Every tool in this guide can spawn a hundred takes. Almost none can tell you which one to keep.
The pitch behind the whole category — Crown's "keep the best of dozens," Variant's "endless designs, just scroll" — is the same pitch we've been making internally for a year: the model's advantage is making drafts at near-zero cost; the human advantage is the taste to pick the right one. That thesis is now correct and widely shared. The interesting question has moved on. If generation is free, the product is no longer the generator. It's the selector.
Watch where each tool puts the judgement, and the field sorts itself into two camps.
| Camp | Curates on | Examples | Closes the loop? |
|---|---|---|---|
| Taste-gated studios | Human judgement, brand fit | Crown, Variant, Banani | No — you decide |
| Performance factories | Measured outcome (CTR, ROAS) | Arcads, AdCreative, Omneky | Yes — the metric decides |
The performance camp is the commercially proven one, because it has a selector that doesn't get tired and can't be argued with: the number. The taste camp is more interesting creatively but carries the harder problem — selection that doesn't scale and has no ground truth.
Crown is the purest expression of the thesis and the best-told story in the category. On our standard P/Q/S radar it lands at 5/9 — a watch, not an act. Two reasons. It's a closed black box: no public API, no MCP, no brand anchors, so it regresses toward generic the moment the brief is brand-led. And it has no feedback loop — it hands you a field and walks away. It is behind Variant on generation depth (Variant writes the running code) and behind the ad cluster on the one feature that makes variation sellable: a selector tied to outcomes.
The unsolved piece sits exactly between the two camps: brand-anchored generation with a scored cull. Variants that are on-voice by construction — generated against a written brand anchor set, not a blank prompt — and then narrowed by a gate that is part model, part measured, part human. That is the thing none of the studios ship and the ad factories don't care about, and it's the thing a 25-year brand operator can build credibly. It's also already most of our stack: the brand-voice and design-anchor work, the Taste Pipeline that turns references into an enforceable skill, and the anti-slop discipline a tool like Hallmark codifies.
So the build, if a client asks for "generate every version of the campaign," is not to resell Crown. It's to run the generation on an ownable harness — variant-ui spins up parallel takes inside Claude Code with our own models and data residency — gate them against the client's brand anchors, and where a real metric exists, let performance make the final call. Closed studio for speed; owned pipeline for brand, privacy and the loop.
Don't sell the ability to make a hundred things. Everyone has that now. Sell the judgement to keep the one that's on-brand and earns its place — and the machinery that enforces it every time.