Wan 2.2 Image-to-Video Prompt Guide: Patterns That Work

Wan 2.2 image-to-video takes one still image and a text prompt and returns a 5-second silent video at 480p or 1080p. The image decides what everything looks like; the prompt decides only what moves and how the camera behaves. Most bad Wan 2.2 results come from prompting it like a text-to-image model — describing the scene the model can already see instead of the motion it cannot. This guide covers the prompt patterns that work, the parameters around them, and the exact API call, at a flat $0.20 per video on Unifically.

TL;DR: In i2v mode, prompt the motion, not the scene: subject action first, then camera movement, then pace. Keep the prompt under a few sentences, push style words into the image itself, use negative_prompt for artifacts you keep seeing, and pin a seed once a result works. prompt_extend is on by default and rewrites short prompts for you; turn it off when you need literal control. Every video is 5 seconds, silent, 480p or 1080p, $0.20 flat. Run it in the playground.

How Wan 2.2 image-to-video works

You send mode: "i2v", a start_image_url, and a prompt to the same endpoint that handles text-to-video. The image becomes the first frame, and the model generates 5 seconds of motion from it. Resolution is 480p or 1080p, output is always silent, and duration is fixed at 5 seconds. There is no length knob to prompt around, so every prompt should describe one continuous beat of motion, not a sequence of shots.

That fixed window shapes everything below: a 5-second video has room for exactly one camera move and one subject action. Prompts that ask for three events produce rushed, smeared motion.

The prompt pattern: motion first

The image already carries the subject, lighting, style, and composition. Repeating them wastes the prompt. Describe, in order:

1. Subject action: what moves, with a concrete verb. "The woman turns her head toward the window", "steam rises from the cup", "waves roll in and break".
2. Camera behavior: one move. "Slow push-in", "handheld drift left", "static camera". Naming a static camera matters; without it the model often invents a drift.
3. Pace and atmosphere: "slow and calm", "sudden gust", "gentle loop-like motion".

A working i2v prompt reads like direction to a cinematographer who is already looking at the still:

"The cat slowly opens its eyes and stretches one paw forward, gentle handheld camera with a slight push-in, calm morning pace."

Compare the common failure mode: "a beautiful orange cat in warm morning light on a windowsill, cozy atmosphere, 4K detailed fur". Every word describes what the image already shows, none of it describes motion, and the model is left to guess.

Two more rules that consistently pay off:

• One subject, one action. "The dancer spins once as confetti falls" beats "the dancer spins, jumps, lands, and the crowd cheers".
• Anchor the end state. "until the door is fully open" or "coming to rest at the center" gives the model a target for frame 120 and reduces drift in the last second.

Negative prompts: fix repeat offenders

negative_prompt tells Wan 2.2 what to avoid. Use it reactively, not preemptively: run the prompt clean first, then add the specific artifact you saw. Terms that earn their place for i2v work:

• "warping, morphing, deformed hands, extra fingers" when people or hands distort mid-motion
• "camera shake, jitter" when a static shot will not hold still
• "text, watermark, subtitles" when overlays appear
• "fast motion, time-lapse" when the model rushes a calm scene

A short, targeted negative prompt outperforms a pasted 40-term blocklist, which mostly dilutes the signal.

prompt_extend: the rewriter is on by default

Wan 2.2 sets prompt_extend: true by default: an intelligent rewriter expands your prompt before generation. For short, casual prompts it usually helps: "she smiles" becomes a fuller motion description. Turn it off when:

• you have written a precise, complete prompt and want it followed literally,
• you are iterating with a fixed seed and want only your own wording to change between runs,
• the rewriter keeps adding motion you then have to negative-prompt away.

With prompt_extend: false, what you write is what runs, which makes prompt debugging deterministic.

Seeds: lock what works

Pass an integer seed to make runs repeatable. The productive loop: iterate seedless until a result is close, read the seed back from the task, then pin it and refine wording one change at a time. With a pinned seed and prompt_extend off, each edit maps to a visible change instead of a re-roll.

The full API call

curl -X POST https://api.unifically.com/v1/tasks \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "alibaba/wan-2.2-video",
    "input": {
      "mode": "i2v",
      "start_image_url": "https://files.example.com/windowsill-cat.jpg",
      "prompt": "The cat slowly opens its eyes and stretches one paw forward, gentle handheld camera with a slight push-in, calm morning pace",
      "negative_prompt": "warping, camera shake, text",
      "resolution": "1080p",
      "prompt_extend": false,
      "seed": 42
    }
  }'

Submission returns a task_id; poll GET /v1/tasks/<task_id> or set a callback_url. The start_image_url must be a reachable image URL (up to 100 MB). Text-to-video is the same call with mode: "t2v" and no image.

Wan 2.2 pricing and where it fits

Wan 2.2 costs a flat $0.20 per video at either resolution on Unifically, pay-per-use with no subscription. That flat rate makes 1080p the default choice: it costs nothing extra over 480p. Draft at 480p only when queue speed matters more than sharpness.

If the brief needs sound or longer output, step up the family: Wan 2.5 adds audio and 5–10 second durations, and Wan 2.6 reaches 15 seconds with reference-to-video. For prompt-sweeping the same image at volume, Wan 2.2 Fast runs the same generation cheaper. A new account's $0.20 free balance covers one full Wan 2.2 video, enough to test your first i2v prompt before topping up.

Frequently asked questions

How do I write a good Wan 2.2 image-to-video prompt?

Describe the motion, not the scene: one subject action with a concrete verb, one camera move, and the pace, in two or three sentences. The input image already fixes the look; the prompt's only job is what happens during the 5 seconds.

Does Wan 2.2 support negative prompts?

Yes. Pass negative_prompt with the specific artifacts you want to avoid ("warping, camera shake, text"). Add terms reactively after you see a problem; short targeted lists work better than long generic ones.

How long are Wan 2.2 videos?

Every Wan 2.2 video is 5 seconds, silent, at 480p or 1080p. For 5–10 seconds with audio use Wan 2.5; for up to 15 seconds use Wan 2.6.

What does prompt_extend do in Wan 2.2?

It rewrites your prompt with more detail before generation, and it is on by default. Keep it on for short prompts; set prompt_extend: false when you want your exact wording followed or when you are iterating with a pinned seed.

We will update this guide when Wan 2.2's parameters change, and the pattern advice as we run more head-to-head prompt tests. For the model's full spec and variants, see the Wan 2.2 model page.