Combining real footage with AI generated material the decisions that have to be taken before the crew wraps

The reflex is that generated material extends what a shoot can do. Read the published documentation and the constraint runs the other way. The shoot has to meet the model's input requirements in resolution, ratio, duration, frame rate, bit depth, codec and file size, and not one of those is negotiable once the crew has gone home. Each of those requirements is cited below against the company that publishes it.

There is a boundary worth drawing first, because it decides which rulebook applies. A generated clip that stands on its own between two filmed shots is a cutaway, and what the platforms, the libraries and the vendors condition on that object is covered in the Journal's article on AI generated b-roll. This article is about the other case: a generated element that has to sit inside a filmed frame, where deleting it leaves a hole rather than a cut. The test is that simple. If the material either side would still join, it is a cutaway. If it would not, it is this.

Property has its own version of the same choice, between a built picture and a filmed one, and the Journal argues it separately in CGI against live action.

Two vendors closed that gap from their own side, and the conditions are the part worth reading.

Luma publishes for Ray3.2, on its product page and its developer documentation read on 6 September 2026, that its EXR export delivers 16-bit EXR frames for colour grading, compositing and visual effects, encoded in ACES2065-1 on the AP0 primaries. ACES2065-1 is the exact encoding the Academy's own glossary defines for the exchange of image data that have not been colour rendered.

The conditionality chain matters more than the headline. Luma publishes that the EXR export requires HDR to be enabled, that HDR requires 720p or 1080p and is rejected at 360p and 540p, that HDR is not supported with a ten-second duration or with looping, and that extending a prior generation is standard dynamic range only with both HDR and EXR export rejected. Stated as mechanics and nothing more: on that published route the conform-grade delivery exists on a five-second, non-looping, freshly generated clip. Extending an existing generation, which is one of the four routes described below, sits outside that window.

Runway publishes two scene-referred ACEScg EXR deliveries, one for ACES 1.3 pipelines and one for ACES 2.0, in its developer documentation retrieved on 6 September 2026. It is the only vendor read that also publishes what to do with them: the delivered picture is inverted through the ACES 1.3 Output Transform at Rec.2100 PQ and 1000 nits, so reading the frames with the stock ACEScg input transform reproduces the delivered picture with no Read-node changes needed, and a colorimetry sidecar names the ACES version and the inverted component. It publishes a 12-bit 4:4:4 BT.2020 and PQ lossless master alongside them, and a half-float OpenEXR sequence described as linear BT.2020 light where 1.0 equals 100 nits. It also publishes that these formats are being enabled progressively per account, which travels with the finding.

Notice what nobody did. No new component was defined for generated material. Two companies made their output arrive in a form that a component which already exists can read. What the vendors publish about their output limits more broadly is in AI video for a brand film.

The rest of the field publishes something else. Google publishes for Veo 3.1, on its Vertex AI model page, video lengths of 4, 6 or 8 seconds, up to four outputs per prompt, aspect ratios of 9:16 and 16:9, resolutions of 720p, 1080p and 4K, 24 FPS and a MIME type of video/mp4, and a case-insensitive search of that page on 6 September 2026 returns zero occurrences of colour space, Rec.709, bit depth, HDR, ProRes, EXR and gamut. MiniMax publishes a table it titles Output Specs naming resolution, duration and aspect ratio and nothing else. Kling's VIDEO 3.0 model user guide, dated 6 February 2026, publishes two resolutions and a duration of three to fifteen seconds, and no frame rate, codec, container, bit depth or colour space. Moonvalley publishes that its output is production-grade, that every frame holds up in studio workflows, and that outputs are ready for final delivery, and its page publishes no resolution, no frame rate, no bit depth, no colour space, no codec and no container.

Take that last one as a pairing rather than an accusation. Production-grade is not a specification. Ask any supplier what the frame rate, the bit depth and the colour space of the delivered file are.

This is the call-sheet section, and it is the clearest published statement of what a shoot has to produce.

OpenAI publishes, in its video generation guide read on 6 September 2026, that an input image acts as the first frame of the video, that it is useful where the output has to preserve the look of a brand asset, a character or a specific environment, that the image must match the target video's resolution, and that the accepted formats are JPEG, PNG and WebP.

For anything that has to reappear across shots it publishes more. A character upload works best as a short two to four second clip, in 16:9 or 9:16, at 720p to 1080p. Character source videos work best when they match the aspect ratio of the requested output, and where the ratios differ the character can appear stretched or distorted. A single video can include up to two characters. And the character name has to be mentioned verbatim in the prompt, because passing the identifier alone is not enough to preserve the character reliably.

The other two routes in are shorter. Luma publishes that a still is passed as a keyframe, one for a start frame and one for an end frame, and that both may be supplied. Google's console route takes a Start image for the first frame and, optionally, an End image for the last.

Three consequences follow without a word of inference. The unit is seconds rather than takes, and two to four seconds of a subject, clean, in the delivery ratio, is a specified deliverable that looks like nothing else on a shot list. There is a published ceiling on how many things can persist, so a brief that needs the product, the founder and the packaging all to travel has met a published limit before it has met a creative one. And the end frame is as much a deliverable as the start frame, because where a generation has to land back on filmed material, the frame it lands on has to exist before the generation is ordered. What a planned day yields on the filmed side is covered in one shoot day and the content that comes off it.

The plumbing decides a timeline more than the craft does.

That 30 FPS input ceiling has a real consequence on food, product and action work. High-frame-rate material shot for slow motion sits outside the published input specification of that route.

Runway publishes input size ceilings of 32MB for a video supplied by URL, 16MB as a base64 data URI and 200MB by ephemeral upload. It separately publishes an accepted codec list in which the QuickTime route takes Apple ProRes from 422 Proxy to 4444 XQ. The professional codec is accepted and the file-size ceiling is what bites, so what reaches the model is a compressed version of the master rather than the master. Nothing read measures what that hand-off does to a picture, so nothing is claimed about it here. It also publishes that the file extension is not considered by the API and that the Content-Type header has to match the asset, with generic values explicitly unsupported.

The return leg has its own clock. Runway publishes that the URLs its API returns are ephemeral, will expire within 24 to 48 hours of accessing the API, and that it expects the customer to download the data and save it to their own storage. Luma publishes that presigned output URLs expire after one hour, that the buyer should download promptly or re-poll to mint a fresh one, and that the generation identifier itself does not expire. On a filmed job the material sits on a card and then on a drive whether or not anybody remembered to fetch it.

Every money question in this subject belongs to AI video against a shoot day, and no figure for it appears here.

Everything in this section binds a named body in a named territory, and none of it binds anybody in the UAE.

EBU R 103, Video Signal Tolerance in Digital Television Systems, Version 3.0, May 2020, publishes its tolerance as a table of code values at four bit depths, describes signals outside the preferred range as having a gamut error, and states that measuring equipment should indicate an out-of-gamut occurrence only after the error exceeds one per cent of the image. The clause that reaches a graded mixed cut is short: for pre-produced and colour graded material, the nominal limits should be followed closely.

It also advises against fixing element by element. Colour gamut legalisers should be used with caution, the EBU writes, because they may create artefacts more disturbing than the errors they correct, and it is advisable not to legalise video signals before all signal processing has been carried out. A mixed timeline is signal processing, so the body's own advice puts the fix at the end of it.

The clearest published proportion sits in the AS-11 delivery specification agreed by the Digital Production Partnership broadcasters, named on its first page as the BBC, BT Sport, Channel 4, Channel 5, ITV, Sky, STV and TG4. It limits material of a lower resolution than the commissioned format to 25 per cent of the programme's total duration and requires that such material not be used for large uninterrupted sections unless the broadcaster agrees. It defines what counts by file property rather than by provenance, naming codecs with bit rates below those specified for HD Tier 2L and material generated or processed on 720-line equipment.

The same specification carries the only published number located anywhere for a composited element that fails to cover the plate. It permits a two-pixel tolerance during CG or complex overlay sequences where key signals, graphic overlays or other effects do not fully cover the background image, and requires that a note of the timecodes and reasons accompany the delivered programme. Two facts in one clause: a quantified tolerance, and a paperwork step the composite itself creates.

Its picture requirements are written as plain conditions. The picture must be free of excessive noise, grain and digital compression artefacts. Movement must appear reasonably smooth and continuous, without break-up or large changes in resolution. The picture must be stable and continuous. There must be no visible contouring or artefacts caused by digital processing, with quantisation noise not apparent.

Then the counterweight, which belongs in the same breath. The same section states that this does not mean low quality material cannot be used, that archive and specialist low quality material used in context is acceptable, and that where there is doubt the broadcaster should be asked. It is a proportion-and-context rule rather than a purity rule. The contractual edge is real all the same: a programme failing quality control is rejected and returned to the supplier for repair.

The buyer-side version names R 103 by number. Clearcast's library upload specification publishes a quality control stage requiring colour space ITU-BT.709, RGB gamut to EBU R103, a photosensitive epilepsy check against ITU-R BT.1702 and integrated loudness to EBU R128, and its accepted commercial formats are all 1920 by 1080, 16:9, 25 frames per second and interlaced. That document's own header reads version 1.5 while its file path reads version 1.6, which is recorded here and not resolved.

Where a non-camera element sits inside an HDR picture was settled long before any of this. Report ITU-R BT.2408-9, March 2026, publishes that SDR graphics should be directly mapped into the HDR signal at the Graphics White signal level of 75 per cent HLG or 58 per cent PQ to avoid them appearing too bright, and defines Graphics White as the equivalent in the graphics domain of a 100 per cent reflectance white card, carrying the same signal level as HDR Reference White at a nominal 203 candelas per square metre. Note the construction. The element that did not come from a camera is anchored to the camera's world, not the other way round.

Then the declaration regime. AMWA's AS-11 UK DPP HD specification requires MXF labelled OP1a with 10 bits per sample in every plane of the decoded picture, and a fixed set of mandatory declarations including a photosensitive epilepsy pass code with values of Yes, No and Not tested. No field in that framework names AI, generation or provenance of any kind. A delivery regime does not ask what a shot is made of. It asks the supplier to sign for what the file is.

Grain closes the section, and it belongs here as image structure rather than as a look. The Alliance for Open Media publishes, in its film grain synthesis technical report of 8 July 2022, that grain is difficult to carry because it is random in the spatial domain, temporally independent, and changes in strength with signal intensity, so preserving it can require prohibitively high bitrates. It publishes the chain the delivery pipeline uses instead: denoise the input, encode the denoised video, analyse the difference to estimate grain parameters, send those parameters alongside, and rebuild the grain at the decoder. The AV1 specification then defines those parameters as numbers rather than as an image, including a grain seed, a scaling function limited to fourteen points or fewer, and separate autoregressive coefficient sets for the two chroma components. Read that against a QC list requiring a picture free of excessive noise and grain and free of visible contouring, and the two together explain why a shot that looks right on a laptop can arrive at the far end of a chain looking as though it came from somewhere else.

Six providers were read on 6 September 2026, and the answers do not line up. What follows is what each publishes, with no ranking and no advice.

Google publishes that its indemnification obligations apply to allegations that an unmodified generated output from a generative AI indemnified service infringes a third party's intellectual property rights. Its own definition limits an indemnified service to one that is not provided free of charge. Across the whole of that document, 176,858 characters, the string unmodified occurs exactly once, in that clause, and combining or compositing generated output with filmed material is nowhere addressed.

Amazon Web Services conditions its modification exclusion on something different. It applies where the customer has fine-tuned, refined, customised or otherwise modified an indemnified generative AI service and the alleged infringement would not have occurred but for that modification. The object is the service, not the output, and the strings unmodified and unaltered do not appear anywhere in that page as retrieved. It also conditions the defence on the customer retaining and providing sufficient records, which is a real record-keeping obligation on a production.

Microsoft's condition is knowledge-qualified rather than a modification rule. The customer must not modify, use or distribute the output content in a manner it knows, or should know, is likely to infringe or misappropriate a third party's proprietary right.

Adobe's current product description, published 27 August 2026, carries no modification exclusion at all, and expressly excludes the placement of an output on the canvas before the project is exported from the covered event. Placing a generated element into a project is precisely the mixed act. Note that an earlier Adobe enterprise legal document of 12 June 2023 does carry a modification exclusion, so two Adobe documents diverge, and which governs a given customer depends on an agreement that is not published.

Runway's public terms of use, last updated 11 May 2026, contain no clause in which Runway defends the customer against a third-party intellectual property claim, and the page points enterprise customers elsewhere, so it establishes nothing about enterprise terms. Luma's terms of 14 May 2026 run the other way: the customer defends claims arising out of the input or the output, Luma's total liability is capped at the greater of fifty US dollars or six months of fees, and it states plainly that output or portions of it may be owned or controlled by a third party.

None of the six defines what unmodified means for a conformed picture, so nothing is established here about where a mixed conform sits in any of them. What a client owns in the finished film is a separate question and is covered in usage rights and ownership.

We have established no figure for how many attempts a matched shot takes, and no vendor read for this article publishes one. We have established no published measurement of whether a generated clip matches filmed material inside one sequence, and the benchmark position is recorded in AI generated b-roll rather than restated here. We located nothing published that says how many brands in this market run filmed and generated material in one cut. Nothing we read measures any generated output against the delivery specifications set out above, so we have established nothing about whether it meets them or fails them. None of the six provider documents we read defines what unmodified means for a picture that has been graded, tracked and conformed into a filmed frame. Nothing was established about what any of these tools costs for this work, which belongs to the cost article, or about their availability in the UAE, since a published API region is a region for an endpoint rather than evidence about commercial availability.

One negative needs its own limit stated. Three regional broadcast hosts were enumerated on 6 September 2026 and nothing naming a technical or delivery specification was found on them. That is a scoped negative about those hosts and paths on that date and not a claim that no UAE broadcaster publishes one. One of those bodies publishes primarily in Arabic, nobody on this project reads Arabic, and an Arabic-only document could not have been used even if it had been found. What a second language costs and changes is a subject of its own, in AI voice and Arabic localisation.

The generated half is bounded by what the filmed half hands it, and the bounds are published by the companies selling the models under headings they call input requirements. The ratio has to be settled before the day, because a frame handed over in the wrong one is cropped to fit from the centre. Seconds of clean subject and an end frame are deliverables of the shoot rather than afterthoughts. Four mechanisms want different things from the same footage and two of them want opposite things, so ask which one the supplier is using before deciding how to shoot. At the far end the file has to be something specific, and the bodies that decide what it has to be have never asked how a picture was made. Somebody has to decide how the generated element enters the colour pipeline, and where nobody has been asked, nobody has decided. And when the output is changed to fit a filmed frame, which it always is, none of the six providers publishes what that does to their cover.