Short answer: 40–60 well-chosen images is the sweet spot for a consistent AI character — but the count matters less than what the count covers. Here's the reasoning, so you can judge any setup yourself.
Why one image isn't enough
A single great portrait anchors the face for shots that resemble it: similar angle, similar distance, similar light. Step outside that envelope — ask for a profile, a full-body shot, dramatic side lighting — and the model has to invent what it can't see. It invents differently every time. That's face drift in its purest form: not randomness, just extrapolation from missing data.
One image is a pin. You need a cage.
The coverage math
Think of your character's visual space as a grid: poses × framings. Three useful poses (front, three-quarter, profile) times three framings (close-up, mid, full body) is nine cells. Add expression variation (neutral, smiling, candid), a few lighting conditions, and a range of outfits, and you arrive at roughly 40–60 images to put multiple verified references in every cell you'll actually use.
That "multiple" matters. With one reference per cell, a single subtly-off image silently owns that entire pose-framing combination. With three or four per cell, the consensus drowns out the outlier.
This is why AI CMO builds 50 references per character by default — generated as a weighted matrix (social content leans on full-body and three-quarter shots, so those cells get extra slots), then filtered with ArcFace face matching before a human curation pass.
Where more images stop helping
Past ~60–80 references, returns flatten fast:
- Redundancy — the new images mostly duplicate covered cells.
- Curation fatigue — every reference must be verified on-model; 200 images means 200 judgment calls, and tired curation lets drifters in. A drifter in the set is worse than a gap.
- Selection dilution — generation conditions on a handful of references per request, chosen to match the prompt. The selector picks from the best matching cell whether the library has 50 images or 500.
A 500-image library isn't a stronger identity; it's a bigger haystack with the same number of needles in play per generation.
Quality gates beat raw count
A 30-image library where every image is verified-on-model outperforms a 100-image library with five drifters. Ranked by impact:
- Every image verified (face-similarity scoring against the seed, then human review)
- Every needed cell covered (poses × framings you'll actually prompt for)
- Multiple references per high-traffic cell
- Raw count
Count is fourth. It's the easiest to market ("trained on 1,000 images!") and the least predictive.
What about LoRA training datasets?
If you're fine-tuning instead of using reference conditioning, the numbers look similar for different reasons: LoRA guides commonly recommend 20–50 high-quality, varied images, and the same failure mode applies — a sloppy dataset bakes drift permanently into the weights, where you can't curate it out afterward. (More on that trade-off in platform vs DIY LoRA.)
The practical takeaway
Don't ask "how many images." Ask: is every pose-framing combination I'll prompt covered by multiple verified references? For social-media content, that answer lands at 40–60 images nearly every time.
Building that by hand takes a day. AI CMO does it in minutes: one seed, a 50-image weighted coverage matrix, automatic ArcFace filtering, one curation pass, locked. First character: $19.