Prompting Wan 2.6
Wan 2.6 (Alibaba) is built for reference-to-video role-play — putting a specific person or character in the video with consistent identity and voice — plus multi-shot narrative and audio-visual sync. This guide shows how to prompt it. Model overview: Wan 2.6 API guide.
Reference-to-Video (Wan's Signature)
Supply a clean reference image (and voice, where supported), then describe the action. The reference sets identity; your prompt sets motion and speech:
curl -X POST https://api.linkmodel.ai/api/v1/video-generation \
-H "Authorization: Bearer $LINKMODEL_API_KEY" -H "Content-Type: application/json" \
-d '{ "model": "wan2.6", "prompt": "The person in the reference presents a product to camera and smiles: \"Here is our new release.\" Natural gestures, studio lighting.", "image_url": "https://your-cdn.com/portrait.jpg" }'Use a clean, front-facing, high-res reference for the best identity retention.
A Prompt Structure That Works
Order: who (reference) → action/speech → camera → setting/mood → audio.
The person in the reference walks along a beach at sunset,
speaks to camera: "Let's begin.",
handheld follow shot, warm golden light,
gentle waves and ambient audioMulti-Shot Narrative
Wan handles short multi-shot sequences with identity held across cuts. Describe shots in order and keep the same reference:
Shot 1: close-up, the person looks up and smiles;
Shot 2: wide shot, they walk toward the camera; consistent appearance and voiceName the language for accurate lip-sync, and keep clips to Wan's length window (up to ~15s) — chain clips for longer stories.
Tips
- One clear action or line per clip.
- Match aspect ratio to your channel (9:16 for social).
- Draft then finalize — iterate before the final render.
- Compare role-play vs all-round in Wan vs Seedance; for avatars specifically, AI avatar video API.
Consent Matters
Only use a real person's likeness and voice with their explicit consent — build consent into your workflow. Confirm parameters in the docs.
Reusable Wan Prompt Templates
These are prompt templates, not claimed output tests. Replace bracketed fields and validate them on the current Wan version.
Product motion
[subject/product] performs [one clear action] in [setting].
Camera: [shot size], [movement], [lens feel].
Lighting: [direction and quality].
Keep [logo/material/shape] consistent. No extra objects or text.Reference character with speech
Use the reference for identity. [Person] says: "[short line]" in [language].
Action: [one gesture]. Camera: locked medium shot.
Natural room tone; preserve face, clothing, and voice across the clip.Two-shot sequence
Shot 1 ([duration]): [establishing action and camera].
Shot 2 ([duration]): [continuation], same person, wardrobe, lighting, and location.
Cut on [motion cue]. Keep screen direction consistent.Failure Patterns and Fixes
| Problem | Likely cause | Rewrite |
|---|---|---|
| Identity drifts | too many characters or changes | use one clean reference and repeat continuity constraints |
| Motion looks chaotic | several actions compete | keep one action per shot and specify start/end state |
| Camera ignores the prompt | style language dominates | put shot size and movement before mood terms |
| Lip sync slips | dialogue is too long | shorten the line and state the language explicitly |
| Text mutates | video diffusion is not a typesetter | add text in post-production when exact copy matters |
Test Parameters One at a Time
Hold the prompt, reference, duration, and aspect ratio fixed while changing one control. Save the task ID, settings, latency, and reviewer score. Without that log, prompt iteration becomes guesswork and a good result cannot be reproduced.
Bottom Line
Lead with a clean reference, describe action + speech + camera, name the language for lip-sync, and chain short shots for narrative. More models in best AI video generation APIs.
Start free with a $1 credit and test your prompts.