By Higgsfield Genjutsu editorial teamPublished Oct 31, 2026Updated Oct 31, 2026
What is a Genjutsu reference video?
A Genjutsu reference video is the source clip that supplies the performance Genjutsu keeps. It provides the body movement, camera, framing and timing while reference images decide who or what replaces the visible subject.
The current uploader accepts a 3 to 30 second MP4 or MOV source. One clean 3 to 10 second take usually produces a more reliable result than a longer montage with several angles.
Source video and reference images have different jobs.| Input | What it supplies | What it should not do |
|---|
| Reference/source video | Motion, camera, timing, framing and scene geometry | It should contain cuts, speed ramps or competing angles |
|---|
| Character image | Identity for the replacement performer | It cannot invent a full performance without a source video |
|---|
| Outfit/product/location images | Specific elements to rebuild | They should not be mixed into one contradictory mood board |
|---|
What length should the source video be?
Use the shortest section that communicates the full action. A 3 to 6 second hook, dance phrase, turn or product reveal is enough for a first test. Add duration only after identity and body motion survive.
The platform accepts videos up to 30 seconds, but the credit cost rises with duration. A longer clip also contains more angles, occlusions and decisions that can drift.
- 3–6 seconds: best first test and social hook.
- 6–10 seconds: one complete phrase or product interaction.
- 10–30 seconds: only when the camera and subject remain stable.
- Trim at cuts, camera whips and blurred transitions.
- Use the same duration you expect in the finished platform placement.
What framing makes a good Genjutsu source clip?
Show the subject clearly and keep the action inside the frame. Full-body dance needs visible feet and hands; a talking-head recast needs a readable face and stable shoulder framing. The source defines the camera geometry Genjutsu must solve.
Choose a clip where the new identity can plausibly own the original silhouette. A compact reference forced onto a very different body or costume creates more drift than a physically compatible reference.
How much camera movement can the reference video have?
Camera movement is part of the value of Genjutsu: you can preserve a real push, orbit or handheld move. The key is consistency. One direction and one speed are easier to reproduce than a shot with several wipes or zooms.
Do not add digital zooms, shake packs or speed ramps before upload. Use the clean camera movement from the source and make those edits after generation.
What lighting and image quality should you use?
Use even exposure with a visible face, silhouette and floor relationship. Keep the main light direction readable so the rebuilt identity can sit in the same scene. Avoid clipping highlights or losing the body in deep shadow.
Resolution helps, but sharpness matters more. A clean 720p source with a stable frame beats a heavily compressed 4K clip with motion blur and rolling-shutter artifacts.
- No captions, stickers, watermarks or edit overlays in the source.
- No strong beauty filter that changes the face frame to frame.
- No severe motion blur around hands, face or product contact.
- Keep the subject large enough to see identity and object boundaries.
- Use a consistent frame rate without dropped-frame stutter.
How do you prepare the source file for upload?
Trim the clip, remove audio-reference files and export MP4 or MOV. Keep the file under 100MB. An H.264 MP4 is the safest compatibility choice for most sources.
If upload fails with 413, reduce the bitrate or trim the clip. Do not submit a long high-bitrate screen recording and assume the platform will automatically cut it to the useful section.
Does the source video need to match the reference images?
The body proportions, visible angles and light direction need to make sense together. If the source turns into profile, include a profile character reference. If the source wears one outfit, keep wardrobe references separate and internally consistent.
The source supplies geometry; images supply replacement appearance. When those two inputs contradict each other, the model has to average them and identity drift increases.
How many performers should the source clip contain?
Start with one performer. One person keeps body ownership unambiguous and gives each character reference more influence. For two performers, keep them separated and prepare distinct identity sets.
Avoid lifts, contact, costume swaps and performers crossing each other in the first test. Those create additional spatial decisions on every frame.
What source footage can you legally upload?
Upload footage you own or have permission to transform, including permission for every real person, location and branded object it contains. A public social video is not automatically licensed for recasting or commercial use.
The output still carries the rights attached to the source, references, music and likeness. Genjutsu changes the visual result; it does not grant permission to use those inputs.
What are signs that you chose the wrong source clip?
The wrong source has internal cuts, insufficient face or body coverage, extreme blur, heavy overlays or several competing subjects. Those conditions produce errors that longer prompts cannot reliably fix.
If a test fails at the same frame, replace the clip before changing the model settings. A cleaner source generally saves more credits than repeated attempts to rescue unusable footage.
- A new camera angle appears halfway through the clip.
- The performer is hidden by hands, hair or another person.
- The body leaves frame during the main action.
- The source already contains large captions or editing effects.
- The clip is long but mostly contains no useful action.
Key takeaways
- The source video supplies motion, camera, timing and scene geometry.
- Use a clean 3–10 second MP4/MOV with one performer whenever possible.
- Choose a source that the replacement identity can plausibly own.
- Trim the source before upload and keep it below 100MB.
- Rights to footage, people, brands and music remain your responsibility.
Higgsfield Genjutsu