Usage Guide

Omni-Reference Mode

Omni-Reference mode is ideal for reliably reproducing a subject, scene, motion, or voice. Assign each asset a clear reference name, such as @Image1, @Video1, and @Audio1.

Example: Multimodal Subject & Audio Reference

Two character images, two audio clips, and a gym scene can together control character identity and dialogue timbre:

Image/Video Subject Reference

Applicable scenarios include character consistency, character cards, clothing, product appearance, scene reproduction, and motion reference.

Your prompt should include at minimum:

The correspondence between subjects and assets;
The hairstyle, facial structure, clothing, accessories, colors, materials, and proportions that must be preserved;
The expressions, poses, camera distance, and environmental dynamics that are allowed to vary;
The pacing, motion logic, and cut style of the video reference.

Video Timing Reference

0–2s: Subject displayed center-frame, slight head turn;
2–2.4s: Horizontal flip along the vertical center axis;
2.4–4.4s: Next subject displayed center-frame;
Repeat this rhythm, ending with the final subject holding steady.

The Chinese-style headwear example uses a fixed "center display — horizontal flip — next subject display" rhythm:

Audio Reference

Subject 1 uses the voice timbre of Audio 1, Subject 2 uses the voice timbre of Audio 2.
Subject 1 turns and says: "What are we training today?"
Subject 2 wipes off sweat and replies: "Let's do back — my legs are still sore from yesterday."

File/Web Page Reference

Specify the role each source plays: character design, scene, storyboard, background knowledge, or visual style. Avoid letting the same asset carry conflicting constraints.

Video editing should distinguish between "what to keep" and "what to change," and mark the time range where the change occurs. Video extension should clearly describe "the ending state of the original clip → the motion in the extended segment → the new ending state."

First/Last-Frame Mode

Describe the first-frame composition, the in-between motion, and the last-frame composition separately. When the first and last frames differ significantly, add more in-between motion and timing cues to avoid subject jumps, discontinuous limbs, or scene resets.