Write a script, pick a presenter, either your own product photo or a stock avatar, and get a talking-to-camera video ad back in minutes instead of booking a creator shoot. Test several hooks and voices on the same script before you spend on media, all from one catalog of avatar and lipsync models.
A UGC-style video ad looks like organic content: someone talking straight into a phone camera about a product, shot loose and unpolished on purpose, because that reads as trustworthy in a way a studio commercial does not. Brands used to pay creators to film dozens of these and test which hook worked, which is slow and expensive at real volume. AI collapses that loop: write the script, pick a presenter, and get a talking clip back in minutes.
The format works because it does not look like advertising. A presenter looking into the lens, mid-sentence energy, a slightly shaky first take, all of that reads as a real person's opinion rather than a brand's pitch, and that reading is worth more clicks than a polished shot ever earns on these platforms. The problem was always production: booking a creator, scheduling a shoot, and reshooting when the hook does not land, for every single variant you want to test.
Picasso IA folds the pipeline into one catalog instead of one creator relationship. Avatar and lipsync models turn a script into a talking presenter, text-to-speech models supply the voice when you are not recording your own, and the same script can run through several presenters in the time it used to take to book one shoot. The full model catalog is worth browsing first, since the right model depends on whether you are starting from a photo, a stock avatar, or a filmed clip you want dubbed.
The presenter decision drives everything downstream, and it comes down to what you are willing to show. A real product photo gives the most authentic-feeling result, because the face and setting are genuinely yours; a stock avatar scales further because nobody has to sit for a photo at all. The table below maps the common paths to the model that handles each one on Picasso IA today.
| Approach | What you provide | Model to use | Best for |
|---|---|---|---|
| Photo to talking video | One portrait photo plus a script or audio | Omni Human (ByteDance) | A single spokesperson clip built from a photo you already have |
| Full stock avatar | Just a typed script | Avatar V (HeyGen) | Running many scripts fast without ever photographing anyone |
| Photo avatar, pick a voice | One portrait photo plus a script | P Video Avatar (PrunaAI) | The same face speaking in several languages or voice styles |
| Dub a real filmed clip | An existing UGC video plus a new script | Lipsync Precision or Video Translate (HeyGen) | Localizing a creator's real footage into another market |
| Voice only, no face on screen | A script and product footage or photos | Text-to-speech voices such as ElevenLabs or MiniMax | Ads that lean on voiceover and on-screen product shots instead of a talking head |
Photo-based presenters generally read as more trustworthy, since the audience is looking at a real person tied to the product. Stock avatars trade some of that authenticity for speed across a whole campaign; run a few candidate avatars on the same script and see which one pulls the highest watch time.
Every part of the workflow depends on the script, since the model performs exactly what you type, no more and no less. The scripts that work read like something a person would actually say out loud: short sentences, a specific problem named in the first line, one claim you can back up, and a plain call to action instead of a slogan. Marketing language, exclamation points and words like revolutionary tend to break the illusion the format depends on, because a real person filming a phone selfie does not talk that way.
The single biggest lever in a UGC-style ad is the first line. Viewers decide whether to keep watching in about a second and a half, so open on the specific problem or the surprising result, never on the brand name; save the product name for the second or third sentence, once the hook has already earned the extra second of attention.
Write the script as if a friend is explaining why they bought something, not as if a copywriter is selling it. Read the line out loud before you send it to the avatar model: if it sounds like something you would say to a friend, it reads as genuine on screen; if it sounds like a billboard, rewrite it.
This is the path from an idea to an exported clip, and most people can run it end to end in under twenty minutes once the script is written. It assumes you already know which product or offer the ad is for.
The advantage over a real creator shoot is not the final clip, it is how cheap the tenth variant is compared to the first. A studio shoot makes every additional hook expensive; here the marginal cost of another version is one more generation. A few habits make that advantage count for something instead of producing a pile of similar clips.
This format is not a free pass around honesty, and it has real limits worth knowing before you build a campaign around it.
Photo-based presenters can show stiffness in the eyes and jaw on longer scripts, and the illusion holds up better in the first ten seconds than across a full minute, which is one more reason to keep scripts short. Audio-driven models such as Omni Human perform best on clips of around fifteen seconds; push much past that and lip-sync quality tends to soften. Stock avatars are limited to the library a given model ships with, so you cannot conjure an arbitrary face, only choose among the ones available. And a synthetic presenter is not a substitute for an actual customer testimonial: if your audience specifically expects proof from a real buyer, a generated presenter reading a script will not carry the same weight, no matter how natural it looks.
There is also a disclosure question that is not optional. Most platforms require you to label ad content as advertising regardless of format, and several are adding rules specifically for AI-generated presenters; treat a synthetic UGC-style ad the same way you would treat any paid endorsement, and disclose it plainly rather than letting it pass as an organic post from a real customer.
It is an advertisement made to look like organic content a real customer posted, someone talking to the camera about a product in a casual, unpolished way, rather than a scripted studio commercial. The format performs well on feed-based platforms because viewers read it as a genuine opinion first and an ad second, at least until the pitch becomes obvious. AI versions recreate that look using a talking-avatar or lipsync model instead of a hired creator.
No, not necessarily. You can upload a single portrait photo you have the rights to use, or pick a stock avatar built into a model like Avatar V, and generate a full talking clip from a typed script. Recording your own short video and dubbing it with a new script is also an option if you already have footage you like the framing of.
Start from what you have. If you have a photo of the presenter, Omni Human or P Video Avatar turns it into a talking clip paired with a script or an uploaded voice recording. If you would rather use a stock face and never photograph anyone, Avatar V generates the whole presenter from a script alone. Both sit in the same catalog, so testing one against the other costs nothing but a second generation.
Fifteen to thirty seconds is the range that performs best for this format, which works out to roughly forty to eighty spoken words depending on pacing. Shorter scripts also render faster and hold lip-sync quality better on audio-driven models, so there is a practical reason to stay brief on top of the attention-span one.
Yes, and this is one of the more useful things about generating rather than filming. Run the identical script through a photo-based presenter and a stock avatar, or through two different avatars, and compare watch time on each; because the words never change, any difference in performance comes from the presenter and the voice, which is exactly the variable you want to isolate.
Only if you have the right to use it. If a customer has given permission, either their photo or an existing short clip can become the base: a photo goes through Omni Human or P Video Avatar with a new script, and an existing clip can be relabeled with new dialogue through a lipsync-dubbing model. Never use a stranger's photo or footage pulled from social media without consent.
Not for every use case. Generated presenters are fast and cheap for testing hooks and scaling script variants, but a real creator brings an actual following, a genuine voice and, in some categories, a level of trust a synthetic presenter cannot match. Many teams use both: generated clips for rapid hook testing, real creators for the campaigns that need an authentic face people already recognize.
Yes, if it is presented as advertising, and most platforms now expect it either way. Ad-disclosure rules apply to the format regardless of how the presenter was produced, and several platforms are adding specific labels for AI-generated people in ads. Treat a synthetic UGC-style ad exactly like any other paid promotion and disclose it plainly rather than letting viewers assume it is an unpaid, organic post.
Lip-sync and facial-motion models are strongest on short, clearly lit, front-facing footage, and they soften a little as a script runs longer or the source photo is angled, low resolution or poorly lit. Keeping scripts under thirty seconds and starting from a sharp, well-lit, front-facing photo fixes most of what causes that stiffness, and choosing a newer model in the catalog when one is available also tends to help.
Cost depends on the model, the resolution and the length of the clip, so any specific figure written here would be wrong within a few months as the catalog updates. Avatar and lipsync video generations sit above single still images and below the heaviest video work in typical cost. Current numbers live on the pricing page, and new accounts start with free credits, enough to test a few scripts and presenters before spending anything.
Stop waiting on a creator's calendar for the next ad test. Open the Picasso IA toolkit, pair a script with a photo or a stock avatar, and have a testable clip back before the meeting where you were going to propose the shoot even starts.
