Skip to main content
Wan 2.2 Image-to-Video Prompt Guide: What Works (2026)
Guides

Wan 2.2 Image-to-Video Prompt Guide: What Works

How to prompt Wan 2.2 image-to-video: motion-first patterns, negative prompts, prompt_extend, seeds. Wan 2.2 retired October 24, 2026; requests move to Wan 2.6.

UnificAlly Model Research TeamHow we research
15 min readUpdated November 16, 2026

Retired: the Wan 2.2 route is no longer available on UnificAlly as of October 24, 2026. Every prompt pattern below carries over to the live Wan models (Wan 2.6, Wan 2.7 and Wan 3.0); the last sections show which request fields change.

Wan 2.2 image-to-video turns one still into a fixed 5-second silent video, and the prompt only has to supply motion and camera: Alibaba's own formula is literally "Motion + Camera movement". Two limits shape every prompt: an 800-character cap, and a rewriter (prompt_extend, on by default) that deletes scene description and trims your text to 100 words.

TL;DR: Write one subject action with a speed word, then one camera move ("fixed camera" if it should not move). Leave out what the image already shows, because the default-on rewriter strips it anyway. Keep prompt_extend on while drafting, then turn it off for the final wording. Use negative_prompt (500 characters max) for artifacts you have seen, not as a 40-term blocklist. A pinned seed makes runs repeatable in practice, but Alibaba does not promise identical output. For new work, Wan 2.6 accepts a Wan 2.2 request almost field for field.

Wan 2.2 prompt guide in five rules

  • Motion and camera are the whole prompt. Alibaba's image-to-video formula has exactly two parts, and the image supplies subject, setting and style.
  • Scene description gets deleted. The open-source Wan 2.2 rewriter is told to remove static details already visible in the image and to keep the result under 100 words.
  • There is a hard length cap. Alibaba's API cuts Wan 2.2 prompts at 800 characters and negative prompts at 500, without an error.
  • "Fixed camera" is the documented phrase for no movement. Leave the camera unnamed and the model picks a move for you.
  • Seeds are repeatable, not guaranteed. Alibaba's reference says the same seed can still produce a different result. We have not measured how often that happens on Wan 2.2.

Wan 2.2 image-to-video limits that shape the prompt

Every number below comes from Alibaba's Wan 2.2 image-to-video API reference, the open-source repo, or our own retired Wan 2.2 docs page; the source type sits in brackets.

LimitWan 2.2 image-to-video
DurationFixed 5 seconds, no duration control (official, our docs)
Resolution480p or 1080p on UnificAlly; default was 480p (our docs)
Frame rate30 fps for Alibaba's hosted wan2.2-i2v-plus; 16 fps and 81 frames for the open A14B weights (official)
AudioNone; silent output only (official, our docs)
Prompt length800 characters, longer text truncated (official)
Negative prompt length500 characters, truncated (official)
Input imageJPEG, PNG without alpha, BMP or WEBP; each side 240 to 8,000 px; up to 10 MB on Alibaba's API (official)
Aspect ratioTaken from the input image, rounded so width and height are multiples of 16 (official)
Prompt rewritingprompt_extend, default on (official, our docs)
Seed range0 to 2,147,483,647 (official)

Each Wan 2.2 image-to-video request is a single 5-second shot from one start frame, so the prompt has room for one action and one camera move, and anything past 800 characters never reaches the model.

How to prompt Wan 2.2 image to video: motion, speed, camera

Alibaba's Wan video prompt guide splits text-to-video and image-to-video. Text-to-video needs subject, scene, motion, lighting and style. For image-to-video the image already settles subject, scene and style, so the formula shrinks to motion plus camera. Speed words such as "slowly" or "quickly" set how hard the motion plays.

Build Wan 2.2 image to video prompts (Wan 2.2 i2v prompts, in community shorthand) in that order:

  1. The action, with a concrete verb and a speed. "The barista slowly pours milk into the cup" gives the model a start, a direction and a pace.
  2. The camera, as one move. "Camera pushes in", "camera moves left", or "fixed camera".
  3. Optional: one secondary motion. Steam, hair in wind, a curtain. One is plenty for 5 seconds.

A portrait of a fisherman on a dock, prompted three ways:

PromptWhat goes wrong or right
"An old fisherman with a grey beard on a wooden dock at sunset, warm light, detailed skin, 4K"Every word restates the image. There is no verb, so the model guesses the motion, and the rewriter would delete most of it.
"The fisherman turns, waves, picks up his net, walks away, and the boat leaves"Five events in 5 seconds. Expect rushed, smeared motion.
"The fisherman slowly lifts his head and squints toward the water, fixed camera, his coat moving slightly in the wind"One action, one speed word, one camera instruction, one secondary motion.

Keep each Wan 2.2 image to video prompt to one beat of motion. A 5-second shot at 30 fps holds about 150 frames, enough for a head turn or a pour, not for a scene change.

Wan 2.2 prompt extend: what the rewriter does to your prompt

prompt_extend defaults to true on Wan 2.2, and Alibaba describes it as an LLM rewriting your input before generation. Alibaba does not publish the hosted rewriter's instructions. The Wan team does publish the ones it uses in the open-source repo, in wan/utils/system_prompt.py, and they explain most of the "describe motion, not the scene" advice you will read elsewhere.

The image-to-video rewriter is told to:

  • keep the moving parts of your prompt and center them on the main subject's action;
  • attach a bare action ("dancing") to the subject it sees in the image;
  • shorten long prompts down to the core action, and pad short ones with plausible motion from the image;
  • keep camera phrases such as "the camera moves left";
  • remove static descriptions of things already visible in the image;
  • stay under 100 words and answer in English, whatever language you wrote in.

A second rewriter handles an empty prompt: it invents the motion from the image alone, under the same 100-word limit. The text-to-video rewriter works the other way. It adds up to four cinematic settings (time of day, light source, shot size, composition) and aims for 60 to 200 words.

So with prompt_extend on, a careful paragraph about lighting and wardrobe is mostly thrown away before generation. With it off, every character you send is used, up to the 800 cap. Alibaba's own API returns the rewritten text in an actual_prompt field when rewriting is on, which is the fastest way to see what the model was really given.

Our call: leave prompt_extend on for the first drafts, especially with three- or four-word prompts, and read the rewrite where your host returns it. Switch it off once the wording is settled, so a change you make is the only change between runs. Two of the top-ranking Wan 2.2 prompt guides also claim that the model pays more attention to the start of a prompt; we found no Alibaba or Wan-team source for that, so we would not build a prompt around it.

Wan 2.2 camera prompts that hold up in 5 seconds

The camera phrases that appear in Alibaba's guide and in the official rewriter examples are short and literal: "camera pushes in", "camera pulls out", "camera moves left", "orbiting camera movement", "fixed camera". The guide attaches an intent to each one: a push-in builds closeness or tension, a pull-out reveals scale, a tracking shot puts the viewer beside the subject, an orbit makes the subject feel important, and a fixed camera signals stillness.

Two documented constraints matter for a 5-second image-to-video shot:

  • Keep an orbit under 45 degrees. Alibaba's guide warns that wider orbits distort space. From a single start frame the model has to invent the far side of the subject, so a half-circle is asking for trouble.
  • Name "fixed camera" when you want stillness. Leaving it out does not mean static; the motion formula expects you to state the camera.

Sequenced moves ("pushes in, then pans right") appear in the guide's text-to-video examples, not its image-to-video ones. For Wan 2.2 camera prompts, one move per video is the safer default.

The Wan 2.2 negative prompt the Wan team ships by default

The open-source Wan 2.2 code sets a default negative prompt (sample_neg_prompt in wan/configs/shared_config.py), written in Chinese. Translated, it asks the model to avoid:

  • garish colors, overexposure, an overall grey cast, blurry detail, JPEG artifacts, worst and low quality;
  • static frames, a still picture, motionless images;
  • subtitles, and anything that reads as "a painting" or "an artwork";
  • ugly, mutilated or deformed bodies, extra or fused fingers, badly drawn hands and faces, malformed limbs, three legs;
  • cluttered backgrounds, crowds of people in the background, and people walking backwards.

That list doubles as a map of Wan 2.2's known failure modes. Three entries are about the image refusing to move, which is the most common complaint with image-to-video. "Walking backwards" and "three legs" are motion errors on people. "Many people in the background" is the model adding extras.

How to use negative_prompt on a hosted Wan 2.2 call:

  • Run the prompt clean first, then add the two or three artifacts you saw: "deformed hands, extra fingers" for hand drift, "camera shake" for a static shot that wobbles, "subtitles, text" for stray captions, "static, motionless" when the subject barely moves.
  • Stay well under 500 characters. Alibaba's API cuts the rest silently, so a long pasted list can lose its most useful terms at the end.
  • Self-hosting with a distilled few-step LoRA at guidance scale 1.0? Classifier-free guidance is off at that setting, so the negative prompt has no effect at all. Raise the guidance scale or drop the negative prompt.

Six Wan 2.2 image-to-video failures and the documented fix

SymptomLikely causeFix
The subject barely movesNo verb, or only scene words, in the promptOne concrete action plus a speed word; "static, motionless" in negative_prompt
The camera drifts when you wanted stillnessCamera left unnamedAdd "fixed camera"
The end of a long prompt is ignored800-character cap, or the rewriter trimming to 100 wordsCut to the action and camera; turn prompt_extend off for literal wording
Output is not the aspect ratio you plannedRatio comes from the input image, rounded to multiples of 16Crop the start image to the target ratio before sending
An orbit warps the subject or backgroundArc wider than about 45 degreesSmaller orbit, or a push-in instead
Same seed, slightly different videoAlibaba says generation stays probabilistic with a fixed seedCompare runs side by side; do not assume byte-identical output

Each row comes from Alibaba's API reference, its prompt guide or the open-source config, not from our own generation tests.

The Wan 2.2 image-to-video API call, and the same request on Wan 2.6

This is the request as the Wan 2.2 API took it before retirement. The parameters match our retired docs page, where resolution defaulted to 480p, so 1080p had to be asked for.

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/fisherman-dock.jpg",
      "prompt": "The fisherman slowly lifts his head and squints toward the water, fixed camera, his coat moving slightly in the wind",
      "negative_prompt": "deformed hands, camera shake, subtitles",
      "resolution": "1080p",
      "prompt_extend": false,
      "seed": 42
    }
  }'

A Wan 2.2 task came back as a task_id with status processing; the finished 5-second MP4 then sat in output.video_url on GET /v1/tasks/<task_id>.

The closest drop-in today is Wan 2.6 on UnificAlly: it still reads start_image_url, negative_prompt (500 characters), prompt_extend and seed. Two things change. Wan 2.6 starts at 720p, and it adds an audio track unless you send "audio": false.

curl -X POST https://api.unifically.com/v1/tasks \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "alibaba/wan-2.6-video",
    "input": {
      "mode": "i2v",
      "start_image_url": "https://files.example.com/fisherman-dock.jpg",
      "prompt": "The fisherman slowly lifts his head and squints toward the water, fixed camera, his coat moving slightly in the wind",
      "negative_prompt": "deformed hands, camera shake, subtitles",
      "resolution": "720p",
      "duration": 5,
      "audio": false,
      "prompt_extend": false,
      "seed": 42
    }
  }'

Moving Wan 2.2 prompts to Wan 2.6, 2.7 or 3.0: what changes

FieldWan 2.2 (retired)Wan 2.6Wan 2.7Wan 3.0
Start framestart_image_urlstart_image_urlfirst_frame_url (+ last_frame_url)first_frame_url (+ last_frame_url)
Prompt cap800 characters (official)Not listed in our docs5,000 characters20,000 characters
negative_promptYesYes, 500 charactersYes, 500 charactersNot listed in our docs
prompt_extendDefault onDefault onDefault onNot listed in our docs
DurationFixed 5 s2 to 15 s2 to 15 s2 to 30 s
Resolution480p, 1080p720p, 1080p720P, 1080P480P, 720P, 1080P
AudioNoneOn by defaultOptional driving audioOn by default

(Our docs pages for each model, checked 2026-11-16.)

A Wan 2.2 prompt moves to Wan 2.6 with one added field ("audio": false) and to Wan 2.7 with one renamed field. Wan 3.0 is the bigger change: its docs list no negative_prompt and no prompt_extend, so any artifact you used to block with a negative prompt has to be handled in the positive prompt instead. With up to 30 seconds per video on Wan 3.0, the one-action rule relaxes too. Alibaba's guide adds a separate multi-shot formula for Wan 2.6 and later.

What Wan 2.2-style prompt drafts cost on the live Wan models

Wan 2.2 billed one flat rate per 5-second video at either resolution, so drafting at 480p saved nothing. The live models bill per second, and resolution changes the price, which makes a draft-then-final loop worth planning:

  • Ten 5-second drafts on Wan 3.0 at 480p: $2.125. The same ten on Wan 2.6 at its lowest tier, 720p: $4.50.
  • One 5-second final at 1080p: $0.85 on Wan 3.0, $0.675 on Wan 2.6.
  • The $0.50 signup balance covers 11 seconds of Wan 3.0 at 480p, enough to try a few prompt variants before topping up.

Our pick for prompt work today: draft on Wan 3.0 at 480p when you can live without a negative prompt, and stay on Wan 2.6 when a Wan 2.2 prompt leans on negative_prompt and you want it to run unchanged.

Wan 2.2 Image-to-Video FAQ

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

Describe what moves and how the camera behaves: one action with a concrete verb and a speed word, then one camera move such as "camera pushes in" or "fixed camera". Skip the subject's looks and the setting, because the image already fixes them and the default rewriter removes them.

Does Wan 2.2 support negative prompts?

Yes. The model takes a negative_prompt of up to 500 characters on Alibaba's API, and the open-source code ships a default one covering static frames, deformed hands, extra limbs and background crowds. Add only the artifacts you actually see; a long list can be truncated.

How long are Wan 2.2 videos?

Every Wan 2.2 video is 5 seconds, silent, at 480p or 1080p, with no duration control. For longer output on UnificAlly, Wan 2.6 and Wan 2.7 run 2 to 15 seconds, and Wan 3.0 runs 2 to 30 seconds with audio on by default.

What does prompt_extend do in Wan 2.2?

It sends your prompt through an LLM rewriter before generation, and it is on by default. In the open-source version, the image-to-video rewriter keeps the action and camera moves, deletes static scene description and caps the text at 100 words. Turn it off for literal wording.

Does Wan 2.2 support reference-to-video?

No. Wan 2.2 only does text-to-video and image-to-video from a single start frame. Reference-to-video, which takes reference images or videos of a character, arrived in Wan 2.6, and Wan 2.7 and Wan 3.0 support it as well.


Every pattern in this Wan 2.2 image to video prompt guide runs today on the Wan 2.6 API, the Wan 2.7 API and the Wan 3.0 API; the retired Wan 2.2 page keeps the original spec.

Last updated: November 16, 2026

Continue reading

Related to Wan 2.2 Image-to-Video