Marketer Ad Variant Factory

Ad variant factory

Take one approved concept and explode it into 10-30 shippable ad variants for A/B testing on Meta, Google, TikTok, and Pinterest. Built for speed (parallel batch) and for direct upload to ad accounts (deterministic naming).

When to Use

• User has a hero / concept and needs many variants across headline × visual × CTA × background for A/B tests.
• "Fan out 30 variants of this ad for Meta" / "generate a test matrix" / "multiply this creative".
• Prepping a new ad-set launch — needs 9:16, 1:1, 16:9 with 3-5 visual variants each.
• Do NOT use for a single hero (use gen-ai generate) or for cross-channel creative (use marketer-campaign-kit). This skill is for depth on one concept, not breadth across channels.

Prerequisites

Ask up front if the brief doesn't cover it (combine into one message):

1. Hero asset — path or URL to the approved concept image.
2. Axes to vary — visual direction (1-5), background/scene (1-5), focal composition (close-up vs wide), optional: color treatment.
3. Platforms / aspect ratios — Meta needs 9:16 + 1:1, TikTok is 9:16, Display wants 16:9. Confirm which.
4. Variant count — how many total? 10-15 is typical for a first test, 30+ for broad exploration.
5. Naming convention — what does the ad platform require (e.g. {campaign}_{axis}_{variant}_{size}.webp)?
6. Brand guardrails — colors (hex), forbidden elements, existing brand.md?

If the user just says "a lot", default to 5 visuals × 3 ratios = 15 variants.

How to Run

1. Anchor on the hero. The hero is the reference image — every variant should feel like a sibling, not a cousin. Upload to Drive first if it's local so downstream jobs can reference a URL.
2. Define the variant matrix. Keep axes explicit. 5 visual directions × 3 aspect ratios = 15 jobs. Don't mix 8 axes — the test becomes unreadable.
3. Write the manifest. One job per variant, unique id that maps to your ad-platform naming convention. The image field references the hero.
4. Estimate + dry-run.

``bash gen-ai batch run variants.json --dry-run ``

1. Run at concurrency 6-8. Image variants are fast and independent — push concurrency higher than the default 3. Watch for 429s; back off to 4 if you see them.

``bash gen-ai batch run variants.json -c 8 -o ./ad-variants ``

1. Audit and resume. Filter results.json for non-completed jobs, retry.

``bash gen-ai batch resume ./ad-variants ``

1. Hand off. Ads platform uploaders expect a flat folder with standard naming — results.json has every path + URL for direct CSV import to Meta Ads Manager / TikTok Ads / Google Ads.

Quick Reference

Naming convention {campaign}_{visual}_{composition}_{size} keeps downstream uploads clean.

{
  "defaults": {
    "model": "recraftv4",
    "negativePrompt": "low quality, watermark, busy background",
    "imageUrls": ["https://cdn-pipeline-output.picsart.com/.../hero.webp"]
  },
  "jobs": [
    { "id": "launch_bright_closeup_9x16", "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "9:16" },
    { "id": "launch_bright_closeup_1x1",  "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "1:1"  },
    { "id": "launch_bright_closeup_16x9", "prompt": "bright daylight variant, close-up focal, warm tones", "aspectRatio": "16:9" },
    { "id": "launch_studio_wide_9x16",    "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "9:16" },
    { "id": "launch_studio_wide_1x1",     "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "1:1"  },
    { "id": "launch_studio_wide_16x9",    "prompt": "studio lighting, wide shot, neutral backdrop",        "aspectRatio": "16:9" },
    { "id": "launch_urban_medium_9x16",   "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "9:16" },
    { "id": "launch_urban_medium_1x1",    "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "1:1"  },
    { "id": "launch_urban_medium_16x9",   "prompt": "urban street setting, medium shot, cinematic",        "aspectRatio": "16:9" }
  ]
}

9 variants from 3 visual × 3 ratios. Scale to 15 or 30 by adding visual rows. Remember: no count field — emit one job per variant.

Quick Reference

| Sub-task | Model | Why | |----------|-------|-----| | Brand-consistent variants from a hero (default) | recraftv4 | Strongest at keeping design language consistent across many renders | | Photoreal product / lifestyle variants | flux-2-pro | Best photoreal adherence, great for Meta/TikTok product ads | | Variants with readable headline text baked in | ideogram-v3 | Only model that reliably renders legible copy — use when you can't overlay | | Face/character continuity across variants | gemini-3-pro-image | Nano Banana Pro locks subject identity best | | Background swaps on a fixed subject | gen-ai change-bg (subcommand) | Keeps the subject pixel-identical, only swaps the scene | | Ultra-cheap exploration before the flagship run | gemini-3.1-flash-image | ~5x cheaper, fast — use to pick winning prompts, then regenerate with flux/recraft |

Confirm IDs with gen-ai models --mode image.

Procedure

• Explicit variant axes. Decide 3-4 axes up front (visual, composition, color, setting). Scattershot prompts make A/B results unreadable.
• Hero as reference image on every job. Use the image field in defaults — every variant inherits the brand look.
• Deterministic naming = direct ad-platform import. {campaign}_{visual}_{composition}_{size} parses cleanly in Meta/TikTok/Google ads CSV templates.
• Draft cheap, upgrade winners. Run 30 variants through gemini-3.1-flash-image for ~$1. Pick top 8. Regenerate those 8 through flux-2-pro or recraftv4 for the final upload.
• Text-safe zones per platform. Meta Stories reserve 250px top + 310px bottom. TikTok reserves ~300px at bottom for UI. Prompt focal into the center 60% of the canvas.
• Concurrency 6-8 for image variants. Images are fast — higher concurrency finishes a 30-variant run in under 2 minutes. Drop to 4 if you see 429s.
• Never overwrite silently. Unique id per variant means unique output filename — resume is safe and collision-free.
• Never claim results in the prompt. "Viral ad, 10M views" doesn't improve output; describe framing, subject, lighting, mood.

Pitfalls

• Too many axes → unreadable A/B. Vary 3-4 at most. If you change visual + composition + color + setting + headline in one variant, you can't isolate the winner.
• Missing the hero reference. Without image in defaults, each variant drifts visually — the bundle doesn't feel like one campaign.
• Wrong aspect ratio for the platform. TikTok is 9:16 full-bleed, Meta Reels is 9:16, Meta feed is 1:1 or 4:5 (not 1.91:1 anymore), Google Display is 300×250 / 728×90 / 160×600 — check the ad set requirements before fanning out.
• Text in the image without localization plan. Baked-in copy blocks localization — keep ad copy in the ads-manager overlay unless it's a one-market run.
• Running 100 variants in one go with no draft phase. 100 full-price flagship renders = wasted credits. Draft → pick → upgrade.
• Overwriting results on re-runs. Keep output dirs per run (./variants-$(date +%F-%H%M)) so resume + audit work cleanly.

Verification

Run gen-ai whoami to confirm authentication, then re-run the failed command with --debug.

Cost & time

| Variant count | Model | Credits each | Total | Wall time @ concurrency 8 | |--------------|-------|--------------|-------|---------------------------| | 9 variants | recraftv4 | ~2 | ~18 | ~45s | | 15 variants | recraftv4 | ~2 | ~30 | ~90s | | 30 variants (exploration) | gemini-3.1-flash-image | ~0.5 | ~15 | ~2 min | | 30 variants (flagship) | flux-2-pro | ~2 | ~60 | ~3 min | | 30 drafts + 8 upgraded | mixed | | ~30 | ~4 min total |

Always confirm with gen-ai batch run variants.json --dry-run and pause if the estimate exceeds the user's cap.

See also

• gen-ai-use/SKILL.md — full CLI reference (flags, model catalog, auth)
• gen-ai-batch/SKILL.md — manifest shape, concurrency tuning, resume
• gen-ai-workflows/SKILL.md — general multi-step patterns
• workflows/marketer-campaign-kit/SKILL.md — chain before this to establish the hero + brand look
• workflows/marketer-localize-campaign/SKILL.md — chain after this to fan winning variants across markets

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: PicsArt
• Source: PicsArt/gen-ai-skills
• License: MIT

Install and usage instructions live in the source repository linked above.