← Back to blog
use case·15 min read·June 3, 2026

How to Make a UGC Ad with AI (Without Filming)

The exact 4-step AI UGC workflow for 2026: model picks per step, hook formulas that convert, and the cost math that makes AI cheaper than one creator test.

Here is the exact 4-step workflow for producing UGC-style ads with AI in 2026, the model picks for each step, and the iteration math that makes this cheaper than hiring creators. You will need a reference image for your avatar, your product images or brief, and about 30 minutes the first time through. After that, each new variant takes under 10 minutes.

TL;DR

Each variant costs $5 to $15 in model credits. You can test 10 variants for less than one creator retainer video.

Why AI UGC works now (and didn't two years ago)

The question used to be whether AI faces were convincing enough. In 2024 they weren't. Faces smeared on movement, eyes tracked wrong, and any viewer who spent time on TikTok could spot it in two frames. That threshold crossed sometime in early 2026 with Higgsfield Soul 2.0 and the identity-locking conditioning that came with it. You feed one reference portrait and the model holds that person's face, skin tone, and micro-expressions across multiple cuts and motion conditions. It's not perfect, but it's past the bar that matters: does the viewer disengage because something looks wrong? For the majority of DTC feed placements the answer is now no.

The second shift was vertical-native generation. Every model worth using in 2026 renders 9:16 natively. You're not cropping a landscape clip and wondering what the algorithm does to it. Seedance 2.0 was specifically trained on short-form vertical content, which shows: the motion curves, the depth of field choices, and the pacing all feel like something shot on a phone in good light rather than resized footage from somewhere else.

The third shift is what makes the economics click. AI UGC isn't better than creator UGC. For hero brand moments, creators still win. But AI UGC is fast enough and cheap enough to run the volume of tests that performance marketing actually requires. A DTC brand needs 20 to 40 unique creatives per quarter for a healthy feed. Creators at $300 to $800 per video can't scale that. AI can.

The 4-step workflow

Step 1: Define the hook

The first 3 seconds are the only seconds that matter for the algorithm decision. Scroll-stopping on TikTok and Reels happens at frame 1 or it doesn't happen. Your hook is not a headline. It's a visual and audio event that makes stopping feel like the correct choice.

Two formats that consistently work in 2026:

Problem-solution hook. Open on the problem, stated bluntly by your avatar. "My skin was breaking out every single week until I figured out what was actually causing it." The viewer who has this problem stops. Everyone else keeps scrolling. That's fine.

Pattern interrupt hook. Open with something visually unexpected: a close-up of the product in an unusual context, a reaction face, a split-screen showing before/after with no buildup. No words for 1 to 2 seconds, then the statement. The silence and the visual do the first work.

Write the hook before you touch any model. It determines the shot structure and the avatar's emotional state in the talking head. A confused avatar delivering a confident hook reads as inauthentic even if the generation is technically good.

For the rest of this guide, we'll walk through a concrete example: a 15-second ad for a DTC skincare brand promoting a vitamin C serum.

Example concept brief:

Hook script (first 3 seconds): Avatar looks directly at camera, slight frustration in expression: "My dark spots were getting worse, not better."

That's the brief. Now generate.

Step 2: Generate the avatar and talking head

Model: Higgsfield Soul 2.0

Upload one reference portrait of your avatar. This can be a stock photo, a photo of a real person (with rights), or a face you've generated with a still-image model like Nano Banana Pro. The reference photo should be front-facing, neutral expression, good lighting. Higgsfield locks identity from this reference and applies it consistently across clips.

Prompt structure that works:

[Avatar description] looks directly at camera in a softly lit bathroom, says "[your hook line]" with a slightly frustrated but relatable expression. Vertical 9:16. Natural light from window left. No music. Clean audio. UGC style, handheld feel.

For our vitamin C serum example:

Young woman with light brown skin, natural wavy hair pulled back, looks directly at camera in a softly lit bathroom, says "My dark spots were getting worse, not better." with a frustrated but relatable expression. Vertical 9:16. Natural light from window left. No music. Clean audio. UGC style, handheld feel, slight natural camera shake.

Generation time: about 75 to 90 seconds per clip. Generate 3 to 5 variants of the talking head with slightly different expressions or head positions. You'll pick the most natural one in assembly. The identity stays consistent across all of them because Higgsfield is holding the reference.

For the rest of the script (10 to 12 seconds of talking head): generate the remaining audio in the same session, using the same reference image. Higgsfield's multi-clip identity conditioning keeps the avatar looking like the same person across all clips. If you switch sessions, re-upload the same reference portrait.

Step 3: Generate b-roll and product shots

Models: Seedance 2.0 (motion clips) and Nano Banana Pro (product stills)

B-roll does three things in a UGC ad: breaks up the talking head so it doesn't feel like a static testimonial, shows the product in context, and visually proves the claim. For the vitamin C serum, you need: the product itself (bottle close-up, texture, maybe the dropper), and skin transformation shots (close-up of glowing skin, morning light on face).

Product motion with Seedance 2.0:

Upload the actual product photo as a reference image. This is Seedance 2.0's multi-reference conditioning doing its best work: the model builds the shot around the product you give it, rather than inventing a generic serum bottle.

Close-up of a vitamin C serum bottle on a white marble surface. A single drop falls from the dropper in slow motion. Warm natural light, golden hour. Vertical 9:16. Product stays in full-focus center frame. 5 seconds.

Generate 2 variants. One with the drop falling, one with just the bottle rotating gently. Cost: about $0.50 to $0.65 per clip.

Skin texture and glow shots with Seedance 2.0:

Close-up of a woman's cheek and jawline with visibly even, glowing skin. Morning light, soft focus background. Warm tone. Vertical 9:16. 3 seconds. No face shown, chin up angle.

This avoids the face-consistency problem (you don't need it to match your avatar) while delivering the visual proof the product claim needs.

Product stills with Nano Banana Pro:

For any frame where you want a static product image (end card, overlay shot, pause moment), generate it in Nano Banana Pro. The model handles product rendering well, especially reflective surfaces and dropper bottles. Upload the product reference, prompt for the shot you need:

Vitamin C serum bottle on a cool white background, studio lighting, slight shadow, hero product shot. Clean and premium. No text.

Nano Banana Pro generates at 2K resolution in about 15 to 20 seconds. Use these as end-card backgrounds or freeze-frame inserts in the edit.

Step 4: Edit and assemble

Tools: 8frame Studio or any NLE (CapCut, Premiere, DaVinci)

You now have: 3 to 5 talking head clips (pick the best 2 or 3), 2 product motion clips, 2 skin texture clips, and 1 to 2 product stills. Total raw footage: 15 to 20 clips for one 15-second ad.

Assembly structure for a 15-second UGC ad:

Seconds Shot Content
0 to 3 Talking head, hook "My dark spots were getting worse, not better."
3 to 5 Product shot Serum drop falling, product close-up
5 to 9 Talking head, solution setup "Then I found out what most serums are missing..."
9 to 12 Skin texture b-roll Glowing cheek, morning light
12 to 15 Talking head, CTA "This vitamin C changed my skin in 7 days. Link in bio."

Auto-captions are not optional. 85% of TikTok plays and roughly 60% of Reels plays are sound-off. If your talking head isn't captioned, you're losing the majority of your audience before they know what you're selling. Every platform's native editor has auto-captions. Use them. Style them in white with a thin black outline: maximum legibility at all scroll speeds.

Color matching. Your talking head, b-roll, and product shots will have slightly different color temperatures. In 8frame Studio you can apply a LUT to all clips in a single step. Keep it warm and soft for skincare: a faint orange-green grade reads as "golden hour organic" to the algorithm and to human eyes.

8frame's workflow templates on /workflows include a pre-built UGC assembly template with the cut timing, caption style, and color grade already set. Clone it, drop your clips into the bins, and export. The first time you use it saves about 20 minutes of setup.

The hook formulas that actually work in 2026

Hook writing is the highest-impact skill in UGC production. A mediocre clip with a great hook outperforms a beautiful clip with a weak opening every time. Here are 10 specific hooks we've tested across DTC verticals. Pull the structure, swap in your product and pain point.

1. Blunt problem statement "My [skin/hair/gut/energy] was getting worse every month until I stopped doing one thing."

2. Embarrassment interrupt "I was embarrassed to wear [shorts/sleeveless shirts/my skin bare] until I tried this."

3. Numbers as proof "I tested [14 serums/8 supplements/6 protein powders] in 60 days. Here's the one that actually worked."

4. Insider framing "Nobody told me that [the real reason you're breaking out/your gut issues aren't food] is actually [root cause]. Here's what to do about it."

5. Unexpected result "I ordered this as a last resort and now I buy it every month."

6. Direct call-out "If you have [dark spots/dull skin/under-eye bags], stop scrolling."

7. The reversal "I used to spend $200 a month on skincare. Now I spend $48 and my skin looks better."

8. Social proof as hook "My dermatologist actually asked me what I was using on my skin. It's a $48 serum."

9. Myth-bust "Vitamin C serums don't fade dark spots. Unless you're using this form of it."

10. Vulnerable admit "I'm going to show you what my skin looked like 30 days ago because I think you need to see it."

The pattern across all 10: specific, immediate, and directed at one person who has the problem. "You" is in the structure even when it's not in the words.

Picking models by use case

Not every UGC format needs the same setup. Here's the routing logic we use.

Talking head testimonial: Higgsfield Soul 2.0, always. It's the only model with identity locking strong enough to cut across 4 to 6 clips without the avatar's face drifting. Pair with Nano Banana Pro to generate the reference portrait if you don't have one.

Unboxing format: Kling 3.0 for the motion clips. Unboxing is mostly about hands and product, not faces, so identity locking isn't the constraint. Kling's speed (roughly 60 seconds per clip) makes it practical to generate 10 variants of "hand opens box, lifts product out" until one looks natural.

Product demo (application, pour, texture): Seedance 2.0. Multi-reference conditioning means you can upload the actual product and get the model to animate it correctly. Product demos live or die on whether the product looks real. Seedance is the right tool.

Problem-solution format with before/after: Seedance 2.0 for the skin/hair/product shots, Higgsfield for any talking head segments. Cut between them in the edit. The models don't need to match perfectly because the format expects a visual jump between "before" and "after."

Lifestyle b-roll (product in use, daily routine context): Kling 3.0 or Seedance 2.0. Kling for speed and volume, Seedance if the product is in frame and needs to look accurate.

For a deeper breakdown of which models to reach for across all video types, see the best AI video generator 2026 guide.

Iteration math

This is where the business case closes.

Traditional UGC creator workflow:

AI UGC workflow:

The math isn't close. The argument against AI UGC is conversion rate, not cost (more on that in the FAQ). The counterargument is that with 40 tests instead of 4, you find the 2 or 3 creatives that convert well enough to spend behind. You don't need every creative to convert. You need enough volume to find the ones that do.

Real numbers from a DTC skincare brand running this workflow in Q1 2026: 38 AI UGC variants produced in 11 days at a total model cost of $290. The top 3 performers were identified by day 18 via Meta's creative ranking. The best-performing creative had a hook-to-click rate of 3.4%, against a category average of 1.8% for creator content from the same brand. The avatar was fully AI-generated.

Common pitfalls

Uncanny face micro-expressions. Happens when the reference image has an extreme or unusual expression, or when you're asking the model to deliver a line that doesn't match the emotion in the reference portrait. Fix: use a neutral or slightly warm expression in the reference, then direct the emotion through the prompt. If a clip looks off, don't use it. Generate 5 variants, not 1.

Jarring cuts between AI shots. Your talking head was generated with warm natural light and your product b-roll was generated with cool studio light. They cut together badly. Fix: specify the same lighting condition in every prompt, or apply a single color grade in post that pulls everything into the same temperature range. 8frame's UGC workflow template includes a "warm organic" LUT that handles this automatically.

Wrong aspect ratio. Generating 16:9 and cropping to 9:16 loses the top and bottom of every frame and often cuts off the speaker's head. Every prompt in this guide specifies "Vertical 9:16." Don't skip it.

Generic stock-feeling output. Comes from generic prompts. "Woman holding product, smiling, white background" generates exactly what that sounds like. Fix: add specific context details (location, lighting source, emotional state, slight imperfections). "Woman in a dim bathroom at 7am, just woken up, holding serum bottle, half-skeptical expression" reads as real because it has specific texture.

No captions. You've read this twice now. Still the most common thing we see people skip.

FAQ

Is AI UGC against TikTok policy?

No, as of June 2026. TikTok's policy requires disclosure of AI-generated content through their "AI-generated content" label, which you apply during upload. This is a disclosure requirement, not a ban. The content can run as a paid ad with the label in place. Meta has the same disclosure requirement for AI-generated visuals in ads. Neither platform bans AI UGC; both require the label. Check each platform's current policy before running a campaign, as these update periodically.

Does AI UGC convert as well as creator UGC?

On average, no. Creator UGC from a real person with an existing audience or strong authenticity cues converts higher per impression. But AI UGC converts well enough when the hook is strong, and you can run 10x the volume at 5% of the cost. The brands getting the most out of AI UGC are using it for top-of-funnel testing and iteration, then putting ad spend behind the winners (which are sometimes AI, sometimes creator, depending on the creative). It's a testing tool first, a production tool second.

What is the best AI model for talking heads?

Higgsfield Soul 2.0, by a clear margin in 2026. The identity-locking feature is what separates it from the field. Kling 3.0 can produce a decent talking head but the face drifts across multiple cuts. Seedance 2.0 is better for product and motion than for character work. If your ad format requires a consistent avatar delivering 3 to 4 lines across 4 to 6 cuts, Higgsfield is the only model where that works reliably right now.

How do I keep the avatar consistent across cuts?

Use the same reference image in every Higgsfield generation session. Don't change the reference between the hook clip and the CTA clip. If you're generating across multiple sessions, save the reference image and reload it every time. The model's identity conditioning is anchored to that image, so as long as the image stays the same, the face stays consistent. Also: generate the same scene with slightly different camera angles rather than dramatically different setups. A medium shot and a close-up of the same avatar cut more naturally than two medium shots from different angles.

How long does it take to produce one AI UGC ad?

First time: 30 to 60 minutes including concept, generation, and assembly. After you've done it once and have a workflow template: 15 to 30 minutes per variant. Running parallel generations (talking head, b-roll, and product shots simultaneously in separate 8frame canvas tabs) cuts the wait time significantly. The generation itself is not the bottleneck. Hook writing and clip selection are.

Build the first one this week

The workflow is: hook, avatar, b-roll, assemble. Each step has a specific model, a specific prompt structure, and a generation cost between $0.30 and $1.20. A full 15-second ad costs $5 to $15 in model credits and takes under an hour the first time.

The testing math means the right way to think about AI UGC is not "can this replace creators" but "can this get me to 40 test creatives before I commit $10,000 to production." The answer is yes, and the DTC brands figuring that out this year are running cleaner ad accounts than the ones still booking one creator a month and hoping it works.

Clone the UGC assembly template on 8frame's workflow library, load your product reference, and run Step 2 first. The avatar is the hardest part. Once you have a talking head you're happy with, the rest of the workflow follows quickly.

Related articles

use caseHow to Make a Shopify Product Video with AIuse caseHow to Make an App Promo Video with AIuse caseHow to Make a Coffee Ad with AI

Your frames start here

Watch the canvas power your creative flow in real time

Stay in the loop

Be the first to hear about our launch and get product updates