Kling 3 for ad creatives: costs, prompts, results

We tested Kling 3 on the WondraKids Meta ad account for three months, from November 2025 through January 2026. Ran it side-by-side with our real UGC creator pipeline and against our static-image ad baseline. Shipped 34 Kling-generated video

We tested Kling 3 on the WondraKids Meta ad account for three months, from November 2025 through January 2026. Ran it side-by-side with our real UGC creator pipeline and against our static-image ad baseline. Shipped 34 Kling-generated video ads in that window, of which 12 ran long enough to produce statistically-useful performance data. The kling 3 ad creatives summary for a mid-size DTC brand is this: synthetic video is now good enough to sustain a real portion of an ad account's creative output, performs at roughly 75% the per-impression ROAS of human-shot UGC, costs roughly 5% of what UGC costs when fully loaded, and is ready for production use inside a disciplined QA workflow — but will not fully replace real humans on camera in 2026, and shouldn't. This is what we tested, the specific prompts and workflow, the performance numbers, and where we'd rather still pay for a real shoot.

The setup

Three ad-creative tracks running concurrently:

1. Real UGC track. 4–6 shoots per month with paid micro-influencers, €800–1,500 per shoot, producing 3–5 short clips per shoot.
2. Kling synthetic track. Text-to-video and reference-image-to-video via Kling 3, 34 ads produced across 3 months.
3. Static control. Our existing static-image ad line running unchanged.

Everything ran through the same Advantage+ Shopping campaign to avoid audience-level confounds. Creative-level attribution via Meta Ads Manager plus Triple Whale cross-check.

The workflow we converged on

After the first week of trial and error, we standardized on a four-step process per clip.

Step 1: Storyboard the 5-second shot. The operator writes a one-paragraph description of the 5-second video — what happens, what's in frame at the start and end, what tone the clip should carry.

Step 2: Generate a reference still. Before generating video, we generate a static image (in Nano Banana Pro or similar) that represents the first frame of the intended clip. This still becomes the anchor image for the Kling generation and the single biggest lever on output quality.

Step 3: Generate 4–6 video variants. Prompted against the reference still, with motion descriptions for the 5-second arc. Select 1–2 usable clips.

Step 4: QA and post-production. Selected clips go through a review for unwanted artifacts (warped hands, physics violations, brand fidelity). Accepted clips get added to an ad with overlays (title cards, CTA, brand logo) in Descript or CapCut.

Total operator time per final ad: ~20 minutes. Four clips per hour is the sustainable pace.

The prompts that worked

Motion prompting in Kling is the lever that separates acceptable output from good. A good motion prompt is specific about what moves, how fast, and in what direction.

Works:

A child's hands carefully adjusting a small mint-colored backpack strap.
Camera: static, tight on hands. Motion: gentle, paced, finishing with
the strap slipping into place. Lighting: warm indoor, afternoon.
Duration: 5 seconds.

Doesn't work:

A kid puts on a backpack and walks away.

The first prompt constrains the motion arc — specific action, camera behavior, light, duration. The second leaves everything open, and the model picks something mediocre.

The specifics that matter most, in descending order of impact:

1. What's moving and how. "Gentle adjustment," "slow reveal," "quick turn." Generic motion descriptions produce generic motion.
2. Camera behavior. Static, slow push-in, handheld, tracking. Without specification, the model often adds unwanted camera motion.
3. What's on-frame at the start versus the end. The model fills in the middle. Specifying the start and end narrows the middle.
4. Lighting and time of day. "Golden hour, indirect" beats "well lit."
5. Duration. Kling honors 5-second and 10-second. Other durations produce either cut-offs or stretches.

What Kling does well

Product-centric close-ups. A child's hands interacting with a WondraKids product at close range is the strongest use case. The model renders products convincingly when the camera stays tight; mistakes are in the background, which is out of focus.

Atmospheric B-roll. Short atmospheric establishing shots — a product on a table, leaves moving in afternoon light, a parent's hand placing something down. These are the lowest-risk generations. Near-100% usability.

Stylized lifestyle moments. When the prompt allows some interpretive latitude ("a parent laughing at something off-screen, casual kitchen setting"), the model produces output that feels authentic, especially when the reference image is carefully chosen.

Short transitions. A hand reaching for a product, a close-up zoom on a detail, a box opening. Transitions work because the motion is simple and constrained.

What Kling doesn't do well

Full-body motion. A child walking, running, jumping — the anatomy drifts. Legs wobble. Arm articulation is uncanny. We stopped generating these at all after three failed attempts in the first week.

Multiple coordinated humans. Two people interacting naturally in frame is beyond the current model's reliability. When we tried, we got output that looked like two people in a video that had been edited together badly.

Specific product details that require accuracy. The model renders products plausibly but not precisely. A specific logo, a specific color trim, a specific seam — these may drift. For ads that depend on the exact product being visible accurately, we still shoot.

Anything where the product is the main actor. "A water bottle falls and splashes" worked once; the other four attempts violated physics in minor but obvious ways.

Extended narrative arcs. 10 seconds is the model's comfortable ceiling. Trying to tell a beginning-middle-end story in 10 seconds routinely fails; the model picks one of the three.

Performance: the actual numbers

Across the 12 Kling ads with enough spend to generate reliable data, and compared against 18 real-UGC ads in the same window, a 6 static-image ads as a further reference point:

Kling's median ROAS is ~75% of real UGC's. Its CTR is lower (viewers click through slightly less often), but its CPM is lower too (Meta seems to bid it slightly cheaper), and the volume of impressions it can sustain at CPM parity is large.

The more important number: Kling's average cost per ROAS-point was €2.29 of ad spend + €24 of production cost per ad. Real UGC was €1.99 of ad spend + €560 of production cost per ad. On a total-cost-per-ad-sold basis, Kling is dramatically cheaper at the scales we tested.

Said differently: if we're running €10,000 of ad spend in a month, the 75% ROAS penalty on Kling ads is nearly erased by not having to pay for €3,000 of UGC shoots. In a mix that uses Kling for 60% of new creative and UGC for 40%, total campaign economics improve despite the individual ROAS per ad being worse.

Where real UGC still wins

Three specific cases where we won't switch.

Testimonial-style ads. A real person saying, on camera, why they bought and love the product. Kling cannot convincingly generate this, and Meta users seem to pick up on synthetic testimonials quickly (we tested two; engagement was poor).

Brand-anchor creative. The core 4–6 ads that define the brand's voice across a quarter. These ads run for months; the small imperfections in Kling output would accumulate and erode the brand's visual identity. For anchor creative, we pay for real.

Ads where a specific, named person is the face of the brand. If the brand has a founder-personality element, that person needs to be real on camera. No amount of reference-image anchoring reproduces a specific human well enough to sustain brand presence.

For the rest — B-roll, product close-ups, atmospheric clips, the long tail of test creatives — Kling runs the show.

The cost model, reconciled

Fully-loaded cost per final Kling ad:

Compare to real UGC:

A 13:1 cost advantage to Kling per final ad. Most of Kling's economic advantage comes from eliminating per-shoot payments to creators and from compressing the briefing/review loop.

Scaling: how many ads per month is reasonable

One operator running the workflow above produces 3–4 Kling ads per hour, sustainably. Over an 8-hour day that's 24–32 ads, but the QA fatigue sets in around hour 5 and the rejection rate climbs. Realistic daily output: 18–22 ads.

We ran 34 Kling ads total across three months with one operator working on it part-time (about 6 hours per week). Scaling to full-time would support 80–100 new Kling ads per month per operator. Whether that much volume is useful depends on the ad account; at our scale, even 40/month would saturate our creative slot availability.

The ad-account rotation rule

Even with the cost advantage, we don't run Kling-only creative. The mix that's worked for us:

Roughly a third of new creative each month is now Kling. The UGC line didn't drop much in absolute terms; we reallocated the "would have been more static images" slot into Kling video. The total creative pipeline is bigger without costing more.

What Meta's algorithm actually did

The concern before we started: would Meta's ad platform penalize synthetic video? Rumors circulated in Q4 about ML-based AI-content detectors downranking generated assets. We didn't see evidence of this.

Kling ads achieved comparable delivery, comparable CPM distributions, comparable learning-phase timing, and comparable fatigue curves to human-shot UGC. The algorithm appears not to care about provenance; it cares about engagement. A boring Kling ad gets killed the same way a boring UGC ad does.

Where Kling ads underperform human UGC — CTR, ROAS — the difference is content-quality-driven, not provenance-driven. When we produce better Kling ads (better motion arcs, tighter reference anchoring, sharper briefs), the gap narrows.

I can't tell which of these ads are Kling and which are UGC unless I pull the creative-slot metadata. Meta's algorithm apparently can't either.

— our media buyer, reviewing the January 2026 numbers

Where we go from here

Three things we're adjusting for Q1–Q2 2026:

More emphasis on product close-ups. The category Kling is strongest in and we've under-invested. We should be producing 25–30 product close-up variants per month, not 8–10.

Stop using Kling for anything with fast motion or multiple humans. The rejection rate in these categories is high enough that operator time is wasted. Workflow is to generate only within Kling's reliable zone and shoot the rest.

Tighter integration with our reference library. We're pushing most brand-image work through Nano Banana Pro for reference stills; the next step is a standardized library of reference stills that Kling generations can pull from, reducing the 2–3 min/ad spent generating the reference from scratch.

The summary for a DTC brand considering this in 2026: Kling 3 is production-ready for short-form ads inside a disciplined workflow, performs at ~75% of real UGC's ROAS per impression, costs 5% of what real UGC costs fully loaded, and should be a third of your ad-creative mix. It will not replace human UGC entirely in 2026, and that's fine — mixed works better than either alone.