Wan·by AlibabaImage to video
Wan-2.6 Image-to-video
A speed-optimized image-to-video option that prioritizes lower latency while retaining strong visual fidelity. Ideal for iteration, batch generation, and prompt testing.
alibaba/wan-2.6/image-to-videoParameters
| Parameter | Type | Default | Range or options |
|---|---|---|---|
duration durationThe duration of the generated media in seconds. | select | 5 | 5, 10, 15 |
resolution resolutionThe resolution of the generated video. | select | 720p | 720p, 1080p |
seed seedThe random seed to use for the generation. -1 means a random seed will be used. | number | -1 | — |
enable prompt expansion enable_prompt_expansionIf set to true, the prompt optimizer will be enabled. | select | true | true, false |
generate audio generate_audioWhether to automatically add audio to the generated video. | select | true | true, false |
shot type shot_typeGenerate video in multi camera angles, only works when set enable_prompt_expansion to true. | select | multi | multi, single |
Inputs
Files this model takes in addition to the prompt.
| Input | Accepts | Max files | |
|---|---|---|---|
Image image | image | 1 | Required |
Audio audio | audio | 1 | Optional |
Prompting guide for Wan
TARGET MODEL: Wan (Alibaba 万相 — image and joint audio-video generation, 2.5 → 3.0). Alibaba publishes the prompt as an additive formula and the model reads it in that order: Entity + Scene + Motion (basic) → + Aesthetic control + Stylization (advanced) → + Sound description (voice / sound effect / background music) → multi-shot as Overall description + Shot number + Timestamp + Shot content. No weighting syntax, no tag-soup; the video rows have no negative prompt at all (write exclusions as "No dialogue." / "No background music." style sentences), and expansion is a parameter, not a word.
Rules that must hold
- Keep the formula order: Entity (appearance — adjectives or short phrases) → Scene (environment) → Motion (amplitude, speed, effect) → Aesthetic control (light source, lighting, shot size, camera angle, lens, camera movement) → Stylization (one style word: cyberpunk, line-art illustration, claymation, felt, pixel, 3D cartoon, black-and-white animation) → Sound. Describe, never enumerate keywords.
- Audio is generated natively; write it as content. Voice = the character's lines + emotion + tone + speed + timbre + accent ("says in a relaxed tone, at a moderate speed, with a clear bright voice, in American English: "…""). Sound effect = source material + action + ambient sound ("a small glass ball falls from the table onto the wooden floor with a 'thud' in a quiet room"). Background music = score + style ("suspenseful background music"). Lines go in double quotes; on-screen text also goes in double quotes.
- Multi-character dialogue: give every speaker a UNIQUE label (never a pronoun or a synonym), anchor each line to a visible action, give each voice its own tone/emotion, and order the lines in time ("first …, then …").
- Multi-shot is written as: one overall description line (story, perspective, the subject's look — repeat the subject here, not inside every shot), then "Shot 1 [0–3 s] …", "Shot 2 [4–6 s] Hard cut transition, fixed camera position, …". Timestamps are honored; 4–6 seconds per shot; state the transition ("hard cut") and the framing of every shot. The labeled storyboard form is also official: "(0:00 - 0:03) Camera: medium shot. Scene: … Action: … Atmosphere: …".
- Name the framing and the camera treatment of every shot ("static shot" / "fixed shot" to hold still; slow push-in, pull-out, tracking shot, orbit, pan, tilt, drone fly-through). Left unnamed, Wan edits on your behalf and cuts inside a single clip. One main camera action per shot.
- Never: a real person by name · rapid scene changes inside one clip · exact text legibility as the point of the shot · very long or complex action chains · a shot that depends on word-exact lip-sync. Quoted lines are performed and synced, but do not build the beat on a specific syllable landing on a frame.
- Write in English (Chinese and English are the supported languages); a non-English line stays inside its quotes, untranslated, with the accent named in the voice description.
- Keep the whole prompt under 2,000 characters so it ports across every Wan row (the 2.6 image rows truncate at 2,000; 2.7 rows at 5,000).
What works best
- The cinematic vocabulary the guide itself uses: light source (daylight, firelight, overcast light, clear-sky light) · light type (soft, hard, side light, high contrast) · time (daytime, night, dawn) · shot size (close-up, close shot, wide angle, extreme full shot) · angle (over-the-shoulder, high angle, aerial) · shot type (clean single, two-shot, group shot, establishing shot) · composition (center, left-heavy) · lens (long-focus, ultra-wide fisheye) · tone (warm tones, low saturation) · effects (tilt-shift, time-lapse).
- Camera moves carry intent: push-in for intimacy or tension, pull-out to reveal scale or isolation, tracking to walk beside the subject, orbit to crown the subject, fixed camera for stillness and focus.
- Match beats to seconds: roughly 5–8 seconds per beat; a 5-second clip carries about 2.5 seconds of actual speech. Under-write and the model fills the gap; over-write and the delivery rushes.
- Verbs beat adjectives for sound ("the soft haptic tick of a thumb tapping the screen" over "realistic phone sounds"); place layers in depth ("foreground clink", "distant chatter", "off-camera door closing"); name the tempo when motion must sit on a beat ("driving bass at 100 BPM, each strike landing on the beat").
- Suppress by sentence, not by field: "No dialogue." "No background music." "No on-screen text, no subtitles." at the tail.
From LUVI's own manual for this family, the same rules the workspace rewriter and the MCP prompt engine apply.