Open-source alternative to Arcads: AI UGC ads

An AI creator talks to camera about your product, holds it, uses it, and sounds like a person filming themselves. Generation runs on varg: the MCP (call_tool) or POST https://api.varg.ai/v2/tools/{image|video|speech}/call. Quote credits only, from estimate_cost, and wait for a yes before spending.

What the data says: in one operator's $100K test (220 videos), human UGC had the higher CTR and AI the higher ROAS, because AI made more variants cheaply. AI works for cold traffic, low-trust products and finding the winning angle; it is weak for high-ticket offers and warm audiences. The script matters more than whether the face is real. Use AI to find the angle, then consider remaking the winner with a real creator.

0. Gates

1. Where does it run? Meta ads and TikTok ads: OK, labelled. TikTok Shop video: stop and use tiktok-shop-video (AI voices as presenter and AI-made material are against Shop rules).
2. Claims: the AI creator may demo and state facts. Never write "I used it for two weeks and…", before/after results, or "my skin cleared". The FTC bans fake testimonials, AI ones included (up to $53,088 per violation); New York requires a synthetic-performer disclosure in ads. Real reviews may be quoted as reviews ("2,100 reviews say…").
3. Product photo: clean, straight-on, label readable. "If the label isn't readable in the reference, the model invents its own text. That's where most people's ads die."
4. Cost: quote the full batch with estimate_cost (assume ~3 takes per usable Seedance clip, ~1–2 for Kling 4).

1. Cast the creator

• Age matches the claim, written as a number ("54-year-old woman"); room matches the problem (energy = kitchen at 6 am, sleep = bedroom lights off, bloating = bathroom mirror, joints = stairs, private = car).
• Real skin: "freckles, visible pores, a small blemish near the jaw, flyaway hairs, no retouching, no skin smoothing".
• One outfit with one specific detail ("collar slightly crooked, sleeves pushed up").
• Generate the creator still with gpt_image_2_5 or nano_banana_2_1: "iPhone selfie photo, slightly below eye level, phone at arm's length…". Pick one, then make an identity sheet (front, ¾, side, expressions) and reuse the same image on every shot. Deeper casting: varg-character-design.