Vidu Reference to Video: Consistent Characters Guide

Most video generators give you two ways in: a text prompt, or a still that becomes the first frame. Vidu (vidu.com, made by Shengshu) adds a third that solves the problem creatives complain about most. Reference-to-video takes images of the things that must stay the same, a character, a product, a location, and lets you write any scene you like around them. This guide explains how the mode differs from image-to-video, how to choose references that work, and how to prompt so the model knows what each image is for.

Reference-to-video versus image-to-video

In image-to-video the upload is the first frame. Composition, camera angle and lighting are fixed and the prompt adds motion. It is the right choice when you have an approved still and want exactly that shot to move.

In reference-to-video the uploads are ingredients, not frames. Vidu reads the identity of what is in them (a face and outfit, a bottle and label, a room) and generates a new shot from your text, placing those ingredients into it. You can ask for a different angle, a different setting, a different action, and the character or product should still be recognisable. That is what makes it suitable for multi-shot sequences: a short ad with the same presenter in three locations, or a product seen on a shelf, in a hand and on a table.

Vidu also supports plain text-to-video and standard image-to-video, with resolution and duration choices and a credit plan; several model versions exist in the current interface and names change, so check what is selected. The Vidu settings guide walks through those controls.

Choosing reference images

The model can only hold what it can clearly see. Spend time on the references and the prompting gets easier.

• One subject per image. A reference showing a person and a product together makes the model guess which you mean. Separate them.
• Clean, even light and a plain background. Dramatic lighting in the reference leaks into the output. Flat light transfers identity best.
• Front-facing and full-body for characters. A clear face plus the full outfit. If the interface allows several references for one subject, add a three-quarter view too.
• Hero angle for products. The angle that shows the label and silhouette. Labels are the first thing to degrade, so keep them simple.
• Consistent source. Generate all references for a character in the same image model and session, so colour and proportion already match before Vidu sees them.
• No text overlays, borders or watermarks. The model may try to reproduce them.

Step by step: a three-shot sequence

1. Prepare references. One character image (front, full body, plain background) and one product image (hero angle). Name the files so you can track them.
2. Open reference-to-video and upload both. Check the interface shows them as references, not as a first frame.
3. Write shot one. Name each reference’s role: "The woman from the reference holds the bottle from the reference up to the light in a bright kitchen. Medium shot, slow push in, morning window light."
4. Generate at the lowest resolution and shortest duration to confirm identity holds. Only then re-run at the settings you need.
5. Write shots two and three changing only the setting, action and camera. Keep the character and product descriptions word-for-word identical.
6. Review side by side. Line the three clips up and check face, hair, outfit colour and label. Regenerate the outlier rather than accepting a near miss.

Prompting with references

The prompt’s job changes when references are present. It no longer needs to describe the character’s appearance in detail; it needs to say which reference does what and build the scene.

• Point at the reference. "The woman in the reference image", "the bottle shown in the reference". Ambiguous prompts produce look-alikes rather than matches.
• Repeat a short identity tag anyway. "The woman (short dark hair, green jacket)" gives the model a second anchor and helps when a reference is partly occluded in the shot.
• One action, one camera move. Identity holds best with calm motion. Fast turns and running are where faces slip.
• Describe the new scene fully. Setting, light and mood come from your text now, not from a first frame, so do not leave them out.
• Keep the character large in frame. Wide shots with a tiny figure give the model too few pixels to hold a face.

For the equivalent approach in other tools, see Kling Elements and Runway Gen-4 References; the reference-image discipline is the same everywhere.

Settings at a glance

Control names may differ in the current interface and vary by model version.

Try it

References: [character: woman, short dark hair, green jacket, front view, plain background] [product: amber glass bottle with a cream label, hero angle]
Prompt: The woman from the reference image stands at a kitchen counter and lifts the bottle from the reference image towards the window, examining it. Medium shot, slow push in, soft morning light, pale wood and white tiles. Calm, natural motion.

Generate, then change only the setting to "a small garden table outdoors, late afternoon sun" and compare the two clips side by side. If the jacket colour or label drifts, re-upload cleaner references before changing anything in the prompt.