You need three things to hit 100 ad variations per week: a modular creative system, the right AI models for each asset type, and a batch export workflow that multiplies every base asset into 5-10 variants without manual rework. A solo creative operator running this system can replace what used to take a 4-person team, assuming you treat each variation as a deliberate test rather than random noise.
Below is the exact workflow we run at Adsome for DTC clients spending €20k-200k/month on Meta and TikTok.
Step 1: Build a Modular Creative Brief
Before opening any AI tool, define your variation matrix. A single product needs testing across these axes:
- Hook (first 1-3 seconds): problem statement, bold claim, UGC-style question, visual demo
- Body format: product close-up, lifestyle scene, before/after, testimonial overlay
- CTA: urgency, social proof, benefit restatement, offer
- Aspect ratio: 9:16 (Reels/TikTok), 1:1 (Feed), 4:5 (Meta Feed)
Four hooks × three body formats × three CTAs × three aspect ratios = 108 unique combinations from one product. You don't render all 108 from scratch. You render base assets and recombine them.
Step 2: Generate Base Visual Assets in Batches
This is where AI models replace photoshoots. For a typical product drop, generate these base assets in one session:
Product stills (15-20 images): Use GPT Image (gpt-image-1) for hero shots with clean backgrounds and consistent lighting. FLUX Kontext handles product-in-context edits where you need the same bottle on a kitchen counter, a gym bag, or a bathroom shelf without re-prompting the product from scratch. Generate 5 background variations per product angle.
Product video clips (8-12 clips): Kling 3.0 Master tier produces the most reliable close-up product motion at 1080p. Feed it a product still and prompt slow rotation, pour shots, or unboxing-style reveals. Each clip runs 5-10 seconds. Runway Gen-4 Turbo works better for lifestyle scenes where a human hand interacts with the product, since its motion coherence on hand-object contact outperforms Kling for that specific use case.
Talking head or UGC-style clips (4-6 clips): Veo 3 generates video with native audio, which means you can produce synthetic testimonial-style clips where the voice matches the speaker without post-sync editing. For brands comfortable with AI-generated presenters, this removes the single biggest bottleneck in UGC production.
One batch session producing these base assets takes 3-4 hours including prompt iteration and quality filtering. Reject roughly 30-40% of outputs on the first pass.
Step 3: Multiply Base Assets Into Variations
This is where 25 base assets become 100+ ad variations.
- Hook swaps: Take each base video clip and pair it with 3-4 different opening frames. A text overlay hook, a zoomed crop of the product, a reaction-style face, or a pattern-interrupt color flash. Edit these in CapCut or Premiere using templates.
- Text and copy overlays: Run your 3 CTA variants and 4 hook copy lines across each visual. This is pure template work. Build 3 text style templates (bold sans-serif, handwritten, minimal lowercase) and apply across all assets.
- Aspect ratio exports: Every finished 9:16 asset gets a 1:1 and 4:5 crop. Auto-reframe tools in Premiere or CapCut handle 80% of these correctly. Manual adjustment takes 30 seconds per clip for the rest.
- Color and pacing variants: Duplicate your best-performing base video. Speed up the first 2 seconds by 1.5x for one variant. Apply a warmer color grade for another. These micro-changes affect thumb-stop rate measurably on Meta.
Step 4: Quality Gate Before Upload
Not every variation ships. Apply a three-check filter:
- Does the hook read in under 1 second with sound off?
- Is the product visible within the first 3 seconds?
- Does the CTA match the landing page offer exactly?
This filter typically kills 10-15% of variations, which keeps your ad account quality score healthy and prevents budget waste on incoherent creatives.
Realistic Weekly Schedule
| Day | Task | Output |
|---|---|---|
| Monday | Brief + prompt writing | Variation matrix, prompt library |
| Tuesday | AI asset generation (images + video) | 20-25 base assets |
| Wednesday | Review, re-generate rejects, select finals | 15-18 approved base assets |
| Thursday | Variation assembly (hook/CTA/ratio swaps) | 80-110 variations |
| Friday | Quality gate + upload to ad platform | 90-100 live variations |
