AI in food and product video the shots it cannot hold, the ones it can, and who says so

Every food and product shot makes two claims at once. It says this is one specific real object, the client's own pack or the kitchen's own dish. And it says that object is doing one specific physical thing: pouring, melting, beading, fracturing, turning under a light. Everything below tests one half or the other.

The most useful opening evidence is mechanical rather than promotional, and it comes from a company selling a model. Amazon's AI Service Card for Nova Reel, which the company publishes as applying to "the release of Amazon Nova Reel that is current as of July 16, 2025", states that the model "is trained from images and videos of objects" and that "It does not store explicit 3D models of objects, or model lighting physics". The same card adds that the model "does not know about every possible object (including living creatures) or every possible variation of an object".

Read that against what a packshot is. A packshot is a three-dimensional object under lighting arranged to describe its surface, its geometry and its colour. The card describes a system that holds neither the geometry nor the lighting model.

Two fences before anything else. Whether a generated depiction is allowed is a separate question with its own answer, covered in AI generated video rules in the UAE. This article asks whether it is possible. And no price appears here for either route, because the cost comparison carries them in both units.

Every source in this section comes from optics, food science or a colour standard, and not one of them has anything to do with generative tools. That independence is the point. A pour, a melt and a bead are the shots a menu launch film is actually bought for, and this is what they are made of.

The pour is a regime problem

A pour looks like one thing and behaves like three. A paper on liquid jet breakup identifies "three distinct breakup regimes, such as dripping, Rayleigh, and wind-induced, each dictated by the interplay of surface tension and aerodynamic forces for various combinations of liquid jet velocity and needle exit diameter". Which regime the shot is in decides what the column does on the way down.

The counter-intuitive part is in the same abstract. "Increasing jet velocity extends breakup lengths for moderate flow rates due to enhanced stability in the Rayleigh regime, but higher velocities induce instability, leading to shorter breakup lengths." Pour faster and the unbroken column first gets longer, then gets shorter. The paper also reports that lower velocities give narrow distributions of breakup length and higher ones give broader distributions, so the length is a spread rather than a number.

Scope it carefully. The work is water from needles under controlled pressure, not sauce from a jug. It establishes that the regimes exist and are measurable. It gives no figure for any product.

Steam, melt and condensation are on a clock

Steam only counts when it moves. A peer-reviewed paper in Food Quality and Preference reports that "animated traces of steam added to food images not only induced hot temperature perception of the food, but also increased food desirability and freshness perception", while "implied animation (i.e., static picture of rising steam) did not produce any such effect". That is the half of a food frame a still cannot carry, and somebody measured it. The provenance travels with the quote: this is the abstract as deposited by the publisher, read through OpenAlex for the DOI and cross-checked against Crossref, because the publisher's own page serves a captcha wall.

The clock has numbers on it. Imaged at 20 ± 0.5 °C, mean ice-crystal diameter in ice cream fell from 66.43 ± 2.07 µm to 45.74 ± 3.92 µm over ten minutes of meltdown, mean air-bubble diameter fell from 88.30 ± 4.52 to 63.25 ± 4.91 µm, and the air volume fraction fell from 41.20 ± 3.90 to 24.25 ± 3.25 per cent. The dessert on the stand is a measurably different object every minute.

Condensation does not reset either. In dropwise condensation the smallest drops grow "principally by the diffusion of water adsorbed on the substrate to the drop perimeter", drops larger than 50 µm grow by direct deposition from the vapour, and "Drop coalescence plays a critical role in determining the drop size distribution". The beading on a cold bottle is a size distribution with a history behind it.

Then continuity, which is where the yardstick gets uncomfortable for everyone. Food scientists building a laboratory test for the mozzarella pull report that "Stretchability is one of the most important functional properties of Mozzarella cheese, but no objective and widely accepted technique exists for evaluation; most of the cheese stretchability tests are influenced by the ambient conditions", and put their own test in an oil bath to escape "subjective stretching speeds and variations in ambient temperature and humidity". If a laboratory cannot repeat a cheese pull without controlling the room, two takes of one are not the same shot either.

Light goes into food before it comes back out

The failure mode here has a name and it is twenty-five years old. A SIGGRAPH 2001 paper on subsurface light transport states that "using the BRDF creates a hard, distinctly computer-generated appearance because it does not locally blend surface features such as color and geometry", and lists the translucent materials that only subsurface scattering describes as "marble, cloth, paper, skin, milk, cheese, bread, meat, fruits, plants, fish, ocean water, snow". That is a food list.

Scope it: a 2001 paper about a rendering model establishes what the phenomenon is and what its absence looks like on screen. It establishes nothing about any tool sold in 2026 and must not be stretched into one.

Gloss is an engineering quantity as well. ISO 2813:2014 defines gloss as the "optical property of a surface, characterized by its ability to reflect light specularly", sets the scale by assigning polished black glass the value 100, specifies the 20°, 60° and 85° geometries, and states that "It is not permitted to interpret and express gloss values as '% reflection'". The same plate under two lamp positions returns two answers. Scope that too: ISO 2813 is a paint and varnish standard for non-textured coatings on plane opaque substrates. It does not measure the gloss of a wet dish, and nothing here says it does.

One more, because it earns its place twice. A February 2026 paper in Foods states that "crispness-related oral sensations arise from fracture, sound emission, and lubrication during mastication". The crunch and the picture of the crunch are the same physical event read by two different instruments.

A brand colour is a registered coordinate

One sentence on the legal half, and no more. The UAE's Trademarks Decree-Law defines a trademark as "everything that takes a distinctive shape of names, words, signatures, letters, Symbols, numbers, addresses, seals, Drawings, Pictures, Engravings, packaging, graphic elements, forms, colour or colours or a combination thereof", which puts packaging and colour on the same list. That text is a Lexis Middle East translation served from the Ministry's own host, and no official English version was obtained.

How a colour gets registered is the part that matters here. WIPO records that offices "normally require, in the context of registration procedures, that the color be identified either by a deposited sample or by reference to a recognized color code (for example PANTONE®)", and that "a mere sample of a color is insufficient, as a sample could deteriorate with time". A brand colour is a code because a code does not decay.

The difference between two colours has an international formula. ISO/CIE 11664-6 specifies CIEDE2000, "incorporating corrections for variation in colour-difference perception dependent on lightness, chroma, hue and chroma-hue interaction". Its own scope excludes stimuli perceived as primary light sources or as specularly reflected light, and a video frame on a phone is a light source.

A tolerance, meanwhile, is a negotiated band with an instrument attached. ISO 12647-2:2013 states that its tolerance values "have therefore been chosen to provide a reasonable balance between customer expectations (meaning small variations), technical production limits and production costs", measured "in accordance with ISO 13655-D50 illuminant, 2° observer, 0:45 or 45:0 geometry". Both ISO texts quoted here are official preview extracts from a standards distributor, because the publisher's own host refused every retrieval route, so the standards are named rather than read in full.

Colour management, meaning how a pipeline carries a colour from one device to another, is a different subject. Matching generated material to filmed material covers it.

Fluids, and the operations a food film is made of

The sharpest published sentence in the whole subject is one clause long and it names three shots. A July 2026 paper on physics-grounded fluid video generation states that video diffusion models "routinely violate elementary physics when the subject involves fluids: liquid columns break apart in mid-air, container water levels fail to rise as liquid is poured in, and splashes disperse without regard to momentum or gravity". Those are the pour, the fill and the splash, in that order.

Set it beside the yardstick. A different discipline established that a liquid column's breakup length is governed by a named regime and moves non-monotonically with velocity. This paper says the columns break in mid-air. Two fields describing the same phenomenon from opposite sides.

Two cautions travel with it. It is a method paper proposing a fix, built on 1,638 simulated pouring and sloshing videos and 2,320 real ones, and its headline improvement of up to 8.75 points on VideoPhy-2 describes its own model rather than any tool a buyer is sold. And its opening sentence is that paper's characterisation of the field on that date, not a statement about any named commercial product.

Procedural food work has been measured directly. PhysVidBench is 383 prompts "emphasizing tool use, material properties, and procedural interactions", run across twelve named text-to-video models including Sora and Veo-2. Its prompts read like a call sheet: pouring flour into a bowl of water and stirring, layering a green liquid over the back of a spoon into a shot glass of brown liquid, separating an egg yolk with a water bottle. It concludes that "current models still fall short of robust, generalizable physical commonsense reasoning", with accuracy roughly between 20 and 36 per cent across its physical reasoning dimensions, scoped to the twelve models it tested on that date.

Drift over the length of a shot has a mechanism rather than being a mystery. A March 2026 paper on hand-object interaction states that models "must generate physically plausible contact and preserve object identity across frames", and proposes mitigating "the temporal decay of 3D RoPE to improve long-range object consistency". A pack that is right in frame one and wrong in frame ninety is a decay somebody has named. Scope it: that paper is about digital humans and reenactment, not food and not a kitchen, and it establishes nothing about a spoon, a dip or a garnish drop.

What the limitation pages do not say

A limitations page is only as useful as what it covers. Google's Veo 3 model card, published 23 May 2025 and updated 13 January 2026, states that it "covers Veo 3 and subsequent versions", and its entire Known Limitations section is one sentence about consistency. Counted across that whole document as read on 6 September 2026 there is no occurrence of colour, logo, brand, trademark or packaging. The Veo 3.1 Lite card, published April 2026, then refers its own Known Limitations back to the Veo 3 card, so the published limitations for the newer model are the January 2026 text of the older one.

OpenAI's Sora 2 System Card, dated 30 September 2025, has eight sections and all eight are safety sections. Counted across its full text there is no occurrence of text rendering, typography, colour or consistency, and no capability limitations are published at all.

Adobe's own answer to what the top use cases are for its text-to-video and image-to-video features lists five, on a page stamped 4 September 2025: atmospheric elements over a black or green screen, animating an image, 2D or 3D animation including claymation, graphics and text videos, and "B-roll focused on plants, animals, and landscapes". Not one of the five is a product, a packshot, a pour or a dish. That is an omission from a vendor's own list on a dated page. It is not Adobe saying it cannot do food, which Adobe does not say.

Sound belongs here in three sentences. A crisp break emits a measurable acoustic envelope, so a crunch is a physical event rather than an effect laid over a picture. Amazon publishes that Nova Reel "does not currently support audio or 3D content", and Microsoft publishes that "Sora 2 supports audio generation in output videos". Nothing located measures whether generated audio matches a generated food action, and voice, dubbing and Arabic localisation are a separate subject with their own article.

The pass class is a real photograph with a generated surround

Three vendors publish a route in which the merchant supplies a finished product photograph and the model generates only what is around it. Google publishes its change background feature as "a fast and cost-effective solution for generating high-quality product images that showcase products in natural, lifestyle-like settings", and its own instruction steps keep the two things apart: select or upload the product image, describe the product, then "Describe the scene in which you want your product placed in". That page prints no last-updated date in its body, which travels with every quote from it.

The mechanism turns up in an operator step, and this is the sentence the section rests on. Adobe instructs the user that "To isolate your product, Select Subject, copy and paste it into a new file created in your chosen size and position it on the canvas where you want it to appear", and only then to type a description of the background. The product survives because a person ran a selection and pasted the real pixels onto a new canvas. Nothing about the model is doing the preserving.

Amazon's contribution is two corporate announcements rather than technical documents, and it is attributed that way. Amazon publishes that advertisers "can use AI creative studio to research, generate, refine, and publish a variety of high-quality ad creatives produced from a single product shot, product page, or existing ad creative", with the studio "currently available in beta for U.S. advertisers".

Now the central negative. Across the three vendor routes above, not one publishes that the supplied product's original pixels are preserved in a generated scene. All three describe the workflow. None specifies the mechanism. That is scoped to those three named documents at the retrieval routes recorded on 6 September 2026. It is not a claim that no such specification exists anywhere, and it is not a claim that the pixels are or are not preserved. It is that a buyer cannot find out from the vendor.

The asymmetry makes the point sharper. Google does publish a preservation sentence, "Only your product will remain on a white background to help shoppers view your product clearly", and publishes it about removing a background rather than generating one.

Stability is not accuracy

The measured picture splits cleanly, and the split is the thing to carry out of this article. LivingWorld, an April 2026 paper benchmarking environmental dynamics across "60 scenes from public images collected from Pexels and Unsplash, covering four categories of environmental dynamics: clouds, water, smoke/fog, and fire, with 15 scenes per category", reports for Veo 3.1 a Motion Smoothness of 0.992 and a Temporal Flickering of 0.979, against a physical-realism score of 0.622. Atmosphere holds together and does not strobe. It is not physically right. Three cautions go with those numbers: it is a method paper whose own model is the headline, fifteen scenes to a category is a small sample, and its source images are stock photographs with no product in frame.

A second pattern points the same way, with both its caveats carried in the same breath. VBench++ defines Subject Consistency by "the DINO feature similarity across frames" and Background Consistency by "calculating CLIP feature similarity across frames", and across the fifteen named models in its reported table the background score is higher on eleven, equal on one and lower on three. The two dimensions use different feature extractors, so this is a pattern across a table rather than a like-for-like comparison, and the paper does not draw it. And Background Consistency measures temporal stability rather than accuracy. A high score means the background does not flicker, not that it is the right background.

The entry condition is published too, and it is narrow. Google publishes that Product Studio "works best with certain products. For example, consumer packaged goods with products front and center", tells merchants to place the product at the centre of the image and leave most of the canvas white, and advises against images that contain humans using the product or that involve hands or other props. A single centred packaged object on a plain field, with no hands, no props and no person. That is a packshot.

One policy line, read as capability rather than as law. Adobe prohibits "entering text prompts designed to generate copyrighted, trademarked, or otherwise infringing content, uploading an input or reference image that includes a third party's copyrighted content". The prohibition is scoped to third-party rights, so it does not restrain a brand generating around its own pack. What it marks out is that generic, non-identifiable surround is the class that survives the policy intact.

The division between a real subject and a synthesised surround is not new to the industry, and it has been specified elsewhere: Epic Games publishes that in-camera VFX "relies on a mixture of LED lighting, live camera tracking, and real-time rendering with off-axis projection", which is real-time rendering in a game engine rather than generative video.

Every absence below is scoped to what was searched on 6 September 2026, and none of it is a statement about the world.

Nobody publishes a success, retry or acceptance rate for a food or product generation, on either side of the line. Three neighbouring Journal articles record the same absence for generation in general, so it is not new here.

No measured colour figure for a generated frame of a real product was located anywhere. No delta-E, no tolerance, no pass rate against a brand colour standard, searched across arXiv, OpenReview and the published documentation of Amazon, Google, Adobe, Runway and OpenAI. Amazon publishes an image guidance feature that lets a user "Control precisely the color palette of generated images by providing a list of hex codes along with the text prompt", and nothing located publishes what comes out of it measured against what went in.

The absence is symmetrical, and a reader is owed that. No measured colour accuracy for any moving image of a pack was located either. Not for film, not for print to screen, not for anything. The filmed route does not have a number attached that the generated route lacks.

No located evaluation scores text rendering on a consumer pack. No located evaluation scores general-purpose video models on hands using utensils on food. Nothing located measures whether a generated background is correct rather than merely stable, and no located evaluation scores abstract texture or pattern generation as a dimension at all. That last one is the easiest place in this whole subject to overclaim, and it is worth naming as a blank rather than promoting as a strength.

Appetite needs its narrowing stated. A peer-reviewed paper does measure desirability, freshness perception and willingness to pay against an animated steam cue on real food images. What is absent is any equivalent measurement of generated material.

A food or product shot is a claim about one specific real object doing one specific physical thing, and both halves are measured by people with no stake in generative tools. A pour is a named regime that behaves non-monotonically with speed. A melt is a clock with micrometres on it. A translucent surface has a documented failure mode that is twenty-five years old and reads as computer-generated when it is missing. A brand colour is a registered coordinate rather than a look. Set the published capability record against that and the tool makers supply most of the answer themselves: Amazon on not knowing a product's dimensions and inventing its unseen sides, Adobe and Microsoft on blocking branded and photorealistic input on named routes, a benchmark on most models scoring below 60 out of 100 on text. And where generated work does pass, it passes because a person took the product out of the model's job first.