How to Make a Reaction-Style Video Ad with AI
The 4-step workflow for AI reaction ads: the split-screen psychology, prompting believable facial expressions, the package-reaction sub-format, and how to keep it inside platform policy.
The reaction ad borrows a mechanic from the entire internet: watching someone else react is more engaging than watching the thing itself. Done with AI, you end up with a 15-to-25-second clip where a person reacts to a product, a result, or an unboxing, usually in split screen so the viewer sees both the trigger and the face at once. This guide is the 4-step workflow: the split-screen psychology, prompting facial expressions that don't read as fake, the high-performing "receiving a package" sub-format, and how to stay inside platform policy. Each variant costs $6 to $16 in credits and ships same-day.
TL;DR
- Step 1: Pick the trigger, the thing being reacted to (a result, a package, a demo). The reaction is only as good as what causes it.
- Step 2: Generate the reacting face with Higgsfield Soul 2.0, locked to one reference, with the emotion set explicitly in the prompt.
- Step 3: Generate the trigger content with Seedance 2.0 (product/package) or Kling 3.0 (fast filler).
- Step 4: Compose as split screen, sync the reaction to the trigger beat, caption, and disclose.
Why reaction ads work
A reaction ad outsources persuasion to a proxy. Instead of the brand claiming the product is impressive, a stand-in viewer reacts as though it is, and the audience mirrors that reaction. The split screen format is what makes it work: showing the trigger and the face simultaneously removes the gap where a skeptical viewer would otherwise disengage. The eye ping-pongs between "what happened" and "how they feel about it," and that loop is sticky.
The reason this was hard for AI until recently is that reactions are all face. A drifting identity or a dead-eyed expression kills the format instantly, because the face is the ad. Identity locking plus explicit emotion prompting is what makes believable synthetic reactions possible in 2026.
The 4-step workflow
Step 1: Choose the trigger
No model yet. The trigger is the causal event: a before/after result, a package arriving, a first taste, a price reveal. The reaction has to be proportional to it. An over-the-top gasp at a mildly nice product reads as fake; a genuine "oh, that's actually good" at a real result reads as true.
For this guide: a home fragrance brand, trigger is opening the box and the first scent.
Match the reaction size to the trigger honestly. This is also the policy line, cover it below.
Step 2: Generate the reacting face
Model: Higgsfield Soul 2.0
Upload one front-facing reference. The whole game here is the emotion word. Vague prompts produce a spokesperson smile; specific emotion words produce a real reaction. Generate the reaction in beats: anticipation, then the hit, then the settle.
Woman in her late 20s sitting on a couch in warm lamp light, holding a
shipping box, looks down as she opens it, face shifting from mild curiosity
to genuine surprise, then a soft "oh wow" and a real smile. Vertical 9:16,
handheld feel, natural home lighting, clean audio, slight camera shake.
Generate five variants and cut the ones where the expression over-sells. The believable one usually has a smaller reaction than you expect. Set the emotion explicitly ("genuine surprise, then a soft smile"), not just "excited." For the deeper mechanics of matching expression to line, see the full AI UGC workflow.
Step 3: Generate the trigger content
Models: Seedance 2.0 (product/package), Kling 3.0 (filler)
The package sub-format is the workhorse. Generate the box and the reveal with Seedance 2.0, uploading the real product so the packaging stays accurate:
[Product reference] home fragrance box being opened on a couch, hands lift
the lid and remove the bottle, warm lamp light, vertical 9:16, product
stays in focus, 4 seconds, authentic handheld framing.
For fast filler (hands, the label, the room), Kling 3.0 at roughly 60 seconds per clip lets you generate several cheaply. The trigger clip is what fills the other half of the split screen, so it has to be clean and readable at half-frame size.
Step 4: Compose the split screen
Tools: 8frame Studio or any NLE
Stack the reaction and the trigger. The sync is everything: the surprise on the face has to hit the exact frame the bottle comes out of the box, or the causal link breaks. A 20-second structure:
| Seconds | Top frame (reaction) | Bottom frame (trigger) |
|---|---|---|
| 0 to 3 | Anticipation, opening the box | Box, lid coming off (hook) |
| 3 to 8 | The surprise hit | Bottle revealed |
| 8 to 14 | First scent, settle | Close-up of the product |
| 14 to 20 | Verdict to camera | Product still + CTA |
Caption the verdict line. Then disclose: this is a synthetic person reacting, so the AI-generated label goes on at upload, and the framing must not present the reaction as a real, unpaid customer's genuine response. The disclosure guide has the specifics.
Cost math
Filmed reaction shoot:
- Talent + product + location: $300 to $900
- Editor for the split-screen sync: $150 to $400
- Turnaround: 3 to 10 days
AI reaction workflow:
- Higgsfield reaction across beats: $4 to $10
- Seedance package/product motion: $2 to $6
- Kling filler: $1 to $3
- Turnaround: same day
The advantage is variant volume: generate five reaction intensities and five triggers, then let the data pick the pairing that lands. Route the winner into real production if the numbers justify it, per the AI UGC vs real creators playbook.
Common pitfalls
Over-acted reactions. The fastest tell. A gasp too big for the trigger reads as staged. Generate several and keep the smallest believable one.
Sync drift. If the face reacts before or after the trigger beat, the causal illusion breaks. Line the surprise frame to the reveal frame precisely.
Dead eyes. Comes from a vague emotion prompt. Set the specific arc ("curiosity to surprise to a soft smile") in the prompt, not just "happy."
Policy framing. Do not present a synthetic reaction as a real customer's unsolicited response. Keep it clearly as branded creative and label it.
FAQ
How do I prompt a realistic facial expression in AI video?
Set the emotion as a specific arc, not a single adjective. "Face shifting from mild curiosity to genuine surprise, then a soft smile" produces a believable reaction; "excited" produces a spokesperson grin. Generate five variants and keep the one with the smallest, most proportional reaction, because over-sized expressions are the most common tell. Lock identity with a single reference image so the face stays consistent across the anticipation, hit, and settle beats.
Is the split-screen reaction format allowed on TikTok and Meta?
Yes, the format itself is allowed. What both platforms require is that AI-generated content carries their AI label at upload, and that you do not present a synthetic person as a real, unpaid customer giving a genuine reaction. Keep the creative clearly framed as an ad, avoid implying spontaneous real testimony, and apply the label. See our disclosure guide for the details.
What is the package-reaction sub-format?
It is the highest-performing reaction variant: someone reacts to opening a shipping box and revealing the product. It works because the box builds two seconds of anticipation before the payoff, and the reveal gives the reaction a clear, honest trigger. Generate the package opening with Seedance 2.0 using the real product as reference so the packaging stays accurate, and sync the surprise on the face to the exact reveal frame.
Pick a trigger, load one reference portrait, and run Step 2 first, the reaction is the whole ad. Clone a UGC template on 8frame's workflow library, or build it on the canvas with every model above side by side.