AI UGC ads prompt templates, model by model.
An AI UGC ads prompt works when it says what each reference image is for, then directs one shot: the framing, one action with its timing, the spoken line and the sound. Below are copy-paste templates for Seedance 2.5, Google Veo 3.1, Gemini Omni Flash and Kling 4.0, each checked against the vendor’s docs on 15 November 2026.
By Babar Ali, founder · Published 15 November 2026 · Sources checked 15 November 2026
On this page: Models · Template · Seedance · Veo · Omni · Kling · Hooks · Shot list · Failures · What not to prompt · FAQ
UGC means user-generated content: an ad made to look like a person filming on their phone. Every template below is a studio template, and every filled example is labelled: the product is a made-up water bottle and the presenter a generated person, so none of it is client work. For one model in depth, from the hooks to the finished cut, read our Seedance UGC ads guide. More guides are in Insights.
Which AI video models work for UGC ads?
Four current models with official prompt guides and generated sound, as their vendors listed them on 15 November 2026. The last row is our reading of those specs.
| Spec | Seedance 2.5 | Veo 3.1 | Gemini Omni Flash | Kling 4.0 |
|---|---|---|---|---|
| Where you run it | BytePlus ModelArk, by API or in its console[1] | Google Cloud’s Gemini Enterprise Agent Platform, in us-central1[2] | The Gemini API, through its Interactions API[3] | Kling AI’s creation workspace[4] |
| Length per generation | 4 to 30 seconds[1] | 4, 6 or 8 seconds; 8 only with reference images[2] | 3 to 15 seconds, 10 by default[3] | 3 to 30 seconds, 10 by default; 4.0 Flash up to 20[4] |
| Sound | Generated with the picture[1] | Sound generation supported[2] | Native audio, made with the picture[3] | Two-channel stereo, made with the picture[4] |
| Languages | 11 listed for speech, English among them[1] | Prompts in English[2] | English fully supported; others not yet evaluated[3] | 9 named, with American, British and Indian English accents[4] |
| A product photo as input | First frame, first and last frames, or up to 30 reference images[1] | First frame, first and last frames, or up to 3 images of one subject[2][5] | An image to animate, first and last frames, or up to 6 reference images[3] | First, last or up to 10 keyframes, or up to 10 reference images[4] |
| Aspect ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 or adaptive[1] | 16:9 or 9:16[2] | 16:9 or 9:16[3] | 21:9, 16:9, 1:1, 9:16 or auto[4] |
| Where it fits in a UGC ad | A whole draft of up to 30 seconds, hook to call to action, in one take[1] | Single shots, one moment per clip, as Google advises[6] | Fast drafts of up to 15 seconds, then edits made by conversation[3] | Whole drafts with dialogue; Kling lists finished ads among its uses[4] |
What changed this autumn
Many prompt guides that rank today were written for models that have since moved. Check the version before you copy a prompt.
- Kling 4.0 rolled out to all users on 15 October 2026[7], and Kling’s guide lists a 4.0 Flash mode for quicker 720p drafts[4].
- Veo 3.1 Preview in the Gemini API shut down on 22 October 2026. Google points developers to Gemini Omni Flash there, or to the generally available Veo 3.1 on its Gemini Enterprise Agent Platform[8].
- Sora 2: OpenAI says its Sora 2 models and Videos API shut down on 24 September 2026, with no one-to-one replacement[9], so we left it out of the templates.
What goes in an AI UGC ad prompt?
We write every shot on the same card before it goes into any model. BytePlus and Kling both start a prompt with the job of each reference asset[10][4], and all three vendors then describe the subject, action, scene, camera and sound[1][11][4], so the card keeps that order. The last two lines are ours. KEEP lists what must not change, in positive words, because Kling and BytePlus both advise describing what you want rather than what you do not[4][10]. DROP holds only the negatives a model supports.
ROLES Image 1: the presenter. Copy face, hair, clothes.
Image 2: the product. Copy shape, colour, label.
FORMAT 9:16, [length] seconds, handheld phone video.
[One real room], [one light source], [one or two props].
SHOT [Shot size], [eye level or angle], [static or one move].
ACTION 0-[n]s: [one visible action].
[n]-[end]s: [one visible action].
LINE [Who] says, in [language and accent], [tone]:
[one product fact, about 2 words per second of shot]
SOUND [Room tone], [one sound effect]. Music: [none, or describe it].
KEEP One presenter. The label reads exactly as in Image 2.
[The grip in plain words: "holds it by the base"].
DROP [Only the negatives this model supports]Two words per second of shot is our budget for a spoken line, so a 4-second shot carries about eight words. It is a rule of thumb from editing, not a vendor limit.
How each model writes the same card
The syntax each vendor documents for references, speech, sound, timing, cuts and negatives.
| Card line | Seedance 2.5 | Veo 3.1 | Gemini Omni Flash | Kling 4.0 |
|---|---|---|---|---|
| ROLES | @Image 1, @Image 2 in upload order, each with its job[1][10] | Up to 3 subject images of one person, character or product; a first frame for anything else[5] | Say in words how each image is used; tags such as <IMAGE_REF_0> are deprecated[3] | @Image1 defines ___; preserve its ___[4] |
| LINE | English: {the line}, the language named first[1] | She says: the line, a colon and no quotation marks[6] | She says: the line; Google’s advice covers Omni and Veo[6] | The line in quotation marks, in its language, with the voice described[4] |
| SOUND | <sound effect> and (music)[1] | Sound effects, ambient noise and dialogue in separate sentences[11] | Describe the audio you want, above all any music[3] | An Audio block: voice, ambience, music, effects[4] |
| ACTION timing | Whole seconds, such as 0-3s; Seedance 2.0 follows shot numbers only[10] | One focused moment per 4 to 8 second clip[6][2] | [0-5s] timecodes or plain words such as “after 3 seconds”[3] | 0-3 s segments under a sequence of events[4] |
| Cuts | Shot 1, Shot 2, or a timed transition[10] | Editing terms such as jump cut or match cut[11] | A few shots by default; write “single continuous shot” for one[3] | Multi-shot by default; name each shot and its seconds[4] |
| DROP | Only for subtitles and audio: No subtitles, No BGM[10] | A separate negative prompt that lists things, without “no” or “don’t”[11] | No negative prompt setting; write “No dialogue” or “Do not…” in the prompt[3] | Only “no subtitles” and “no background music” are needed[4] |
Or have 4 finished ads made for $600
We write the hooks, run the prompts, check every frame against the failures below and cut 9:16, 1:1 and 4:5 versions with captions and the AI label. Delivered within 10 business days, with no commitment.
Seedance 2.5 UGC prompt template
BytePlus sets the order: subject, action, scene, visual style, camera movement or shot cuts, then sound, leaving out what a shot does not need. Refer to each upload as @Image 1 and say what it provides and what it should not[1]. Seedance 2.5 can make up to 30 seconds in one pass[1], so this template writes three shots of a draft in one prompt. Our Seedance guide has the single-shot version.
@Image 1 is the presenter: copy her face, hair and clothes.
@Image 2 is the product: copy its shape, colour and label.
The presenter from @Image 1 shows the product from @Image 2 in
[one real room], handheld phone video, [the light].
Shot 1 [0-3s]: [shot size]. [The hook action].
Her line ([emotion]), in English: {[the hook line]}
Hard cut.
Shot 2 [3-7s]: [shot size]. [The demo action]. <[one sound]>
English: {[fact one]}
Hard cut.
Shot 3 [7-10s]: close-up of the product from @Image 2 alone,
label facing the camera, slow push-in.
English: {[fact two]}
One presenter. The label stays exactly as in @Image 2.
No subtitles. No BGM; only room tone and action sounds.@Image 1 is the presenter: copy her face, hair and clothes.
@Image 2 is the bottle: copy its shape, matte green finish and
label.
The presenter from @Image 1 shows the bottle from @Image 2 at a
kitchen counter, handheld phone video, soft morning window light.
Shot 1 [0-3s]: close-up. She lifts the bottle into frame and
taps the lid twice.
Her line (warm), in English: {Three things to check before you
buy a water bottle.}
Hard cut.
Shot 2 [3-7s]: medium shot. She flips the lid open with one
thumb and clicks it shut. <soft click>
English: {One, the lid opens with one thumb.}
Hard cut.
Shot 3 [7-10s]: close-up of the bottle from @Image 2 alone on
the counter, label facing the camera, slow push-in.
English: {Two, it's steel inside and out.}
Hard cut.
Shot 4 [10-13s]: medium shot. She slides the bottle into the
side pocket of a backpack.
English: {Three, it fits a backpack side pocket.}
One presenter. The label stays exactly as in @Image 2.
No subtitles. No BGM; only room tone and action sounds.What Seedance handles differently
- Dialogue and lip sync. Dialogue goes in
{ }, with the language named first for anything but Chinese[1]. Put the emotion in the line’s label, such asHer line (warm), because BytePlus says tone or action notes attached to single words can trigger subtitles[10]. - Sound. Effects go in
< >, music in( )and subtitles in【】[1]. If music still appears, BytePlus suggests listing every word for it (music, BGM, score, melody) and repeating the rule at the start and the end[10]. - Shot length and cuts. 4 to 30 seconds per generation, 4 to 15 on Seedance 2.0[1]. Mark segments with
Shot Nor whole-second timestamps with no gaps; 2.0 ignores timestamps and follows shot numbers[10]. - Aspect ratio. A first-frame task keeps the ratio of the image, so the setting must be adaptive[1]. That is how to get 4:5 for a Meta feed: any ratio between 0.4 and 2.5 can come from the input image[10].
- A product photo. Use it as a first frame, or as a reference image named in the prompt; up to 30 on Seedance 2.5 and 10 on 2.5 flash[1]. Real human faces cannot be uploaded directly: the presenter is a preset digital character, a face from a trusted model’s output, or a real person whose likeness is authorized[1].
- Negatives. Only for subtitles and audio; describe everything else positively[10].
Veo 3.1 UGC prompt template
Google breaks a prompt into subject, action, scene, camera angle and movement, lens, style, timing and audio, and says not every prompt needs every part[11]. For short clips it advises one focused moment per prompt[6], and Veo 3.1 clips are 4, 6 or 8 seconds[2], so this template is one shot. It starts from a composed first frame of the presenter holding the product and follows Google’s image-to-video advice: prompt for the motion only, and call the person “the woman” or “she”[6].
First frame: a 9:16 still of the presenter holding the product,
made with an image model, 720p or larger.
[Camera move], [shot size]. The woman [one action] and
[one small second action].
She says: [the line, with no quotation marks]
[Room tone]. [One sound effect].
Negative prompt: subtitles, captions, text overlay, background music
Duration: [4, 6 or 8] seconds. Aspect ratio: 9:16.First frame: the presenter at a kitchen counter holding the
matte green bottle at chest height, label facing the camera.
Slow handheld push-in, medium close-up. The woman lifts the
bottle toward the camera and taps the lid twice with one finger.
She says: Three things to check before you buy a water bottle.
Quiet kitchen room tone. Two soft taps on a steel lid.
Negative prompt: subtitles, captions, text overlay, background music
Duration: 6 seconds. Aspect ratio: 9:16.What Veo handles differently
- Dialogue. Write
She says: the line. Google says to use a colon and leave out quotation marks, which can make the model render the words as on-screen text[6]. Describe the audio in separate sentences[11]. - The same presenter across shots. Google’s method is a named character with a fixed description and voice, pasted unchanged into every prompt, and the same seed[6].
- Shot length and ratio. 4, 6 or 8 seconds, in 16:9 or 9:16 only[2]. Cut 1:1 and 4:5 from the 9:16 master in the edit.
- A product photo. As a first frame, use 720p or larger in 16:9 or 9:16; other sizes may be resized or cropped[12]. As subject references, up to three images of a single person, character or product, and Veo 3.1 then returns 8-second clips only[5]. That leaves two routes: compose presenter and product into one first frame, as in the template, or give Veo product photos and write the presenter in words.
- Negatives. A separate negative prompt that names what you do not want, such as “text overlay”, rather than “no text”[11].
- People. The default setting generates adults only, and Google’s filter rejects photorealistic depictions of prominent people[5][13].
Gemini Omni Flash UGC prompt template
Gemini Omni Flash is Google’s video model in the Gemini API now that the Veo 3.1 Preview there has shut down[8]. It makes 3 to 15 seconds with native audio and can edit or extend a result across turns[3]. Google asks for the scene, camera movement, lighting and mood, and for timing in plain words or timecodes[3].
Use the first image for the woman's face, hair and clothes.
Use the second image for the product: keep its shape, colour
and label exactly.
Vertical handheld phone video in [room], [light].
[0-3s] [Shot size]: [the hook action]. She says: [hook line]
[3-7s] [Shot size]: [the demo action]. She says: [fact one]
[7-10s] Close-up of the product alone, label facing the camera.
She says, off screen: [fact two]
Audio: her voice, [room tone], [one sound effect].
No background music. No subtitles or on-screen text.Use the first image for the woman's face, hair and clothes.
Use the second image for the bottle: keep its shape, matte green
finish and label exactly.
Vertical handheld phone video at a kitchen counter, soft morning
window light.
[0-3s] Close-up: she lifts the bottle into frame and taps the lid
twice. She says: Three things to check before you buy a water
bottle.
[3-7s] Medium shot: she flips the lid open with one thumb and
clicks it shut. She says: One, the lid opens with one thumb.
[7-10s] Close-up of the bottle alone on the counter, label facing
the camera, slow push-in. She says, off screen: Two, it's steel
inside and out.
[10-13s] Medium shot: she slides the bottle into the side pocket
of a backpack. She says: Three, it fits a backpack side pocket.
Audio: her voice, quiet kitchen room tone, a soft click on the
lid. No background music. No subtitles or on-screen text.What Omni handles differently
- Cuts by default. Omni tries to make a video with a few shots. For one continuous take, write “in a single continuous shot” or “no scene cuts”[3].
- Dialogue and sound. Write the line after a colon, as for Veo[6]. Omni makes an audio track by default, so say what you want, above all whether there is music[3]. Uploaded audio and voice editing are not supported[3].
- Length and ratio. Any whole second from 3 to 15, 10 by default, and 16:9 or 9:16[3]. With both a first and a last frame the limit is 10 seconds[3].
- A product photo. Up to 6 reference images, with up to 3 per character for up to 2 characters. Describe each image’s job in words and leave out tags such as
<FIRST_FRAME>[3]. Images of certain recognizable people are not supported[3]. - Negatives and text. There is no negative prompt setting: write “No dialogue” or “Do not…” in the prompt. Google says Omni renders text you ask for correctly and readably[3], and we still read every frame.
- Edits by conversation. Google says simple edit prompts work best; end with “Keep everything else the same”. Extensions add 3 to 10 seconds a turn, up to 30 seconds in all[3].
Kling 4.0 UGC prompt template
Kling’s 4.0 guide gives a four-part structure: asset roles, overall settings, a sequence of events with seconds, and audio. Its rule is concrete description with few abstract words and few negatives: “premium look” becomes the lighting, the background and the surface you would see[4]. Kling 4.0 makes 3 to 30 seconds, with multi-shot switched on by default[4].
[Asset roles]
@Image1 defines the presenter; preserve her face, hairstyle and
clothing.
@Image2 defines the product; preserve its shape, colour and
label text.
[Overall settings]
[Room, with three concrete details], [light]. Handheld vertical
phone video with small natural shake. Natural skin texture.
[Sequence of events]
0-3 s: [Shot size]. [The hook action]. She says: "[hook line]"
3-7 s: [Shot size]. [The demo action]. She says: "[fact one]"
7-10 s: Close-up of the product alone, label facing the camera.
She says, off screen: "[fact two]"
[Audio]
A woman in her [age] with a [tone] voice speaks [language and
accent] at a [pace] pace. [Ambience]. [One sound effect].
No subtitles. No background music.[Asset roles]
@Image1 defines the presenter; preserve her face, hairstyle and
clothing.
@Image2 defines the bottle; preserve its shape, matte green
finish and label text.
[Overall settings]
A kitchen counter with white tiles, a wooden chopping board and
a kettle behind her, in soft morning window light. Handheld
vertical phone video with small natural shake.
[Sequence of events]
0-3 s: Close-up. She lifts the bottle into frame and taps the lid
twice. She says: "Three things to check before you buy a water
bottle."
3-7 s: Medium shot. She flips the lid open with one thumb and
clicks it shut. She says: "One, the lid opens with one thumb."
7-10 s: Close-up of the bottle alone on the counter, label facing
the camera. She says, off screen: "Two, it's steel inside and
out."
10-13 s: Medium shot. She slides the bottle into the side pocket
of a backpack. She says: "Three, it fits a backpack side pocket."
[Audio]
A woman in her late 20s with a warm, clear voice speaks natural
American English at a relaxed pace. Quiet kitchen ambience, a
soft click on the lid. No subtitles. No background music.What Kling handles differently
- Dialogue and lip sync. Write the line in quotation marks in the language it should be spoken in, and describe the voice: age, tone, pace and emotion[4]. Kling lists nine languages and accents such as American, British and Indian English[4].
- Sound. Two-channel stereo. You can upload a voice profile, but not music or sound effects: describe those in the audio part of the prompt[4].
- Shot length and cuts. 3 to 30 seconds in 1-second steps, 10 by default; 4.0 Flash makes up to 20 seconds at 720p. Multi-shot is on by default and the model splits shots by itself, so name each shot and its seconds, or ask for one continuous shot, as Kling’s own examples do[4].
- Aspect ratio. 21:9, 16:9, 1:1, 9:16 or auto; with keyframes, the ratio follows the first keyframe[4].
- A product photo. In Omni Reference, up to 10 images, each named
@Image1,@Image2in upload order with its role stated. In Keyframes, writeUse @Image1 as the first frame. The two entry points cannot be combined in one task[4]. - Negatives and text. Kling says only “no subtitles” and “no background music” are truly needed, because the model may add them by default. It also lists text, emoji and logo rendering among 4.0’s upgrades[4].
Hook formulas, written as prompt lines
A hook is the first 1 to 3 seconds: one action and, usually, one line. Each formula below is the first ACTION line of the card and the first spoken line. On our own explainer channel the rule is an event in the first frame, in six words or fewer, which came from watching what viewers did; the video pipeline case study has the numbers.
| Hook | First prompt line (0 to 3 seconds) | First spoken line |
|---|---|---|
| The checklist | Close-up. She holds the [product] to the lens and taps it twice. | Three things to check before you buy a [category]. |
| The problem, on camera | Close-up. [The everyday problem happens in one visible action]. | If your [category] does this, look at the [part that fixes it]. |
| The question | Medium close-up. She looks up from [the task] straight into the lens. | Does your [category] still [the problem]? |
| The result first | 0-2s: extreme close-up of [the product doing its one job]. No dialogue. | From 2s: That's the [feature]. Here's how it works. |
| Side by side | She holds [the usual alternative] in one hand and the [product] in the other. | Same size as a [alternative], but [one fact you can prove]. |
| The unboxing | Overhead shot. Her hands pull the tab and lift the lid of the [box]. | Here's everything that comes in the box. |
| The silent opener | Static medium shot. She [does the task]. No dialogue. | None. Add six words or fewer of on-screen text in the edit. |
None of them has the presenter claim to have used the product or describe a result. Fill each bracket with a fact the brand can prove.
A shot list for a 15-second AI UGC ad
The labelled water-bottle ad from the templates. The first four rows are generated; the end card is built in the edit.
| Seconds | Shot | The line | Check before it ships |
|---|---|---|---|
| 0 to 3 | Hook, close-up: she lifts the bottle into frame and taps the lid twice | Three things to check before you buy a water bottle. | The label is readable when the bottle stops moving. |
| 3 to 7 | Demo, medium shot: she flips the lid open with one thumb and clicks it shut | One, the lid opens with one thumb. | Five fingers; the thumb never merges into the lid. |
| 7 to 10 | Insert, close-up: the bottle alone on the counter, slow push-in | Two, it's steel inside and out. | The label matches the packshot letter by letter. |
| 10 to 13 | Use, medium shot: she slides it into a backpack side pocket | Three, it fits a backpack side pocket. | The bottle keeps its size against the bag. |
| 13 to 15 | End card: the product, the offer and the call to action | None, or a voice-over added in the edit. | Text sits inside every placement's safe zone. |
How to split it by model
- Seedance 2.5, Kling 4.0 and Gemini Omni Flash: the first 13 seconds can be one generation, as in the labelled examples, since each allows 13 seconds or more[1][4][3]. If one shot fails, regenerate that shot alone and cut it in.
- Veo 3.1: two clips, an 8-second clip for the hook and the demo and a 6-second clip for the insert and the use, trimmed in the edit[2].
The failures to prompt against
Our checklist, with the fix each vendor documents where there is one. We run it on every take before the edit and again on the finished cut.
| Failure | Prompt against it | Check before it ships |
|---|---|---|
| Label drift | “The label stays exactly as in Image 2”, and a close-up insert made from the product photo alone. Generated text can be misspelled; BytePlus suggests spelling a word letter by letter or giving the text as an image[10]. We add any words that must be perfect in the edit. | Freeze the frame and read the label letter by letter against the packshot. |
| Hands | One hand action per shot, with the grip in plain words: “holds the bottle by its base, label facing out”. Kling advises concrete description over negatives[4]. | Five fingers, a believable grip, and no hand merging into the product. |
| Lip sync | One speaker per shot, the line in the model’s syntax, the language named. In Seedance, put the emotion in the line’s label rather than on single words[10]. | The mouth lands on every word, above all on p, b and m. |
| Stray subtitles | “No subtitles” in Seedance and Kling[10][4]; no quotation marks in Veo and Omni dialogue[6]. A reference video with subtitles can override the instruction[10]. | Scrub every frame, including the last half second. |
| Stray music | “No BGM; only room tone and action sounds”, repeated at the start and end in Seedance[10]; “No background music” in Kling[4]; in Omni, describe exactly the audio you want[3]. | Listen on headphones with the picture off. |
| Cuts you did not ask for | “In a single continuous shot” in Omni, which makes several shots by default[3]; in Kling, name each shot or ask for one continuous shot, since multi-shot is on by default[4]. | Count the cuts against the shot list. |
| The wrong face on the wrong role | Upload references in the order the people first appear and bind each one in the text, not by writing names on the image[10]. Kling says to state each asset’s role[4]. | The presenter matches Image 1 in every shot, with the same clothes. |
| Seedance artefacts | Neutral emotion words, such as “amazed” rather than “extremely shocked”, against glowing eyes; references no larger than the output resolution, against fingerprint-like textures; no water, waves or echo, against bubbling audio[10]. | Eyes in every close-up, textured areas at full size, and the audio solo. |
What not to prompt
Lines and requests we leave out of every ad, and the reason for each.
- A customer’s experience. Not “I’ve used this every day for a month” or “it changed my mornings”. The FTC’s rule bars a business from creating a consumer testimonial that materially misrepresents that the person exists, used the product or had that experience[14], and FTC staff say an AI avatar’s message can count as a testimonial, which is barred when it is fake or false[15]. Vendor showcase prompts are not written for this: Kling’s own guide has a promotional character say she honestly loves the jacket she shows[4]. In an ad styled as a customer video, we rewrite that as a product fact.
- A rating or quote that no customer gave. No generated star rating, review count or quote card on screen, for the same reason[14].
- A real person without a signed release. No real face, voice or name without written consent. BytePlus does not take real faces as direct uploads[1], Google filters photorealistic prominent people[13], Omni does not support images of certain recognizable people[3], and TikTok bars likenesses used without permission[16].
- A generated “after”. TikTok bars before-and-after results that may give a false or distorted impression[16], and a generated result is one nobody had.
- Absolute claims. TikTok does not allow absolute terms about a product in relation to time, region or brand, such as results “right away”[16].
- Another brand’s logo, product or character. Our rule: if it is not the client’s, it is not in the prompt.
- Syntax that backfires. Quotation marks around Veo and Omni dialogue[6], “no” or “don’t” inside Veo’s negative prompt[11], abstract words such as “cinematic feel” in Kling[4], and tags such as
<IMAGE_REF_0>in Omni[3].
The labels and testimonial rules
As the platforms and regulators stated them on 15 November 2026. This is not legal advice: rules differ by country and category.
- TikTok: label it. Use TikTok’s AIGC label or your own clear disclaimer, caption, watermark or sticker; an undisclosed AI ad is rejected or restricted[16]. In Ads Manager, switch on “This ad contains AI-generated content”; Spark ads follow the community guidelines instead[17].
- Meta: assume it can tell. Meta adds an “AI info” label to ads made or significantly edited with its own AI tools, and says it will detect ads made with other AI tools through industry-standard signals[18]. Every Omni video carries an invisible SynthID watermark[3], and Veo 3.1 supports Content Credentials[2].
- The FTC: a presenter is not a customer. 16 CFR Part 465 covers fake testimonials[14]; AI avatars are not banned outright, but a fake or false avatar testimonial is[15]. Our presenters demonstrate and explain.
New York’s synthetic performer law and the likeness rules are in the rules section of our Seedance guide.
What does a finished AI UGC ad cost?
Creators, self-serve tools and a studio, priced side by side from their published rates, with what each price leaves out.
Not for you if you would rather pick an avatar than write prompts: our Arcads and Creatify comparisons say where a self-serve tool is the better buy. Our studio prices are on the pricing page, and the AI content service covers the other formats.
Free download · PDF, 4 pages · no email needed
AI UGC ad prompt templates
The one-shot card, the syntax for each model, the Seedance 2.5, Veo 3.1, Gemini Omni Flash and Kling 4.0 templates, the hook lines and the failure checklist from this guide.
Download itQuestions about AI UGC ad prompts
Say what each reference image is for, then direct one shot: the framing, one action with its timing, the spoken line and the sound. Add what must not change, such as the label, and only the negatives the model supports. Write the dialogue in the model's own syntax.
It depends on the shot. Seedance 2.5 and Kling 4.0 make up to 30 seconds with sound in one generation, Gemini Omni Flash 3 to 15 seconds, and Veo 3.1 clips of 4, 6 or 8 seconds. A long model suits a whole draft, and a short clip suits a single shot that has to be exact.
Veo 3.1 is generally available on Google Cloud's Gemini Enterprise Agent Platform, but the Veo 3.1 Preview in the Gemini API shut down on 22 October 2026, and Google points developers there to Gemini Omni Flash. OpenAI says its Sora 2 models and Videos API shut down on 24 September 2026. Check that a guide's model still exists before you copy its prompts.
Use the model's own syntax: curly brackets after the language in Seedance, a colon with no quotation marks in Veo and Omni, and quotation marks in Kling. Keep one speaker per shot and a line short enough to say in the shot. Then check the mouth word by word, above all on p, b and m.
Give the model a sharp product photo as a reference and say that the label must stay exactly as in that image. Make the close-up insert from the product photo alone, and add any words that must be perfect in the edit. Freeze the frame and read the label before the ad ships.
Some models add them by default: Kling's guide says to exclude subtitles and background music when you do not want them, and BytePlus lists both as known Seedance issues. Write 'No subtitles' and 'No BGM' where the model supports negatives, and in Veo avoid quotation marks, which can make it render the words as text.
No. The FTC's rule bars a business from creating a testimonial that materially misrepresents that the person exists or used the product, and FTC staff say an avatar testimonial is covered when it is fake or false. Script the presenter to demonstrate and state product facts instead.
On TikTok, yes: use its AIGC label or your own clear disclaimer, or the ad can be rejected or restricted. Meta adds an AI info label to ads made with its own tools and says it will detect ads made with other AI tools. Plan for the label rather than against it.
Our trial is 4 finished ads for $600 and our pilot 12 for $1,800, each delivered within 10 business days. That covers the scripts, the takes we discard, the edit, the checks and the labels. Our UGC ad cost page sets it against creators and self-serve tools.
Sources
Every outside fact on this page comes from the pages below, read on 15 November 2026. Our own prices are the ones published on our pricing page.
- BytePlus ModelArk documentation: Dreamina Seedance 2.5 series tutorial (updated 15 November 2026). docs.byteplus.com/en/docs/modelark/seedance-2-5. Read 15 November 2026.
- Google Cloud documentation: Veo 3.1 on Gemini Enterprise Agent Platform (updated 12 November 2026). docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate. Read 15 November 2026.
- Google AI Studio: Generate and edit videos with Gemini Omni Flash. aistudio.google.com/docs/omni. Read 15 November 2026.
- Kling AI: Kling 4.0 User Guide (4 November 2026). kling.ai/quickstart/kling-4-model-user-guide. Read 15 November 2026.
- Google Cloud documentation: Generate videos from references. docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/generate-videos-from-references. Read 15 November 2026.
- Google Cloud documentation: Best practices for Gemini Omni and Veo. docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/best-practice. Read 15 November 2026.
- Kling AI blog: Kling 4.0 at Work: Bringing Creative Vision into Production (27 October 2026). kling.ai/blog/kling4-work-creative-vision-production. Read 15 November 2026.
- Google AI Studio: Video generation in the Gemini API. aistudio.google.com/docs/video. Read 15 November 2026.
- OpenAI API documentation: Video generation with Sora. developers.openai.com/api/docs/guides/video-generation. Read 15 November 2026.
- BytePlus ModelArk documentation: Dreamina Seedance 2.5 series prompt guide, with its FAQ of known issues (updated 15 November 2026). docs.byteplus.com/en/docs/modelark/seedance-2-5-prompt-guide. Read 15 November 2026.
- Google Cloud documentation: Video generation prompt guide for Gemini Omni Flash and Veo. docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide. Read 15 November 2026.
- Google Cloud documentation: Generate videos from an image. docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/generate-videos-from-an-image. Read 15 November 2026.
- Google Cloud documentation: Responsible AI and usage guidelines for Veo. docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/responsible-ai-and-usage-guidelines. Read 15 November 2026.
- eCFR: 16 CFR Part 465, Rule on the Use of Consumer Reviews and Testimonials. www.ecfr.gov/current/title-16/chapter-I/subchapter-D/part-465. Read 15 November 2026.
- Federal Trade Commission: Consumer Reviews and Testimonials Rule, questions and answers (November 2024). www.ftc.gov/business-guidance/resources/consumer-reviews-testimonials-rule-questions-answers. Read 15 November 2026.
- TikTok Advertising Policies: Misleading and false content (updated November 2026). ads.tiktok.com/help/article/tiktok-ads-policy-misleading-and-false-content. Read 15 November 2026.
- TikTok Ads Manager help: Add disclaimers to ads (updated September 2025). ads.tiktok.com/help/article/add-disclaimers-to-ads. Read 15 November 2026.
- Meta: Expanding GenAI transparency for Meta's ads products (February 2025, updated 1 June 2026). about.fb.com/news/2025/02/gen-ai-transparency-metas-ads-products/. Read 15 November 2026.
Keep reading
Rather have the ads than the prompts?
A trial is 4 finished ads for $600 and the pilot 12 for $1,800, each delivered within 10 business days, with no commitment. Send a product link and the ads you run now, and we reply within one business day with the hooks we would test first.