Every consistent AI character traces back to one decision: a single image, chosen from a batch of candidates, that becomes the canonical answer to "what does she look like?" That's the seed image, and because every reference, every filter decision, and every future generation inherits from it, it's worth understanding what makes a good one.
What the seed actually does
The seed is the root of a dependency tree:
- The reference library is generated from the seed — each reference is the model's attempt to render the seed's person in a new pose, framing, or light.
- The face-similarity filter scores every candidate reference against the seed — the seed is literally the definition of "on-model."
- Ongoing generation conditions on references that all descend from the seed.
Change the seed and you've changed the character. Everything else is downstream plumbing.
What makes a strong seed
When you're picking from a candidate batch, prefer:
- A clear, well-lit face — front or slight three-quarter angle, no heavy shadow across the features. The filter measures facial geometry; occlusion and murk make every downstream similarity score noisier.
- Distinctive but reproducible features. Strong characteristics (a sharp jawline, distinctive eyes, memorable hair) survive re-rendering better than a generically pretty face, which the model can quietly swap for another generically pretty face without tripping any alarms.
- Neutral-to-mild expression. An extreme expression bakes itself into the identity; references generated from a mid-laugh seed fight to keep the laugh.
- The look you want at its most typical. The seed should show the character's default — signature hairstyle, usual makeup level. Variants belong in the reference set; the seed is the baseline they vary from.
What doesn't matter much: background, outfit, framing beyond the face. Those are scene properties, and scenes are disposable.
Picking from candidates: commit
The standard workflow generates a batch of candidates from your brief — AI CMO gives you 10 — and you click one. The psychological trap is regenerating forever in search of perfect. Two things to know:
- Candidates are starting points, not finished aesthetics. Lighting, styling, and vibe all get explored across the 50-image reference build. You're choosing a person, not a photo.
- A committed "very good" beats an endless search for "perfect." Character accounts live or die on consistency and cadence, not on whether the founder face was a 9.3 or a 9.6. (Regenerating a fresh batch costs a few credits if the whole set genuinely misses your brief — the cure for a bad brief, not for indecision.)
Can you bring your own seed?
Tempting — and where the legal landmines live. Using a photo of a real person (a celebrity, an ex, even yourself-but-idealized) creates publicity-rights and impersonation exposure that no disclosure label fixes. Generated-from-text seeds are wholly fictional by construction — nobody owns that face — which is why brief-to-candidates is the compliant default on AI CMO.
Seed quality compounds
A mediocre seed doesn't fail loudly. It fails as a reference set with mushy similarity scores, a filter that can't tell drift from variance, and a character who's 92% the same person in every shot — which followers register as "something's off" without knowing why. Spend your care at the root of the tree.
Write a one-line brief, get 10 candidate seeds, and pick yours — first character for $19.