A reference file helps only when the prompt says what it controls and what it must leave alone.
A practical workflow for deciding what each image, video, and audio reference should control before you generate in Nsketch.
Give Every Reference One Job
Uploading more files does not automatically make a video brief clearer. A portrait can fight a motion clip over wardrobe, location, or identity. An audio track can set a pace that the requested camera move cannot finish in time.
Before writing the scene, decide what each file is allowed to control. Then state that job in the prompt and name the details that must not transfer. This turns a loose pile of references into a usable production brief.
| Reference | Best used to control | Example instruction |
|---|---|---|
| Image | Identity, object shape, composition, or visual treatment | Use this image for the woman’s face and coat. Do not copy its background. |
| Video | Movement, camera path, action timing, or edit rhythm | Use this clip for the lateral tracking pace only. Do not copy the performer or location. |
| Audio | Voice, beat, ambience, or cue timing | Use this audio for delivery and pause timing. Keep the room tone quiet. |
| Prompt | Priority and conflict resolution | The image controls identity. The video controls motion. Preserve the coat and station. |
Let the Image Reference Lead Identity and Look
An image is the cleanest anchor for a face, product, outfit, or composition. In this Seedance 2.5 test, the still image defined the woman, her silver hair streak, the rust-orange coat, the transparent umbrella, and the station geometry. The prompt gave motion a separate job: turn, lower the umbrella, walk three steps, and track sideways.
The generated clip keeps those visual anchors while introducing a simple, readable action. The instruction also names what cannot change, which gives the model fewer chances to improvise around the subject.
Seedance 2.5 · 8 seconds · 720p · Generated in Nsketch
Prompt
Use the uploaded image as the sole authority for the woman's identity, short black bob with one silver streak, rust-orange coat, transparent umbrella, and brutalist station. Begin in the same three-quarter framing. She turns toward the empty platform, lowers the umbrella, and walks three measured steps as the camera tracks sideways at her pace. Blue-hour light stays cool and the wet floor reflections remain stable. Preserve her face, coat, umbrella, station geometry, and the single-subject composition. No new people, no text, no wardrobe change, no cuts.
Use Video References for Motion and Timing
A video reference carries more signals than a still image. It can lend body mechanics, camera speed, action timing, and edit rhythm at once. That makes it useful, but also more likely to overpower the subject or setting you meant to preserve.
In a discarded motion-transfer test, the source performer kept too much authority and the identity reference lost. The useful lesson was the failure itself: specify that the video controls only the movement or camera path, then say that its performer, clothing, and location must not transfer. If the instruction still conflicts, split the job into two generations.
The Seedance clip above could serve as a motion reference in a second pass. A new image could control the subject, while the clip contributes only the measured turn, three-step walk, and lateral tracking pace.
- Name the exact motion or camera behavior to borrow.
- State which performer, wardrobe, objects, and location must not transfer.
- Keep the requested action short enough to fit the output duration.
- Use separate passes when identity and motion instructions continue to compete.
Let Audio Set Timing and Delivery
Audio can control a performance without changing the image design. The portrait below supplies the face, wardrobe, and studio background. The Nsketch voice track supplies the spoken line, its pause, and the delivery length.
Treat music and ambience the same way. Describe whether the camera should follow the beat, whether dialogue must stay intelligible, and which sound cues need visible reactions. Avoid asking one short clip to obey several unrelated rhythms.
Nsketch Talking Photo · 6 seconds · Nsketch voice + Talking Photo
Prompt
Script: For this shot, keep the room quiet. Let the pause before the final line carry the tension. Direction: calm, measured delivery with a deliberate pause before the second sentence. Keep the portrait identity, green sweater, radio-studio background, and direct camera angle unchanged.
Combine References Without Creating a Fight
A multimodal brief works when every input has a narrow responsibility and the prompt settles priority before generation begins. Repeating the same visual idea across several files rarely adds useful control. It usually adds another interpretation the model must reconcile.
- Assign one primary role to each file.
- State which reference wins if two files disagree.
- Repeat the scene’s non-negotiable details once in the final prompt.
- Remove redundant files that do not contribute a distinct instruction.
- Build difficult transformations in stages and inspect each result before the next pass.
Choose the Nsketch Model That Matches the Reference Pack
Choose a model by the files you actually need and the duration of the shot. These limits reflect the Nsketch configuration on August 9, 2026.
| Model | Reference types | Good fit | Current Nsketch limits |
|---|---|---|---|
| Seedance 2.5 | Image, video, audio | Longer multimodal briefs and reference-heavy scenes | 4–30 seconds · 480p or 720p · native audio |
| MiniMax H3 | Image, video, audio | Compact mixed-reference shots with high-resolution output | 4–15 seconds · fixed 2K · audio cannot be the only reference |
| Gemini Omni Video | Image, video | Reference-led transformations without an audio input | 4–10 seconds · 720p to 2160p · no audio in the current Nsketch integration |
Check the Reference Pack Before Rendering
A quick review catches most reference conflicts before they consume a generation. If a file has no distinct job, remove it. If a likeness or voice is not licensed for your use, replace it before rendering.
- Can you describe each file’s job in one sentence?
- Do any two references disagree on identity, wardrobe, location, or object design?
- Can the camera move and action finish inside the selected duration?
- Do you have permission to use every likeness, voice, and source asset?
- Does the prompt say what each reference must leave unchanged?
A Reusable Multimodal Reference Brief
Write the authority of each file before the scene description. Then close with the continuity rules that apply to the whole clip.
- Image reference: controls subject identity, wardrobe, and object design. Do not copy its background unless requested.
- Video reference: controls movement, camera path, and action timing. Do not copy its performer, clothing, or location.
- Audio reference: controls voice, beat, ambience, and cue timing. Do not let it change the visual design.
- Priority: image wins for identity, video wins for motion, audio wins for delivery timing.
- Continuity: preserve subject count, wardrobe, travel direction, lighting logic, and the final end state.
FAQ
Frequently asked questions
Not by themselves. Additional files help only when each contributes a distinct instruction. Redundant or contradictory references can make identity, motion, and setting less stable.
Seedance 2.5 and MiniMax H3 currently accept all three reference types in Nsketch. Gemini Omni Video currently accepts image and video references, but not audio.
No. In the current Nsketch integration, MiniMax H3 audio references require a visual anchor such as an image or video reference.
It can, but a primary job keeps the brief easier to interpret. A portrait may also suggest lighting, for example, while identity remains its main responsibility.
Remove the lower-priority reference or state which file wins for the disputed detail. If the conflict remains, generate the identity and motion stages separately.
