How to Use Genjutsu (Motion Transfer Studio)

Genjutsu is an AI video motion transfer studio: you upload a character (or product) image and a reference video that already contains the performance you want. The model maps pose, timing, and often camera energy from the clip onto your subject. You are not asking a text prompt to invent choreography from scratch.

Use this when you need controlled movement — dance, walks, gestures, product demos — while keeping a recognizable identity. Genjutsu on genjutsu.online is an independent product experience; it is not an official Higgsfield property, even when the underlying motion ideas feel familiar to creators who have used similar pipelines elsewhere.

Lock the still first. A sharp, well-lit image with a readable face and body outline beats a stylish but occluded crop. Full-body or three-quarter framing usually transfers better than extreme close-ups when the reference is a dance or walk.

Pick the motion file second. Prefer a short clip with clear silhouettes, stable framing, and a single primary subject. Busy backgrounds, rapid cuts, and heavy occlusion force the model to guess. Match camera height roughly between still and reference — a top-down still with a side-on walk reference is a common failure mode.

Decide whether you need Motion Transfer (identity from the still, motion from the clip) or Object Swap (person/scene already in the video, replace a held item or garment). Mixing those intents mid-test wastes credits.

1) Sign in on genjutsu.online so generation and credit balance are available. The marketing demo on /motion-transfer only shows the upload layout; real renders run from the homepage studio after login.

2) Upload your character or product image. Keep the subject large in frame; avoid tiny figures in wide empty scenes.

3) Upload the reference video. Credits follow the length of this motion clip, so trim dead air before upload when you can.

4) Choose resolution and optional prompt guidance if you need style hints, then generate. Review the preview, then iterate with small changes — crop, duration, or a cleaner still — instead of swapping every variable at once.

5) Save what works in history and reuse the same identity package across new references so performance changes, not the person.

Motion length drives cost and how much choreography you ask the model to hold. Short hooks (a few seconds) are cheaper for A/B tests; longer walks need cleaner references.

Resolution trades detail for speed and credits. Start at a working resolution to validate identity hold, then up-res once the motion read is right.

Prompt fields are optional seasoning, not a replacement for a good reference. If the dance is wrong, fix the clip — do not write a paragraph hoping the model invents new footwork.

Identity: does the face and wardrobe still read as the same person across the clip? If not, tighten the still (better lighting, less stylization) before changing motion.

Motion fidelity: do peak poses land on the beats you care about? If the reference is muddy, re-export a cleaner source rather than regenerating endlessly.

Camera: if framing drifts oddly, prefer references with fewer whip pans. For product work, keep the item large in both still and plate.

Publish policy: label synthetic talent where your platform or brand guidelines require it — especially for UGC-style ads and ecommerce models.