Wan 2.6 API Guide: Text-to-Video, Image-to-Video, and Reference Video for Production

This guide is written for CTOs and engineers shipping generative video into real systems: async orchestration, budget guardrails, reliability patterns, and integration strategy (including a clean EvoLink.ai path at the end).

1. Wan 2.6 Model Family: Choose the Right Endpoint

• Start with T2V for concept exploration.
• Switch to I2V when you have a "source-of-truth" frame you must respect.
• Use R2V when you need identity continuity across shots/scenes.

2. The Production Workflow: Async Tasks (Not Real-Time)

Key operational details:

• You must send the async header: X-DashScope-Async: enable (DashScope HTTP mode).
• You receive a task_id and poll status until it succeeds/fails.
• task_id is valid for 24 hours (store it immediately; do not "re-submit" to recover).

• Submit task from an API worker
• Persist task_id + request hash + user/job metadata
• Poll with exponential backoff (or a scheduler/queue)
• On success, persist the returned video_url and download/replicate it (URLs are often time-limited by providers)

3. Multi-Shot Storytelling: What Actually Changes in Wan 2.6

How to enable it (T2V example)

Practical prompt guidance for multi-shot:

• Write your prompt like a short "shot list"
• Keep the main subject description consistent across shots
• Specify shot transitions ("cut to", "wide shot", "close-up") only if needed; otherwise let the model auto-segment

How it works in Wan 2.6 R2V (character references)

4. Audio: What You Can Safely Rely On

Wan 2.6 supports audio in different ways depending on the endpoint:

T2V / I2V

• Audio support includes auto voiceover or passing a custom audio file URL to achieve audio-visual sync.
• When providing a custom audio file, the platform documents practical constraints (format/size) and that audio may be truncated/left silent if it doesn't match the requested duration.

R2V

• Audio is generated via prompt, and can reference the input video's voice timbre (useful when you want continuity of voice feel).

Unless you have validated it end-to-end, avoid claiming "lip-sync" or "phoneme-accurate mouth matching." The official docs describe audio generation and audio-visual sync, but don't guarantee lip-level alignment.

5. Cost Model: Know Your Per-Second Pricing Up Front

T2V pricing (Alibaba Cloud / Bailian)

• wan2.6-t2v: 0.6 RMB/sec (720P), 1 RMB/sec (1080P)

I2V pricing (first-frame)

• wan2.6-i2v: 0.6 RMB/sec (720P), 1 RMB/sec (1080P)

Wan 2.6 R2V pricing (reference video)

• Failures are not billed
• Input video billing duration is capped (documented as "not exceeding 5 seconds" for billing)

• Dev/test default: 720P + shortest duration your UX allows
• Add server-side caps: max duration, max resolution, max jobs/user/day
• Require reference-video validation before R2V submission (format/size/duration) to reduce waste

6. Wan 2.6 Reliability Friction You'll Actually Hit

Region binding

Beijing and Singapore have independent API keys and request endpoints; mixing them can cause auth failures.

SDK gaps (I2V)

URLs and assets

Across workflows, you'll be passing media via URLs (HTTP/HTTPS), and you may need an upload step to produce temporary URLs for local files.

7. Using Wan 2.6 Through EvoLink.ai (Unified API + Clean Task Model)

• POST https://api.evolink.ai/v1/videos/generations
• Wan 2.6 models (examples):
  • wan2.6-text-to-video
  • wan2.6-reference-video
• Asynchronous processing with task IDs, and generated video links valid for 24 hours (save promptly).

Example: Text-to-Video via EvoLink

Full doc: https://docs.evolink.ai/en/api-manual/video-series/wan2.6/wan2.6-text-to-video

curl --request POST \
  --url https://api.evolink.ai/v1/videos/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "wan2.6-text-to-video",
    "prompt": "A cinematic multi-shot sequence of a runner crossing a neon-lit city bridge at night, rain reflections, dramatic camera cuts, realistic motion."
  }'

Example: Reference Video via EvoLink (copy-paste)

curl --request POST \
  --url https://api.evolink.ai/v1/videos/generations \
  --header 'Authorization: Bearer YOUR_API_KEY' \
  --header 'Content-Type: application/json' \
  --data '{
    "model": "wan2.6-reference-video",
    "prompt": "character1 walks into a bright cafe, orders a drink, then turns and smiles to camera; multi-shot narrative.",
    "video_urls": [
      "https://your-cdn.example.com/reference_character.mp4"
    ]
  }'

This endpoint accepts up to 3 reference videos and documents requirements like format (mp4/mov), file size (≤100MB), and duration range (2–30s).

8. Ship Wan 2.6 Faster

If you're building production video features—UGC creation tools, marketing automation, product visualization, or storyline generation—the hard part isn't "can the model generate video?" The hard part is operationalizing it: task orchestration, spend control, and evolving model/provider choices over time.

• One API surface for Wan 2.6 (and other video models as you expand your stack)
• A clean async task pattern you can standardize in your backend
• A practical path to reduce integration churn when providers update parameters or add new endpoints

9. FAQ (Production Notes)

1) What durations does Wan 2.6 support for each mode?

• Text-to-Video (wan2.6-t2v): 5 / 10 / 15 seconds
• Image-to-Video (wan2.6-i2v): 5 / 10 / 15 seconds
• Reference Video (wan2.6-r2v): 5 / 10 seconds

2) Can I bring my own audio? What are the constraints?

• Formats: wav / mp3
• Duration: 3–30 seconds
• Size: ≤ 15MB
• If audio is longer than the requested video duration, it is truncated; if shorter, the remaining video is silent.

3) How do I force silent output (no auto audio)?

4) What are safe prompt length limits?