endpoint takes up to ten images of the same product, uses only the first as the primary reference, and returns four to fifteen seconds. Runway also publishes where its own product-swap breaks: it works best like for like and struggles when the reference video shows an action the new product would not perform.
OpenAI describes sora-2 as suited to rapid iteration, concepting and rough cuts, and directs anyone needing a 1080p export to sora-2-pro. That is the line drawn by the company selling it.
Generative video passes one delivery test on a vendor's published account. Runway publishes professional output formats on Gen-4.5 and Aleph 2.0: ProRes to 4444 XQ, PNG sequences, 10-bit SDR Rec.709 masters, HDR10, HLG and OpenEXR. The container and colour half of a delivery specification is a real gate, and one vendor publishes that it meets it.
The published prompt ceilings draw a boundary of their own. Amazon publishes 512 characters for single-shot generation, Adobe 1,800 for Firefly video, Kling 3,072. Those are limits on the instruction, not on the brand guideline it has to carry. The Journal's article on CGI against live action covers the render-versus-proof argument.
The Dubai layer, and the four things this research did not establish
A brand film is made to show real, named things. A tower with a name, a lobby a guest will walk into, a founder whose face is on the licence. Beyond Google's MENA personGeneration restriction, a vendor capability limit rather than a rule, four gaps are worth stating plainly. The first is availability: we have established nothing about whether any of these tools is available in the UAE.
We have established nothing about how the Dubai authorities treat a film that was generated rather than shot. Everything regulatory sits in the Journal's article on AI-generated video rules in the UAE, and the permission a brand advertisement needs before it runs is in the article on the UAE advertiser permit.
On real places, five vendor policy documents were read in full and none names a building, a landmark or a real place, each framing the question as third-party rights in general. What was established is a silence in those documents on that date, neither permission nor prohibition.
On price, SL Media publishes AED 1,500 to 3,000 for a 30-second listing reel from existing photos, 5,000 to 8,000 for a walkthrough with AI staging and voiceover, and 12,000 to 20,000 for AI avatar or narration with one real interior shoot, on 29 July 2026. Motion Labs publishes monthly plans at USD 3,099 and USD 5,099, on 29 August 2026. Two suppliers' own rate cards, and averaging them would manufacture a figure nobody published. We have established no market rate for an AI-made brand film here, and no figure for how many brands here use generative video.
Common questions
Can AI generate our founder on camera?
Four companies answer differently. OpenAI publishes a flat block with an eligibility route through an account manager. Runway ships a per-request dial. Amazon publishes safeguards that deter and filters that cannot be reconfigured. Adobe sells an enterprise likeness route for images. A contract question, settled before the prompt.
How long a film can these tools actually make?
Google publishes an extension route reaching 148 seconds and OpenAI 120, so the length is there. What is switched off at the ceiling is the point: OpenAI publishes that extensions support neither characters nor image references, the features that hold a subject looking like itself.
We were quoted a per-second rate. Is that the real cost?
It is a real published rate: Google USD 0.40 a second for Veo 3.1 Standard with audio at 1080p, OpenAI USD 0.10 for sora-2. Divide Adobe's monthly credit allocations by its 1080p24 rate, arithmetic on one page, and the four plans buy 20, 40, 100 and 500 seconds a month. None of those rates covers the brief, edit, grade or approvals.
Will YouTube label the film?
YouTube requires disclosure where realistic content has been altered or synthetically generated, labels photorealistic AI content in the player, and may apply that label automatically to content carrying C2PA metadata the generator wrote. Scoped to YouTube, the only platform document retrieved.
If the model produces something that infringes, who carries it?
Every indemnity read runs from the model provider to its own direct customer, and AWS expressly excludes claims that output infringes a third party's trademark. Whether any reaches a brand at the end of an agency chain was not established. Ask for the terms of the account the film was made on.
Our agency says they use Runway. Is that enough to know?
No. Runway's API exposes fifteen video-generation model identifiers and only four are Runway's own, the rest third-party models resold under Runway's billing. Adobe publishes that three other companies' models run inside Firefly, and publishes separately that its indemnity excludes capabilities the interface denotes as powered by non-Adobe trained models. Ask which model, on which surface, on what date.
What to hold onto, in one pass
A generation is seconds long, both flagship vendors publish an extension route past two minutes, and the features that hold a product or a person consistent are documented as unavailable on that route. Whether a real person can be generated has four published answers, by company, tier and region. One vendor ranks misrepresenting your product as an anticipated failure of its multi-shot mode. The per-second price and the monthly credit allowance tell opposite stories off the same page.
All of it has a shelf life the vendors publish themselves. Kling publishes that nine named models retire on 15 September 2026. Google marks Veo 3.1 generally available on one surface and Preview on another, retiring 17 November 2026 or later. So the durable thing is not the numbers, it is what to ask: which model made the film, on which surface, on what date.
Grolez Films: the questions that get asked before the quote
Before we quote a job, we work out who has to give permission and who has to apply for it. That is the same question this article has put to the vendors and the platforms, asked of a live-action shoot instead.
Luis Grolez, the founder, pitches and reviews every job personally. Post-production, the edit and the grade, is about a third of a job. The client owns the finished film, and raw files are supplied on request.
See the work. Then get in touch about a specific brief and a specific quarter.
Sources and methodology
Everything here about what a model does was read from the company that builds it, on that company's own documentation, on 30 August 2026, and every figure carries that date because the vendors themselves publish retirement dates measured in months. No production-company blog, agency listicle, tool-comparison page or AI newsletter is cited as authority for how a model behaves, and every one returned by ordinary searching was set aside. The two Dubai suppliers below are cited only for the prices they publish about themselves. The retrieval ladder was worked per host and in both directions, and four hosts that looked unreadable to one route were readable by another on the same day, which is why nothing here is written up as a company publishing nothing. OpenAI's general usage policies, Kling's commercial-use terms, SAG-AFTRA's material, Midjourney's documentation and TikTok's and Meta's labelling pages were not retrieved at any rung attempted, so no claim is made about any of them in either direction. No benchmark located had tested the current flagship models, so no benchmark number is attached to one. Two circulating market figures were traced to a single cited page containing neither and are excluded as facts. The two named Dubai suppliers' AED and USD tiers are what those suppliers publish about themselves; they are not averaged and not generalised. No effective date is quoted from Adobe's product-description page, because two independent readings of it on the same day returned conflicting dates, and no update date is quoted for Microsoft's Customer Copyright Commitment because two of its URLs carry two different ones. The ASA material is United Kingdom and the Copyright Office material is United States, and neither is evidence about the UAE, which the Journal's article on AI-generated video rules covers. This article is informational and is not legal advice. Confirm your position with the relevant vendor, platform or counsel before relying on it. Prepared 30 August 2026.
- Google, Veo 3.1 model documentation, for clip lengths, aspect ratios, frame rate, resolutions, prompt language, sound, Content Credentials and the GA and retirement dates: docs.cloud.google.com
- Google, Gemini API Veo documentation, for the Limitations section including the MENA
personGenerationrestriction, the 141-to-148-second extension route, two-day retention and reference images: ai.google.dev - Google, Gemini API pricing, for the per-second Veo rates and the note that only successful generations are charged: ai.google.dev
- Google DeepMind, Veo 3 model card, published 23 May 2025 and updated 13 January 2026, for the consistency limitation and the skin-tone and semantic bias findings: storage.googleapis.com
- Google Cloud, Generative AI Indemnified Services, last modified 20 July 2026, for Veo's presence on that list: cloud.google.com
- Google, Generative AI Prohibited Use Policy, last modified 17 December 2024: policies.google.com
- YouTube Help, disclosing altered or synthetic content, for the triggers, the exemptions, where the label appears and the automatic C2PA route: support.google.com
- OpenAI, video generation developer guide, for the guardrails block, the character feature, image references conditioning the first frame, and extensions supporting neither characters nor image references: developers.openai.com
- OpenAI, sora-2 model page, for the per-second pricing and the exploration-versus-production framing: developers.openai.com
- OpenAI, Sora 2 System Card, 30 September 2025, for C2PA on all assets and the visible moving watermark on downloads: cdn.openai.com
- Microsoft, video generation with Sora 2 on Azure, for the 1-to-20-second range, the content quality limitations and the Latin-script prompting guidance: learn.microsoft.com
- Microsoft, Customer Copyright Commitment required mitigations, for the defence obligation and the retained testing and evaluation report. Two URLs carry two different update dates and no single date is published here: learn.microsoft.com and learn.microsoft.com
- Amazon Web Services, AI Service Card for Amazon Nova Reel, current as of 16 July 2025, for the ranked error types, the novelty-by-design statement, composition and camera-angle variation, on-screen text, languages, public figures and the prompt ceilings: docs.aws.amazon.com
- Amazon Web Services, Nova user guide, for the model specification and regional availability: docs.aws.amazon.com
- Amazon Web Services, Service Terms, last updated 20 August 2026, for the indemnity naming Nova Reel and the trademark exclusion: aws.amazon.com
- Runway, models guide and OpenAPI specification, for the fifteen model identifiers, the per-model duration and ratio enums, the moderation dial and the professional delivery formats: docs.dev.runwayml.com and docs.dev.runwayml.com
- Runway, pricing guide and moderation documentation, for the per-second credit costs, the USD 0.01 credit and moderated generations costing the same as successful ones: docs.dev.runwayml.com and docs.dev.runwayml.com
- Runway, recipes documentation and API reference, for the seven brand-job endpoints, the product-ad inputs, the published product-swap limitation and the likeness obligation: docs.dev.runwayml.com and docs.dev.runwayml.com
- Runway, Gen-4.5 research page, 1 December 2025, for causal reasoning, object permanence and success bias: runway.com
- Runway, Terms of Use, last updated 11 May 2026, for the training licence in section 4.4 alongside the no-ownership and commercial-use statements: runway.com
- Adobe, Firefly product description, for the indemnity conditions, the eligible features and the non-Adobe-model exclusion. No effective date is quoted from this page, per the methodology note: helpx.adobe.com
- Adobe, generative AI user guidelines, last updated 15 May 2026, for the prohibition on removing, altering or disabling Content Credentials: adobe.com
- Adobe, Generate Video help page, last updated 4 September 2025, for the Known issues block and the 1,800-character prompt ceiling: helpx.adobe.com
- Adobe, generate videos using non-Adobe models, last updated 7 August 2026, for Luma Ray3, Runway Gen-4.5 and Veo 2, 3.1 and 3.1 Fast inside Firefly: helpx.adobe.com
- Adobe, generative credits FAQ, page stamped 20 August 2026, for 100 credits per second at 1080p24, the four monthly allocations and the no-rollover statement: helpx.adobe.com
- Adobe, custom models overview, last updated 11 December 2025, for the enterprise photoshoot-of-a-person use case: helpx.adobe.com
- Luma, Ray 3.2 generation guide and pricing, for the duration enum, the frame grid, the resolutions and the per-generation prices: docs.agents.lumalabs.ai
- Luma, FAQ and billing documentation, for moderated and failed generations being refunded and for named public figures and reference images depicting real people among the common causes of a false positive: docs.agents.lumalabs.ai
- Kling, capability map and base pricing, read in a browser, for video extension unavailable across its six current video models, the per-second prices, the 3,072-character prompt ceiling and the 15 September 2026 retirement notice: kling.ai
- Netflix Partner Help Center, using generative AI in content production, no date printed on the page and attributed to the retrieval date, for the written-approval trigger, the five guiding principles and the Digital Replica consent definition: partnerhelp.netflixstudios.com
- ASA and CAP, United Kingdom, 11 June 2026, for advertiser responsibility for automated output: asa.org.uk
- United States Copyright Office, Copyright and Artificial Intelligence Part 2, 29 January 2025, United States jurisdiction only: copyright.gov
- Content Credentials and the C2PA specification 2.2, for the governance statement and the provenance definitions: contentcredentials.org
- VBench-2.0, arXiv:2503.21755, tested on HunyuanVideo, CogVideoX-1.5, Sora at 480p and Kling 1.6. The two dimension scores quoted are read from the full text rather than the abstract, which carries neither the model names nor the figures: arxiv.org
- T2VTextBench, arXiv:2505.04946, ten named systems as current in May 2025, for the on-screen text findings: arxiv.org
- T2VPhysBench, arXiv:2505.00337, for all models tested scoring below 0.60 and prompt hints failing to remedy physics violations. No model is named in that abstract: arxiv.org
- SL Media, SkyLight Studio LLC, Dubai, article dated 29 July 2026, primary for that supplier's own published AED tiers and nothing else: slmedia.ae
- Motion Labs, Dubai, published 29 August 2026, primary for that supplier's own published subscription prices and nothing else: motionlabs.agency
- The citation chain behind the two circulating market figures, listed as the trail rather than as evidence: lumalabs.ai and sellerscommerce.com