11 min read
How to Create Consistent AI Characters Before Motion Transfer
Identity locks, sheets, and still discipline so Genjutsu changes performance — not the person — across dance, UGC, and product clips.
Consistency is a stills problem first
Motion transfer cannot rescue a mushy identity. If two stills already look like cousins instead of the same person, every dance clip will amplify the drift. Treat character consistency as a pre-production gate: only after the face and wardrobe read as one talent should you introduce Genjutsu.
This guide is the prep layer under /ai-influencer and the influencer video how-to. Skip it and you will burn credits debugging “why does she look different?” when the real bug was the sheet.
Build a minimal character package
Hero portrait: sharp eyes, natural skin texture, brand-correct makeup and hair. Avoid heavy beauty filters that melt under motion.
Body still: three-quarter or full-body with the wardrobe you will reuse. Note shoes and accessories — extremities are where models invent details.
Style rules: write three bullets (palette, era, do-nots). Reuse them whenever you regenerate stills so mid-week posts do not invent a new jacket.
Stress-test before you scale
Run one calm motion transfer (walk or simple gesture) and one harder clip (dance or fast hands). If identity fails the calm test, do not batch ten dances.
Compare outputs at the same frame times. Drift in eye shape, age, or wardrobe color means freeze a better still — not a longer prompt.
Hand-off into Genjutsu use cases
Once the package passes, route by job: dance socials → /dance-video, ad variants → /ugc-video-generator, pack and hold demos → /product-video. Keep the same still across those routes when you want one influencer universe.
For tool shopping and “what else exists,” use /genjutsu-alternatives rather than thin vs-pages. Motion vocabulary lives on /motion-transfer as a support page, not the primary acquisition URL.
Common identity failure modes (and fixes)
Age drift: the still is over-smoothed. Re-export with natural skin texture and less beauty filter.
Wardrobe melt: colors in the still conflict with lighting in the reference. Either match lighting mood or simplify the outfit to flatter shapes.
Hair chaos: flyaways plus fast dance spins. Prefer a tied or simpler hairstyle for the first validated package, then add complexity.
Twin problem: you swapped stills mid-week. Name files `persona_v1_hero.png` and refuse to generate with anything else until v1 is retired on purpose.
Keeping a package alive for months
Store the hero still, body still, style bullets, and two golden reference clips that always look like “her.” When onboarding a teammate or agency, send that zip — not a Slack screenshot.
When trends change, add a new outfit still rather than regenerating the face. Identity equity compounds; wardrobe can fashion-cycle.
Practical parameters
- Minimum package
- Hero face still + full/three-quarter body still
- Style lock
- 3 written rules: palette, wardrobe, do-nots
- Gate test
- 1 calm motion + 1 hard motion before batching
- Failure signal
- Age, eye shape, or wardrobe color drift across takes
- Archive
- Zip: stills + style bullets + 2 golden motion clips
FAQ
- Do I need a full LoRA or fine-tune?
- Not for Genjutsu’s image + motion workflow. A disciplined still package is enough for many short-form pipelines. Fine-tunes help when you regenerate stills often across tools.
- Why do hands look wrong even when the face is fine?
- Hands are hard. Prefer references with visible, unoccluded hands and avoid tiny props crossing fingers on key frames.
- Can I change outfits often?
- Yes — treat each outfit as a new still in the package. Do not ask one still to invent a wardrobe the model never saw.
- Where should I generate after the sheet is ready?
- Log in and open the homepage studio. Use use-case pages to plan the job; use this guide to keep identity stable.
- What if the brand wants a seasonal look?
- Version the package (v1 spring, v2 summer) with explicit stills. Do not mix v1 face with v2 wardrobe mid-batch without a new gate test.