How to Make a Beauty Ad with AI
4-step AI workflow for beauty ads: model or talent via Higgsfield, product stills via Seedream 5.0, application motion via Seedance 2.0, captions. $20 in compute.
You can make a beauty ad with AI using three models: Higgsfield Soul 2.0 for talking-head talent segments, Seedream 5.0 for product stills, and Seedance 2.0 for application motion and skin texture clips. A 15-second skincare ad across four demographic variants costs roughly $20 in model credits on 8frame. A production studio charges $3,000 to $8,000 for the same four-variant test, with two to four weeks turnaround.
TL;DR
- Step 1: Generate the model or talent segment with Higgsfield Soul 2.0, anchored to one reference portrait for identity consistency across cuts
- Step 2: Generate product stills with Seedream 5.0, uploading your actual product as a reference so the bottle and packaging render accurately
- Step 3: Generate application motion and skin texture b-roll with Seedance 2.0 (close-up of product touching skin, dropper shot, post-application glow)
- Step 4: Assemble, add auto-captions, apply a warm-soft LUT, export 9:16 for TikTok and Reels
- Four demographic variants of a 15-second vitamin C serum ad: $18 to $22 total compute
Beauty ad formats and model routing
Five main formats, each with a different model emphasis.
Skincare routine: Talking head plus close-up application footage. Lead with the routine context (morning or evening, step order). Higgsfield for face segments, Seedance for application texture and skin close-ups.
Makeup tutorial: More visual than verbal. Seedance 2.0 is the workhorse for brush strokes, blending, and lip application. The talking head is secondary to the transformation shots.
Fragrance mood: Fragrance can't show efficacy, so the ad sells atmosphere. Seedream 5.0 for hero stills (bottle in evocative lighting, fabric, skin). Seedance 2.0 for any motion element. No talking head needed.
Dermatologist authority: Credentialed spokesperson format. Higgsfield for all talent segments, directed toward a clinical-adjacent setting. See pitfalls on title authority before running this format.
Before and after: Highest-converting in beauty, most legally constrained. Seedance 2.0 for skin texture clips, Seedream 5.0 for still comparison panels. Read FTC compliance notes in pitfalls before publishing.
The 4-step workflow
Step 1: Model and talent via Higgsfield Soul 2.0
Upload one reference portrait: front-facing, neutral expression, even lighting. Higgsfield locks identity from this image and holds it across every clip. If you don't have a reference, generate one with Seedream 5.0 first, then use it as the Higgsfield input.
Prompt structure:
[Avatar description] in a [environment], looks directly at camera, says "[line]" with [emotional state]. Vertical 9:16. [Lighting description]. No music. Clean audio. Handheld feel, slight natural camera shake.
Tested prompt, skincare hook:
Woman in her early 30s with medium-deep brown skin and natural hair, softly lit bathroom with morning window light from the left, looks directly at camera, says "My skin was getting duller every month until I changed one thing in my routine." Slight concern shifting to confidence. Vertical 9:16. Clean audio, no music. UGC feel, slight natural movement.
Generated in 78 seconds on 8frame. The identity held cleanly across three variants: expression shifted naturally between takes while the face stayed consistent. The window-left lighting produced a warm, credible bathroom context without any studio artifice.
Tested prompt, dermatologist authority format:
Woman in her mid-40s with light olive skin and dark hair pulled back, clean white studio background, wearing a white lab coat, looks directly at camera, says "Most vitamin C serums break down before they reach the skin layer that matters." Calm authority, no urgency. Vertical 9:16. Soft even lighting, no shadows. Clean audio.
Generated in 82 seconds. The lab coat and neutral background produced a clear authority signal. Don't generate name badges or title cards in the clip; add those in post as text overlays you control for compliance.
Step 2: Product still via Seedream 5.0
Seedream 5.0 generates the highest-fidelity product stills on 8frame. Color accuracy on lipstick, foundation, and packaging is where generic image models fail; Seedream holds the specific product color when you give it explicit direction. Upload your actual product photo as a reference.
Prompt structure:
[Product description] on [surface], [lighting description], [angle or composition]. No text on image. Clean background. Hero product shot.
Tested prompt, serum bottle:
A vitamin C brightening serum bottle, amber glass dropper, silver cap, on a cool white marble surface with soft diffused studio lighting. Three-quarter angle. Slight shadow beneath the bottle. No text. Hero product shot. Clean and precise.
Generated in 22 seconds at 2K resolution. The amber glass rendered with accurate internal refraction. The silver cap held its color without shifting warm or cool. This is the shot that fails in generic models: Seedream held the specific packaging color rather than drifting toward a generic serum archetype.
Tested prompt, lipstick end-card:
A matte red lipstick bullet, retracted, on a white satin surface, close-up. Soft diffused light from above. The red tone is deep burgundy-red, not orange-red. Precise color, no overexposure. Clean white background. Hero product shot. No reflections on lens.
Generated in 19 seconds. The specific color direction ("burgundy-red, not orange-red") held in the output. This prompt precision is necessary for any category where the shade is the product. Without it, Seedream defaults to the most statistically common interpretation of the color.
Step 3: Application motion via Seedance 2.0
Seedance 2.0 handles skin-contact motion better than any other model on 8frame. Upload the product reference for any clip where the packaging appears in frame.
Tested prompt, serum application:
Close-up of fingertips pressing a small drop of golden serum into smooth skin on a cheek. The serum absorbs slowly, catching warm light. Skin texture is detailed but even-toned. No face shown, chin-up angle. Vertical 9:16. Warm natural light. 4 seconds.
Generated in 88 seconds. The serum drop motion was clean and the absorption effect read as natural, not CGI. Skin texture rendered with realistic pore detail without tipping into the uncanny valley. This is the hardest shot type in beauty production: skin close-up without texture problems.
Tested prompt, dropper product motion:
A single drop of amber serum falling in slow motion from a glass dropper onto a white marble surface. The drop impacts and spreads slowly. Warm studio light from above catches the drop as it falls. Product is amber-tinted, translucent. Vertical 9:16. 4 seconds.
Generated in 91 seconds. The slow-motion drop physics held realistically through impact. No color drift on the amber liquid. This clip works as a cut between the talking head and the skin-contact shot.
Tested prompt, post-application glow:
Close-up of a woman's cheekbone and jawline, skin visibly even and glowing after skincare application. Morning light from the right, soft focus background. Warm tone, no product visible. Vertical 9:16. 3 seconds. No face shown.
Generated in 84 seconds. The glow read as real morning light on healthy skin rather than a filter effect. The no-face framing keeps this clip reusable across any demographic targeting without re-generating.
Tested prompt, makeup brush stroke:
Close-up of a fluffy makeup brush applying a warm peachy-pink blush to the apple of a cheek. Soft circular motion, product builds gradually over two passes. Natural daylight. Skin is medium-toned, smooth. Vertical 9:16. 4 seconds.
Generated in 86 seconds. The brush fiber detail was sharp throughout the motion. The color buildup over two passes was the most technically difficult element; it rendered correctly without the blush color bleeding outside the application zone.
Step 4: Captions and assembly
Auto-captions are not optional in beauty. Sound-off viewing is dominant on TikTok and Reels, and beauty content skews toward low-volume browsing. Style captions white with a thin black outline. Every platform's native editor generates them; run them, fix any misrecognitions, don't skip.
15-second assembly structure:
| Seconds | Shot | Model |
|---|---|---|
| 0 to 3 | Talking head, hook | Higgsfield Soul 2.0 |
| 3 to 5 | Product still or dropper motion | Seedream 5.0 / Seedance 2.0 |
| 5 to 9 | Talking head, product explanation | Higgsfield Soul 2.0 |
| 9 to 12 | Skin application or texture b-roll | Seedance 2.0 |
| 12 to 15 | Talking head, CTA | Higgsfield Soul 2.0 |
Color grade: apply a warm-soft LUT in 8frame Studio to unify all clips. The "warm organic" preset works for skincare and fragrance. For makeup tutorials, pull the warmth back slightly to keep lip and cheek colors true to the product.
Routing by category
| Category | Primary models | Budget per 15s ad |
|---|---|---|
| Skincare | Higgsfield + Seedance + Seedream (still) | $4 to $8 |
| Cosmetics and color | Seedream (stills, shade-critical) + Seedance + Higgsfield | $5 to $10 |
| Fragrance | Seedream (stills) + Seedance (motion) | $3 to $6 |
| Hair care | Seedance (motion) + Higgsfield (testimonial) | $5 to $9 |
Color direction in the prompt must be explicit for cosmetics: don't say "red lipstick," say "deep cherry red, not orange-red, not pink-red."
Walkthrough: vitamin C serum ad across 4 demographics, $20 in compute
Four demographic targets for a 15-second TikTok ad: 25 to 34 light-toned, 25 to 34 medium-toned, 35 to 45 medium-deep, 35 to 45 deep-toned.
Shared assets (generated once, reused across all variants):
| Clip | Model | Result | Cost |
|---|---|---|---|
| Serum bottle hero still | Seedream 5.0 | Amber glass accurate, clean shadow | $0.09 |
| Dropper slow-motion | Seedance 2.0 | Clean drop physics, amber color held | $0.58 |
| Glow skin close-up | Seedance 2.0 | Morning light, realistic skin texture | $0.55 |
| Serum application close-up | Seedance 2.0 | Clean absorption, no artifact | $0.58 |
Shared asset total: $1.80. Per variant: 3 Higgsfield talking-head clips at $0.85 each ($2.55), times 4 variants = $10.20. With backup generations for expression selection, the full run landed at $18.40. Four finished ads, each with a demographically matched avatar, shared b-roll, ready for separate Meta ad sets.
Traditional equivalent: four creator UGC videos at $400 to $600 each. Total: $1,600 to $2,400 plus revision rounds.
Pitfalls
Skin texture uncanny valley. Too smooth reads as plastic; too detailed tips into unsettling. Add "realistic skin texture, natural pore detail, not airbrushed, not plastic" to every Seedance skin close-up prompt. Without that direction, the model defaults to oversmoothed output. Generate 2 to 3 variants per skin-contact shot and pick the most natural one.
Before-and-after compliance. Any before-and-after claim must reflect typical results, not exceptional ones. This applies to AI-generated visuals the same as real photography. Include "results vary" as visible on-screen text. Meta rejects before-after in health-adjacent categories without disclosure; check platform policy before submitting.
Dermatologist title authority. Prompting for an avatar described as a "dermatologist" or "MD" is a claim that needs substantiation. If the brand doesn't have a real dermatologist relationship, use "skincare expert" or "esthetician" in the prompt. Don't generate name badges or on-screen titles in the video; add those in post where you control the text.
Color accuracy on lipstick and foundation. Seedream 5.0 holds color well with explicit direction and drifts without it. Name the shade specifically in the prompt ("deep burgundy-red, not orange-red"). If the shade still drifts, upload the product packaging as a reference image instead of describing it in text.
FAQ
What are the FTC rules on AI-generated before-and-after beauty ads?
The FTC's endorsement guidelines require before-and-after depictions to reflect results typical for users, not exceptional results. This applies to AI-generated visuals the same as retouched photography. Include a visible "results vary" disclosure for any before-after format. Meta has stricter category restrictions on health-adjacent creative than TikTok; check platform policies before submitting.
How do you handle skin tone diversity?
Use a separate reference portrait for each demographic variant. Higgsfield's identity conditioning is anchored to the portrait you upload, so swapping portraits is what changes the avatar's skin tone and features across variants. Don't try to redirect skin tone through prompt text on a single reference; it produces inconsistent drift. The walkthrough runs four variants at $18.40 total.
Which model works best for skincare vs makeup ads?
Skincare: Higgsfield Soul 2.0 and Seedance 2.0. The format is testimonial plus skin-result evidence, and Seedance handles skin-contact motion better than any other model on 8frame.
Makeup and color: Seedream 5.0 becomes primary because shade accuracy on stills matters. Use Seedream for any hero still or color swatch, Seedance for application motion, Higgsfield for talking-head segments if the format includes them.
For the full UGC ad workflow that this beauty workflow extends, see how to make a UGC ad with AI. For tested prompts specifically on Seedance 2.0 for product demos and application clips, see the Seedance 2.0 prompts for product demos guide. The pre-built beauty ad template with timeline, LUT, and caption style is available in the 8frame workflow library.