A frame from our master clip, generated by Kling 4.0 from a single studio photo of the product.
Personalisation is one of the most reliable ways to improve marketing performance. A message that reflects the customer’s own context, whether that is a gym, an office or a school run, simply feels more relevant than a generic one. Customer data platforms have made it easy to decide which audience should see which message. What has remained hard is producing the message itself, especially in video. Every audience-specific version of a video ad has traditionally meant another location, another shoot and another edit.
Video generation models promise to remove that bottleneck. If a model can take one photo of your product and place it, faithfully, into a different scene for each audience, the cost of a new variant drops from thousands of dollars to about one. We tested that promise with Kling 4.0, using a fictional product: a sage green insulated water bottle with a bamboo lid. From one studio photograph we generated a master clip, four audience-specific contexts, two edits of the master and an extension, and recorded exactly what each one cost.
The eleven clips cost $14.00 in total. Most of them would be usable in a campaign. One detail, which we found only by zooming in, changed how we think about reference images, and it is the most important finding in this article.
How Kling 4.0 handles references
Kling 4.0, model ID kling-v4, combines several modes behind one endpoint. Text-to-video needs only a prompt. Image-to-video uses an image as the first frame. Reference-based generation treats images as references for a subject that the model then places in a new scene. Video editing changes an existing clip, and video extension continues one. The mode is selected by the assets you send and an optional omni_reference_task_type field. Every Kling 4.0 clip comes with stereo sound, and resolution is either 720p or 1080p. We used 720p throughout.
For personalised ads, reference-based generation is the key mode. It lets you describe any scene and say “the bottle from the reference”, rather than being locked to the composition of your product photo.
{
“model”: “kling-v4”,
“prompt”: “The sage-green water bottle with a bamboo lid from the reference sits on a bench in a modern city gym, a woman finishes a set on the rowing machine, grabs the bottle and drinks, gym ambience.”,
“image_with_roles”: [{“url”: “https://…/bottle.jpg”, “role”: “reference_image”}],
“omni_reference_task_type”: “reference”,
“duration”: 8,
“resolution”: “720p”,
“aspect_ratio”: “16:9”
}
Reference clips took about two minutes each to generate. Edits took two to three minutes, and the extension five.
Test 1: the master clip
Our reference was a clean studio photo of the bottle on a light grey background, itself generated with an image model for this test. The master clip placed it in the brand’s hero setting.
The water bottle stands on a wet mossy rock beside a clear mountain stream, morning sun, a hand reaches in and picks it up, droplets fall, sound of running water and birds.
Test 1: the reference photo and two frames from the 8-second master clip.
The result looked like a real outdoor product shoot. The bottle stood on the moss with water flowing around the rock, the camera moved in close, and a hand picked the bottle up, leaving a small pool of water where it had stood. The colour, the matte finish, the proportions and the bamboo lid all matched the reference. Like every Kling 4.0 clip, it came with a stereo sound track. The eight-second clip cost $1.256, which is $0.157 per second at 720p.
Test 2: four audiences, four contexts
With the master in place, we generated four audience-specific variants from the same reference photo: an athlete in a gym, a professional in an office, a parent on a school morning and a hiker on a trail, the last one in 9:16 for stories and short-form placements.
Test 2: four contexts generated from one reference photo.
Each variant told its own small story and kept the product recognisable:
- Gym. A woman finishes her set, picks up the bottle from the bench and drinks. The gym looks contemporary and the lighting is convincing.
- Office. A man in a shirt at a laptop picks up the bottle and takes a sip, a natural moment that fits a “stay hydrated at work” message.
- School morning. A parent slides the bottle into a backpack in a sunny kitchen, and a child in a yellow hoodie zips the bag and heads out.
- Hiking, vertical. A close-up of the bottle clipped to a backpack, then a hiker walking away along a forest trail at sunset. The model recomposed the scene for the vertical frame rather than cropping a wide one.
In every clip the bottle kept its colour, shape and bamboo lid. The one visible deviation was in the vertical variant, where the lid gained a small metal loop for clipping to the bag, a plausible detail that the real product does not have. The model added it because the prompt asked for the bottle to be clipped on. If an accessory does not exist, do not write a scene that requires it.
Each variant cost $1.256. Four audience-specific video ads from one product photo cost just over five dollars.
Test 3: the fine print on the reference
When we zoomed into the master clip, we found a line of small text under the logo, “WANDER OFTEN STAY WILD”, that we had not written into any prompt. Our first assumption was that the video model had invented a slogan. It had not. When we zoomed into the original reference photo at full resolution, the same two lines were there, printed under “PURE TRAIL”. The image model that created our product photo had added them, and we had simply not noticed at normal viewing size.
Test 3: the tagline was on the reference all along, and Kling 4.0 reproduced it faithfully, even when the prompt said “no other text”.
That makes the result impressive rather than worrying. Kling 4.0 reproduced the label in close-up, logo, name and four words of fine print, accurately enough to read, in every clip where the label was visible. To see whether a prompt could override the reference, we generated the master again with an extra sentence: “The bottle label shows only the small mountain logo and the words PURE TRAIL exactly as in the reference, with no other text.” The tagline was still there. When the prompt and the reference disagree about the product, the reference wins.
For a brand, this cuts both ways. It is exactly what you want when the reference is your real, approved product photo: the packaging comes through intact. But it also means that any mistake in the reference, an outdated label, a retouching artefact, or, as in our case, text added by an image model, will be copied into every variant. You cannot fix it in the prompt. Audit the reference image at full resolution before you generate anything from it.
Test 4: editing the master instead of regenerating
Many variants are not new scenes but small changes: a different colourway, a seasonal setting, a market-specific detail. Kling 4.0 can edit an existing clip while keeping its motion. We tried two edits of the master.
Change the colour of the water bottle in @Video1 to a matte coral pink, keep the bamboo lid, the stream, the hand and all motion exactly the same.
Test 4: the colour edit was precise; the winter edit added snow but did not freeze the stream.
The colour edit was the most impressive result of the test. The bottle became coral pink while the bamboo lid, the moss, the flowing water, the hand and the timing of every movement stayed exactly as they were in the master. A product team could use this to show every colourway of a product in the same hero clip, with perfect consistency between versions.
The winter edit was good but incomplete. Snow covered the rock and the banks, the trees were dusted white, and the bottle and the hand movement were unchanged. But we had also asked for the stream to be partly frozen, and it flowed as before. Large environmental changes worked; specific secondary details were less reliable.
Both edits came back one second longer than the source, at nine seconds, and both cost $1.607. Kling bills an edit for the output length plus the input length, so per second of output an edit costs about $0.18, slightly more than generating a new clip. The premium buys exact continuity: the variants are the same shot, not similar shots.
Test 5: extending the master
Ads often need a longer cut. We asked Kling 4.0 to continue the master for five more seconds with a specific action.
Continue the shot: the hiker unscrews the bamboo lid, drinks from the bottle and looks out over the valley, sound of the stream.
Test 5: the extension continued the scene but skipped the close-up action we described.
The returned clip was 13 seconds long: the original eight seconds followed by a new five-second segment. The extension cost $0.746 and was billed only for the new seconds. The new segment did continue the scene, cutting to a wide shot of a hiker sitting on a rock by the river and drinking. But it skipped the unscrewing of the lid, and at that distance the bottle is too small to recognise. As atmosphere at the end of an ad it works; as a product demonstration it does not.
Test 6: plain image-to-video
For comparison, we also used the studio photo as a first frame, the simplest mode, and asked for a slow orbit as condensation formed on the bottle.
Test 6: image-to-video kept the studio shot almost static.
The clip was faithful to the photo, but almost static: a soft light sweep across the bottle, with no visible orbit and no condensation. It was cheaper, $0.785 for five seconds, and it took over five minutes to arrive. For a product listing it could work as a subtle animated image. For an ad it tells no story. Reference mode was far more useful, because it frees the product from the composition of its photo.
Test 7: 720p or 1080p for a hero clip?
Finally, we regenerated the master at 1080p with the same prompt and reference, to see what the higher resolution buys.
Test 7: 1080p added visible water droplets and crisper texture; the label was correct in both.
The 1080p version was noticeably richer at the same moment: water droplets beaded on the bottle, the matte finish showed more texture, and the moss was sharper. The label was reproduced correctly in both versions. The 1080p clip cost $1.786, or $0.223 per second, 42 percent more than 720p, and took 152 seconds to arrive. For the hero clip that leads a campaign, the extra detail is worth it. For the audience variants, which will mostly be seen at phone size in a feed, 720p is the better value.
What it costs, and a plan for a real campaign
Billed cost per job, and per second of output.
Before you scale this up, it helps to compare the Kling 4.0 API pricing for each mode with what you actually need. In our test, a new reference clip and an image-to-video clip both cost $0.157 per second at 720p. An edit cost about $0.18 per output second because the input is billed too. An extension cost $0.149 per new second. Every task returned its billed cost with the result, which is what made it possible to report these numbers to the cent.
A practical plan: one master, four audience contexts and one colour edit for under eight dollars.
For a real campaign we would generate one master in reference mode, four audience-specific contexts from the same reference, and one colour edit of the best-performing variant: six clips for $7.89. Doubling that to allow for regenerations still leaves the full set under $20, a fraction of the cost of a single day of location filming. The constraint moves from production to review: someone has to approve the reference image and check every clip at full resolution, for invented accessories like the clip loop in Test 2 and for anything inherited from the reference.
Connecting variants to audiences
Generating variants is only half of personalisation. The other half is making sure each audience sees the right one. In practice, that means naming every clip after the segment it is made for, storing the prompt and the reference image next to it, and uploading each variant to the ad platform or the customer data platform with the matching audience rule. Because the API returns a task ID and a cost for every clip, it is straightforward to log all of this automatically: which reference was used, which prompt, which segment, what it cost and when it was approved.
That record also protects you. If a reference turns out to contain an error, such as an outdated label, you can find every clip generated from it and replace them all, rather than discovering the problem one complaint at a time.
The walkthrough video
Walkthrough: one product photo, eleven Kling 4.0 clips (about 3 minutes), including the real footage.
Recommendations
- Use reference mode, not image-to-video, for ads. It places the product in a new scene; image-to-video barely moved it.
- Write one context per audience. Four distinct stories from one photo cost about five dollars.
- Audit your reference at full resolution. Kling copied the fine print of our reference faithfully, including text we had not noticed, and the prompt could not remove it.
- Do not write scenes that need features the product lacks. Asking for a clipped-on bottle produced an invented clip loop.
- Edit for colourways and seasons. Edits kept the exact motion of the master, which keeps a campaign consistent.
- Budget edits correctly. They are billed for input and output seconds, about $0.18 per output second at 720p.
- Use extensions for mood, not demonstration. The extension skipped the specific action we asked for.
- Log everything. Prompt, reference, segment and cost for every clip make review and correction fast.
Limitations
We generated one take of each variant, so we cannot say how often a given prompt would succeed, and our judgements of product consistency were made from frames extracted from each clip rather than with a formal measure. Our product is simple, a single-colour bottle with a small label; products with complex packaging, transparent materials or many small parts may be harder to keep consistent. The extension and edit results describe one attempt each.
Appendix: the prompts we used
For anyone who wants to repeat these tests, here are the prompts as we sent them, grouped by clip. Parameters such as duration and resolution are listed after each prompt.
k3-master (model: kling-v4, duration: 8, resolution: 720p, aspect_ratio: 16:9)
The water bottle stands on a wet mossy rock beside a clear mountain stream, morning sun, a hand reaches in and picks it up, droplets fall, sound of running water and birds.
k3-ref-gym (model: kling-v4, duration: 8, resolution: 720p, aspect_ratio: 16:9)
The sage-green water bottle with a bamboo lid from the reference sits on a bench in a modern city gym, a woman finishes a set on the rowing machine, grabs the bottle and drinks, gym ambience.
k3-ref-office (model: kling-v4, duration: 8, resolution: 720p, aspect_ratio: 16:9)
The sage-green water bottle with a bamboo lid from the reference stands on a tidy office desk next to a laptop, a man in a shirt picks it up during a video call and takes a sip, quiet office ambience.
k3-ref-school (model: kling-v4, duration: 8, resolution: 720p, aspect_ratio: 16:9)
The sage-green water bottle with a bamboo lid from the reference is packed into a child’s school backpack by a parent in a sunny kitchen in the morning, the child zips the bag and runs out, kitchen sounds.
k3-ref-vertical (model: kling-v4, duration: 8, resolution: 720p, aspect_ratio: 9:16)
The sage-green water bottle with a bamboo lid from the reference on a hiking trail at sunset, a hiker clips it to her backpack and walks on, wind sound.
k3-i2v (model: kling-v4, duration: 5, resolution: 720p)
The camera slowly orbits the bottle as fine condensation forms on its surface, soft studio light sweep.
k3-edit-coral (model: kling-v4, resolution: 720p)
Change the colour of the water bottle in @Video1 to a matte coral pink, keep the bamboo lid, the stream, the hand and all motion exactly the same.
k3-edit-snow (model: kling-v4, resolution: 720p)
Edit @Video1 so the scene is a snowy winter forest: snow on the rock and banks, the stream partly frozen, keep the bottle and the hand movement identical.
k3-extend (model: kling-v4, duration: 5, resolution: 720p)
Continue the shot: the hiker unscrews the bamboo lid, drinks from the bottle and looks out over the valley, sound of the stream.
k3-master-notext (model: kling-v4, duration: 8, resolution: 720p, aspect_ratio: 16:9)
The water bottle stands on a wet mossy rock beside a clear mountain stream, morning sun, a hand reaches in and picks it up, droplets fall, sound of running water and birds. The bottle label shows only the small mountain logo and the words PURE TRAIL exactly as in the reference, with no other text.
The bottom line
Kling 4.0 can turn one product photo into a set of audience-specific video ads that look like they came from a real production, for about $1.26 per eight-second variant. Its reference mode kept our product consistent across five scenes, and its editing mode changed the product’s colour while preserving every frame of motion. Its fidelity to the reference is so strong that the reference itself becomes the thing to check: fine print we had overlooked came through in every clip, and a prompt could not remove it. With an approved product photo and careful review, it makes personalised video practical for teams that could never have afforded it before. If you want to build your own variants, or compare Kling 4.0 with other video models through one account and API key, all of them are available on APIMart.

























Email
SMS
Whatsapp
Web Push
App Push
Popups
Channel A/B Testing
Control groups Analysis
Frequency Capping
Funnel Analysis
Cohort Analysis
RFM Analysis
Signup Forms
Surveys
NPS
Landing pages personalization
Website A/B Testing
PWA/TWA
Heatmaps
Session Recording
Wix
Shopify
Magento
Woocommerce
eCommerce D2C
Mutual Funds
Insurance
Lending
Recipes
Product Updates
App Marketplace
Academy