How to improve AI character quality
Last updated:
How to improve AI character quality
The biggest factor in whether your AI character looks like the person you trained it on is the reference photo set — not the prompt, not the scene, not the camera style. Most quality issues trace back to the reference photos, so that’s where the leverage is.
What good reference photos look like
A strong reference set has:
- One era of the person’s life — typically a 5-10 year span. Mixing photos from when they were 30 and when they were 70 confuses the AI.
- 8-16 photos — 3-4 is the minimum that produces a recognizable result; 8-16 is the sweet spot where the AI has enough variety to learn the face robustly.
- Mix of distances — close-up (face + shoulders), medium (waist-up), full-body. At least 2-3 full-body shots help duo-mode composition.
- Mix of angles — front-facing, 3/4 angle, side profile. The most common mistake is all front-facing selfies.
- Mix of expressions — smiling, neutral, candid. All-smiling reference sets produce characters that always smile.
- Mix of lighting — indoor, outdoor, natural, artificial. Helps the AI generalize.
- Mix of outfits — different clothing teaches the AI not to embed clothing as part of the character.
What to skip
- Heavily filtered Instagram photos — beauty filters and face-modification filters give the AI bad information about real features.
- Sunglasses, hats, masks — anything covering the face hurts training.
- Group photos — the subject is small in frame and other people contaminate the training set.
- Mixed eras — pick one era and stick to it.
- HEIC files — convert to JPG first (Apple Photos → Share → Save Image).
How to evaluate a trained character
After training (~30 minutes), don’t go straight to burst mode. Run a single test generation in a neutral scene (e.g., “studio portrait, soft lighting, neutral background”) and check:
- Does the face look unmistakably like the person?
- Do the proportions look right?
- Are there visible AI artifacts (extra fingers, distorted features)?
If the character passes the test, run burst generations. If it doesn’t, retrain with a different reference set — tweaking the prompt won’t fix a weak character.
When to retrain
- 75%+ of generations look generic or “wrong”
- Faces are recognizable but proportions consistently feel off
- Duo-mode results put characters at wrong scales
- You realize you mixed eras in the original training set
Retraining costs the same as initial training (counts as one of your monthly character allowance). Delete the old character first so you don’t hit the limit.
Common reasons characters look “off”
| Symptom | Likely cause | Fix |
|---|---|---|
| Generic-looking face | Too few reference photos | Retrain with 8+ photos |
| Always-smiling, always-same-pose character | All reference photos similar | Retrain with varied expressions |
| Hair color/style inconsistent | Mixed eras in reference set | Pick one era and retrain |
| Body proportions wrong in duo mode | Only close-ups in reference | Add 2-3 full-body shots and retrain |
| Skin tone inconsistent | Heavily filtered references | Use natural-lighting photos only |
| Visible AI artifacts (fingers, eyes) | Generation variance | Run in burst mode and pick keepers |