Making a product advert with a neural network

An advert shows a product in action, explains what it is good for and, in a few seconds, moves the viewer towards buying. That job used to need a set, a camera operator, lighting, props, an editor and several days of work. Now a large part of the process can be done with a neural network: shaping the idea, writing the script, building a background, animating a photograph, adding camera movement, voicing the copy and assembling the finished video in the format a website, a marketplace or a social network needs. AI does not replace understanding of the product and the audience, though. Upload one photograph and ask for "a nice advert" and the result may look striking without explaining why anyone needs the thing. This guide walks through the whole process, from preparing the source material to the final check. The example prompts can be adapted for clothing, cosmetics, electronics, homeware, food, accessories, sports equipment and other categories.

A neural network can turn a text description into a short scene, add movement to a product photograph or alter footage that has already been shot. A still image of a bottle, for instance, can become a shot in which the camera slowly moves closer, a soft highlight slides across the surface and water or light mist appears in the background.

An advert is usually assembled from several short fragments. The first shot catches attention, the second shows the problem or the situation of use, the third presents the product, the fourth reveals the benefit, and the closing one carries the name, the price, the offer or the call to action. Generating a long clip in a single request is not always convenient, because the network may change the shape of the packaging, the colour, the lettering and small details. It is more reliable to create scenes separately and then combine them in an editor.

AI is particularly useful when a company cannot organise a proper shoot. It helps test several concepts quickly, prepare variants for different audiences and replace a background without reshooting. The accuracy of the product image still has to be controlled, though. For an advert it matters that shape, colour, logo, contents and the way it is used match reality.

What to prepare before making the clip

State the goal of the video. One clip should solve one main task. That might be introducing a new item, showing a benefit, explaining how to use it, comparing it with the familiar solution or inviting people to use a discount. Trying to tell the brand story, list every specification and demonstrate five use cases at once produces an overloaded video.
Define the audience. The same item can be advertised in different ways. A robot vacuum can be shown as a helper with daily cleaning or as the answer to pet hair for animal owners. The more precisely the audience is described, the easier it is to choose the plot, the setting and the wording.
Choose the main benefit. Viewers rarely remember a long list of properties, so a short video is better built around one strong point: compactness, durability, time saved, simple controls, a safe composition, convenient packaging or unusual design. The rest can go into the next clips.
Collect good product images. Prepare photographs from the front, the side, above and at an angle. The product should be well lit and its edges clearly readable. If you plan to remove the background, a plain surface without complex shadows works best. Low resolution, glare on the packaging and hidden elements all increase the risk of distortion.
Record the mandatory details. Write down the exact name, colour, material, contents, dimensions, the lettering on the packaging and the elements that must not change. That list will serve both when building prompts and at the final check. If the product has several variants, decide in advance which one this clip shows.
Prepare the brand assets. Keep the logo, typefaces, colours, slogan, packaging photographs and the rules for placing the mark separately. Lettering and the logo are usually safer to add at the editing stage than to ask a neural network to draw inside the scene, because generated text often comes out wrong.
Decide on the platform and the length. A vertical clip suits phone viewing, a square one suits certain posts and cards, a horizontal one suits a website, a presentation or a video platform. Ten to twenty seconds is enough for a short advert; demonstrating a complex product may need thirty to sixty.
Gather references. A reference is an example of mood, light, camera movement or composition. You can save a few frames from an advert, a catalogue or a film without copying someone else's idea wholesale. References help explain precisely whether you want a minimal studio piece, a domestic scene, an energetic sports treatment or a calm premium atmosphere.

The format depends on the product and on how familiar the audience already is with it. For something new it helps to start with a clear demonstration: show the object close up, then explain where and how it is used. For a familiar category you can build the clip around a problem — tangled cables, awkward cleaning, an overstuffed bag — and then present the product as the answer.
The "before and after" format suits care products, cleaning, repairs, organising space and teaching. It matters not to exaggerate the result or create an effect the product cannot deliver. The demonstration has to stay believable.
An unboxing clip shows the packaging, the contents and the first impressions. That scenario suits electronics, cosmetics, toys and gifts. If the network renders hands or fine details inaccurately, use real close-ups and apply AI to the background, the transitions and the overall styling.
Another option is a product scene with no people. The item sits on a table, a podium, stone, fabric or another surface, the camera moves smoothly, the light emphasises the form, and thematic elements appear around it. That format is easier to control, which makes it a good first attempt.

Step by step

State the core message. Write it as a single sentence with no advertising ornament. For example:

"A compact steamer gets clothes back in shape quickly, at home and on a trip"

That sentence becomes the semantic basis of the script. If it comes out too long, the clip is trying to do several jobs at once.

Choose the structure. For a fifteen-second clip you can use: problem — product — demonstration — benefit — call to action. For a calmer advert this sequence works: atmospheric shot — close-up — use — detail — end card. Set an approximate length for each scene.
Write a short script. Describe not the general idea but what the viewer will see on screen. Instead of "show convenience", write:

"A person takes a compact device out of a travel bag, switches it on with one button and smooths a crease on a shirt"

That kind of description is far easier to turn into a visual prompt.

Build a storyboard. Split the script into individual shots and sketch the composition roughly. You don't need to draw well. It is enough to mark where the product is, which shot size is used, where the camera moves and where text will appear. A storyboard helps spot repeated scenes and keep the logic.
Prepare the source image. Remove the unwanted background, fix the brightness, level the horizon and check the colour. Don't over-sharpen or over-smooth the surface, or the product will look plastic. If the packaging carries small print, keep the original file for overlaying later.
Build the prompt for the first scene. Include the object, the action, the place, the lighting, the camera movement, the mood and the constraints. Start with a simple description and then add detail. An over-long request with contradictory requirements can make the result worse.
Generate several variants. Even a precise prompt doesn't guarantee a usable shot first time. Compare not only how attractive the image is but how closely the product matches reality. Pick the version where the form, the proportions and the interaction with the surroundings look convincing.
Create the scenes separately. Use its own request and source image for each scene. Keep the same light, background and camera characteristics if the shots are meant to read as one clip. Repeat the key features in every prompt where needed.
Assemble the video in an editor. Put the fragments in order, trim the pauses and set the speed. Don't overload the clip with elaborate transitions. In product advertising a plain cut, a soft fade or a match on a similar shape usually looks tidier.
Add text and graphics. The headline must be readable in one or two seconds. Use a large typeface, good contrast and short wording. Don't cover the product or the important action. The name, the price and the terms of the offer must match what is on the purchase page.
Record the voice-over or add subtitles. The spoken text shouldn't repeat every caption word for word. The voice-over explains the meaning; on-screen text highlights the key facts. Since many people watch without sound, the main benefits must be clear from the picture and the subtitles.
Choose music and sound effects. The tempo should match the edit and the character of the product. A calm rhythm suits a household item, a more energetic one suits sports. The sound of a button, a spray, packaging opening or fabric moving strengthens the sense of reality, but must not be louder than the voice.
Make the end card. At the end show the product, the name, the key benefit and the action: "find out more", "choose a colour", "order" or "see the specifications". Don't put too much small print there. Promotion terms can go into the post description or on the page.
Export several versions. Save the video in the required proportions and check that the platform interface doesn't crop the captions. Sometimes it pays to rebuild the composition slightly rather than simply crop a horizontal clip into a vertical format.

Writing a script that holds attention

The first shot has to say quickly what the video is about. That doesn't require a hard cut or a large caption. You can show a recognisable problem, an unusual angle on the product or an action that raises a question. A drop of water lands on a waterproof case, a crease disappears as the device passes over it, a compact container slips easily into a small bag.

The script has to develop in sequence. If a problem is shown first, the next shot must lead to the solution. If the clip opens on a beautiful product shot, a function should follow, or the viewer will only have seen the looks. Each scene should add new information rather than repeat the previous one from another angle.

Real perception time matters. In one second a person cannot read a long phrase and study the product at the same time. When text appears on screen, the visual action at that moment should be simple. A complex demonstration is better left without a large caption, or split across two shots.

A five-scene script suits a wide range of products. The first shows the situation, the second introduces the product, the third demonstrates the action, the fourth reveals the result, the fifth carries the call to action. That structure reads clearly and is easy to shorten or extend.

Building prompts for video generation

A prompt should describe a visible scene. It is easier for a network to satisfy a request naming a specific object, action and setting than an abstract wish to "make it look expensive and beautiful". It helps to move from the main thing to the secondary: product and action first, then place, light, camera, mood and constraints.
In describing the product, state the shape, material, colour and position. If you are using a source photograph, you can write that the object's appearance must stay unchanged. For the action, pick one clear verb: rotates, opens, sprays, slides, folds, switches on. Several complex actions in one short scene raise the risk of errors.
Camera movement should be described directly too. "Slow push in", "smooth orbit around the product", "camera moves left to right", "static close-up" give a far more predictable result than asking for something "cinematic". Words about style can come after the main parameters, but they must not replace the description of the scene.

Example prompts for different products

A product shot of a cosmetic

"A glass bottle of clear serum stands on a pale stone podium, a few drops of water and soft reflections around it, the camera slowly pushes in towards the bottle, a neat highlight travels across the glass, clean minimal background, natural cool tones, realistic product photography, the shape of the bottle and the colour of the cap stay unchanged, no extra lettering and no new objects."

A kitchen appliance

"A compact hand chopper stands on a wooden kitchen worktop, a person puts pieces of vegetable inside and presses the button, the contents are chopped quickly, the camera shows the action in close-up and then pulls back smoothly, bright modern kitchen, daylight from the window, realistic hands, the device keeps its original shape, colour and button layout."

Clothing

"A person in a light mid-season coat walks along a city street, the fabric moves naturally as they walk, the camera tracks alongside smoothly, then shows the texture of the material and the neat stitching in close-up, soft morning light, neutral architecture behind, natural body proportions, the colour and cut of the coat do not change."

A sports accessory

"A sports bottle stands on a bench in a modern gym, a hand picks it up, opens the cap in one movement and takes a drink, then the bottle returns to the bench, dynamic but smooth camera movement, contrasty directional light, light drops of water on the surface, realistic scene, the logo and the shape of the bottle do not change."

Homeware

"A folding organiser lies folded on a bed, hands open it out and fit it into a wardrobe, then clothes are placed neatly into the compartments, the camera shows the process from above and moves to a wide shot of the organised shelf, bright bedroom, soft daylight, calm domestic atmosphere, the dimensions and the number of compartments match the original product."

Headphones

"Wireless earbuds sit in an open charging case on a dark matte surface, the camera slowly orbits the case, the indicator lights up softly, one earbud rises gently to show its shape, directional studio light, neat reflections, minimal technological atmosphere, no change to the design, the colour or the placement of elements."

Food

"A crisp biscuit lies on a pale ceramic plate, a cup of a hot drink beside it, one biscuit breaks in half to reveal the texture inside, a few crumbs fall naturally onto the plate, warm morning light, the camera pushes in slowly, realistic food photography, the packaging in the background keeps its original design."

A children's product

"A wooden construction set is spread out on a light rug, a child's hands join the large safe pieces and build a small figure, the camera shows the process from above, then moves smoothly to the finished model, cosy children's room, soft daylight, natural hand movement, no change to the shape or colour of the pieces."

Jewellery

"A slim silver bracelet lies on dark velvet, the surface of the fabric moves slowly, the bracelet turns towards the light source, the camera shows the clasp and the texture of the metal in close-up, soft highlights, restrained premium atmosphere, high detail, the design of the piece stays unchanged."

A cleaning product

"Marks and dust are visible on a glass surface, a hand sprays the product and wipes once with a cloth, the surface becomes clean and clear, the camera captures a readable before-and-after transition, bright domestic setting, realistic reflections, no excessive shine, the product packaging keeps its original colour and shape."

A prompt for the end card

"The product sits centred against a clean plain background, the camera pushes in slowly, free space remains beside it for a headline and a button, soft studio light emphasises the form, symmetrical and calm composition, no text inside the image, no additional logos, the original appearance of the product is preserved."

A prompt for vertical video

"Vertical 9:16 composition, the product occupies the central part of the frame, safe empty space remains at the top and bottom for interface and captions, smooth forward camera movement, the background does not distract from the object, important details are not placed near the edges, realistic advertising photography."

Refining a prompt after a failed generation

If the product changes shape, cut the number of actions and forbid changes to the appearance outright. Rather than describing the whole clip, make a close-up and move the complex demonstration into the next scene. This wording helps: "preserve the exact shape, proportions, colour and placement of details of the source object".
If the camera moves too fast, replace the vague "dynamic" and "striking" with something concrete: "slow even push in over four seconds" or "static camera, only the object moves". The fewer simultaneous movements, the more stable the scene.
If the hands look unnatural, try changing the angle. Instead of a close-up of fingers you can show the action from above, from the side, or partly hide the hands behind an object. For a complex unboxing it is sensible to shoot real footage and use AI to improve the background and the transitions.
If the background distracts, strip out the extra description and leave two or three elements. A request for "a modern beautiful kitchen with lots of detail, crockery, flowers, appliances and a view from the window" creates visual noise. Far more reliable: "bright kitchen worktop, neutral blurred background, daylight".
If packaging or lettering comes out distorted, don't keep generating in the hope of readable text inside the frame. Build the scene without small type and add the exact label image, the name and the price in the editor over the video.

The voice-over

The voice-over should complement the picture. If the viewer can see the device folding, the voice needn't say "the device folds". Better to explain the benefit: "Fits in a travel bag and takes up no room at home." That way you show the property and reveal what it means at the same time.

The script for a fifteen-second video should normally be short. It is read at a calm pace with small pauses. An over-packed voice-over turns the advert into a tongue-twister and stops people looking at the product. Reading the text aloud with a stopwatch before recording helps.

An example for a portable steamer: "Creases turn up at the worst possible moment. A compact steamer heats up in no time and gets your clothes back in shape at home or on a trip. One device, for when looking neat matters." The call to action can be shown separately on the end card.

For a product with several benefits it is better to make a series of clips: speed in the first, compactness in the second, easy care in the third. Short themed videos are easier to take in and let you see which message works best with the audience.

On-screen text and subtitles

On-screen captions should be shorter than the voice-over. Instead of "This device heats up quickly and helps save time", write "Ready to go fast" or "Less prep time".

Don't place important text too close to the edges. The platform interface may cover the lower part of the video with buttons, a description or controls. It is best to check the safe area in advance, especially for vertical formats.

Subtitles are useful even with a voice-over. They let people watch without sound and improve comprehension. Break the text into short meaningful phrases and sync it with the voice. Automatic transcription needs checking: names, numbers and specifications are often recognised incorrectly.

Keeping the product looking the same

One of the main difficulties with generated video is the object changing between shots. An extra button appears in the first scene, the shade shifts in the second, the packaging gets wider in the third. To reduce the divergence, use the same source photograph, repeat the key characteristics in the requests and don't overload a scene with too many actions.

It helps to choose a reference image of the product in advance. It becomes the basis for every scene and for the end card. Compare each generated fragment against it, watching not only the overall silhouette but the placement of seams, ports, lids, decorative elements and markings. A change the author doesn't notice may matter a great deal to a buyer who has already seen the product listing.

If the generation is right in mood but the packaging has shifted slightly, it is sometimes easier to replace the object in the editor with the exact image than to keep trying. For that, leave a clear background and free space around the product in advance. This combined approach keeps things truthful while still using the advantages of an artificial setting.

For catalogue and technical products, accuracy matters more than artistic effect. You must not show a function that doesn't exist, change the number of parts, exaggerate the size or create a false result. If the network can't handle a complex demonstration, shoot the real action on a phone in good light and use AI for cleaning the background, building the opening, the voice-over and adapting the format.

Common mistakes

No specific goal. A video can look good and still fail to explain the value of the product. Before starting, decide what action you expect from the viewer and what message leads to that action. One clip needn't answer every question a buyer has, so trying to fit in a full description usually weakens the main point.
Too long an opening. A few seconds of atmospheric background are fine in image advertising, but in a short selling video the object or the problem must appear almost immediately. If the viewer spends too long not understanding the subject, they may move on before the demonstration starts.
An overloaded prompt. A large number of objects, actions, effects and angles in one request makes generation harder. Better to split the idea into short scenes, each with one clear action, and then join them in the edit.
Changed product characteristics. A network can add a detail, change a material or make packaging more symmetrical than it really is. Every scene has to be compared against the real item. Colour, shape, the number of elements, the way it opens and what's included deserve particular attention.
Trying to generate small print. Names, contents, prices and offer terms come out distorted. It is more reliable to add them in the edit, using the original logo, the correct spelling and the brand typefaces. That also makes updating a price easy without regenerating the whole shot.
Overdoing the result. Implausible shine, instant removal of stubborn dirt or impossible working speed reduce trust. Advertising should emphasise benefits without promising what the product doesn't do in ordinary use.
The wrong frame format. A horizontal video cropped automatically at the edges can lose the product and the captions. Check the composition for each platform and, where needed, make a separate version in which the main object sits inside the safe area from the start.
Small, long captions. On a phone screen, detailed specifications become unreadable. Leave one short message in frame and put the extra information in the description or on the product page. The typeface must stay readable on a small screen too.
Music louder than the voice. Even a good voice-over loses its meaning when the background drowns it. Check the balance in headphones and through a phone speaker before exporting. The voice should be effortless to follow, with the music track supporting the mood.
No check after saving. The final file can differ from the preview: compression artefacts, drift out of sync, cropped elements or sound that is too quiet. Watch the video all the way through after export rather than only inside the editor.

Watch it without sound first. It should be clear what is being advertised, how the product is used and what its benefit is. If the meaning rests entirely on the voice-over, strengthen the demonstration or add short subtitles — without turning the clip into a several-paragraph presentation.

Then turn the sound on and assess the sync. Voice, music and effects should support what is happening on screen. If the voice-over announces the result before it is shown, the viewer finds it harder to connect the information. Sometimes moving a phrase or a sound effect by a fraction of a second is enough.

Check the product against the reference photographs. Compare shape, colour, material, logo, buttons, lettering and contents. Any noticeable error can raise questions for a buyer. For children's, food, medical and technical products, the truthfulness of the use and the result deserves particular care.

Then open the clip on a phone. Make sure the text is readable, the product isn't covered by the interface and the important details aren't at the edge. A video that looks good on a large monitor sometimes turns out too small on a mobile screen.

Using AI honestly

An advert must not show properties the product doesn't have. A neural network produces a spectacular result easily, so there is a risk of accidentally exaggerating effectiveness, speed or scale. Before publishing, check every claim against the specifications, the instructions and real experience of use.
Rights to music, images, typefaces, voices and other material need checking. Being able to download a file doesn't always mean permission to use it commercially. For advertising, choose content with clear terms or the company's own material.
If a realistic person appears in the video, don't create the impression that a specific well-known individual endorses the product without their consent. Nor should you use an ordinary person's likeness in a context that could damage their reputation or mislead the audience.
Certain categories carry specific requirements for advertising, wording and warnings. Before publishing medical, financial, children's, food and other regulated products, check the current rules of the country and the platform. AI speeds up production, but responsibility for the content stays with the advertiser.

The best results come when the neural network plays a specific role: building the setting, animating a photograph, suggesting camera movement or speeding up the preparation of variants. Meanwhile the person keeps control of the meaning, the truthfulness and how well the product matches reality.