A question from technically-curious users: "which image model does the platform use?" The honest answer is two different ones, on purpose — and the reasoning says a lot about what character consistency actually requires. Here's the architecture, vendor-neutral.
Two jobs, two optimization targets
An AI character's life has two phases with opposite requirements:
Phase one: minting. From one seed image, produce ~50 references across a coverage matrix — this pose, this framing, this lighting, this outfit tier. The job rewards instruction-following and edit-style control: the model must take an image plus a precise textual directive and execute the directive faithfully. Identity slips are acceptable here — every output gets face-similarity filtered anyway, so the pipeline can afford a model that occasionally drifts but reliably obeys.
Phase two: production. Every day, take a scene prompt plus a handful of verified references and render the character into a new scene. The job rewards multi-reference fidelity and photorealism: how well the model copies a face from example images into novel contexts. Instruction precision matters less — the prompt describes a vibe, not a matrix cell — but identity transfer is everything, because these outputs ship straight to a feed with no filter gate.
These are genuinely different benchmarks, and as of 2026 no single model leads both. Editing-class models (the Gemini image family being the well-known example) excel at "take this person, now show her in profile, full body, golden hour"; reference-conditioned generators (the Seedream/FLUX-style families) excel at "here are photos of a person, render her at a café, indistinguishable from an iPhone shot."
Why the filter makes the split safe
The architecture only works because of the quality gate between phases. Minting with an obedient-but-occasionally-drifty model is fine when every output is scored against the seed and the misses are discarded — the golden set that survives is verified by measurement, not by trusting the generator. Production then conditions on those verified references, so phase-two fidelity is anchored to ground truth.
Skip the filter and the split becomes dangerous: drifted references poison production silently, and the drift compounds. The two-model design and the verification gate are a package deal.
The strategic bonus: model independence
Because identity lives in a library of photographs rather than in any model's weights, each phase can upgrade independently. A better reference-fidelity model ships? Swap phase two; the same golden set conditions it. A better editing model? Swap phase one for future characters. Compare this with the LoRA architecture, where identity is married to the base model it was trained on and every base-model generation leap means retraining.
This is also a useful lens for evaluating any character tool: ask where the identity lives. In weights → you inherit retraining costs forever. In verified references → the identity outlives every model in the stack.
What this means if you're not an engineer
Practical translations: a platform built this way mints characters in minutes (no training), survives model churn without your character changing, and can show you why each reference was kept (it has a similarity score, not vibes). Those properties — speed, durability, auditability — are the user-visible symptoms of the architecture underneath.
AI CMO runs exactly this design: an instruction-class model builds the matrix, ArcFace gates it, a reference-fidelity model renders your daily content. You experience it as "she always looks like her" — $19 to see it on your own character.