If you read about AI character pipelines you'll hit the term golden set: the curated library of reference images that defines who a character is. It's the single most important asset in an AI influencer workflow — more important than the model, more important than the prompts. Here's what it is and why it works.
The definition
A golden set is a collection of images of one character — typically 40–60 — that have all been verified to be the same person and that collectively cover the visual situations you'll generate later: different poses, camera distances, expressions, lighting, and outfits. Every future image is generated conditioned on this library, so the model copies the face from verified pixels rather than re-imagining it from a text description.
The "golden" part is the curation: these aren't just any images of the character, they're the ones that survived filtering. One off-model image in the set poisons everything generated from it, so the bar for inclusion is high.
What goes into a good golden set
Coverage is the design principle. A generator asked for a full-body gym shot can't extrapolate well from ten close-up portraits, so the set is built as a matrix:
- Poses — front, three-quarter, profile (and a back shot or two for hair)
- Framings — close-up, mid shot, full body
- Expressions — neutral, smiling, candid mid-motion
- Lighting — daylight, golden hour, indoor warm, flash
- Clothing range — from neutral basics to fitted outfits, so the model learns body shape rather than one silhouette
AI CMO generates 50 of these automatically from a single seed image, weighted toward the shots social content actually uses (full-body and three-quarter angles get extra slots).
The filtering step most people skip
Here's the catch: the reference library is itself AI-generated from the seed — which means some candidates drift. If a slightly-off face makes it into the set, every image conditioned on it inherits the error, and drift compounds over time.
The fix is automated face verification. Each candidate reference is scored against the seed using ArcFace embeddings — face-recognition technology — and anything below a similarity threshold is rejected before a human ever sees it. The user then reviews the survivors and removes anything that feels off even if it scored well. Machine filter for accuracy, human pass for taste.
Locked means locked
Once curated, the set is locked: generation pulls from it, but it doesn't change. This is deliberate. A library that silently grows with new generations would gradually absorb drift — each slightly-off image becoming a reference for the next. Locking the set is what makes month-six content match month-one content.
(You can still extend a locked set deliberately — generating a new batch that goes through the same filter-and-curate gate — which is different from letting it grow on its own.)
Why not just use one great photo?
A single reference can anchor a face for similar shots, but it starves the model everywhere else: ask for a profile when your one reference is frontal and the model invents the missing angle — differently each time. The golden set exists precisely to remove invention. Whatever the scene calls for, there's a verified reference nearby.
How many images you actually need — and where the diminishing returns kick in — is its own question; we wrote up the numbers here.
Build one in minutes instead of days
Assembling a golden set by hand — generating hundreds of candidates, eyeballing faces, organizing coverage — takes a working day with good tooling. AI CMO does the whole loop in minutes: pick a face from 10 candidates, we generate the 50-image coverage matrix, ArcFace filters the drifters, you do one curation pass and lock. Your first character costs $19.