VidoGridHow Wan 2.2 prompts are put together
Alibaba's guide builds a prompt from three parts: subject, scene and motion. For finished shots it adds aesthetic control (light, shot size, angle, lens, camera movement) and a visual style. Each part works best as concrete adjectives and short phrases rather than abstract language.
The official prompt rewriter in the GitHub repo opens with a few comma-separated cinematic settings, then describes the subject, action and background in plain sentences, aiming for roughly 60 to 200 words. A named style such as claymation goes first. On Model Studio, Wan 2.2 prompts longer than 800 characters are truncated.
Soft lighting, side lighting, close-up shot, cool colors. A matte black steel water bottle stands on wet slate beside two sprigs of mint. Beads of condensation slide slowly down its side. A thin trail of cold mist drifts across the frame from left to right. The camera pushes in slowly toward the brushed cap. Plain charcoal background.
Lighting and framing vocabulary
Wan 2.2 was trained on aesthetic data labeled for lighting, composition, contrast and color tone, so naming those qualities directly pays off in wan 2.2 prompts. The official English rewriter adds no more than four such terms and defaults to a medium or wide shot with center composition. It picks from groups like these:
- Time and light source: day, night, dawn, sunrise; daylight, moonlight, firelight, practical or overcast lighting
- Light quality and angle: soft or hard; top, side, under or edge lighting
- Color: warm, cool or mixed colors
- Shot size and angle: extreme wide shot to extreme close-up; low, high, aerial, Dutch or over-the-shoulder
- Composition: center, balanced, symmetrical, left-heavy, right-heavy
Dusk, practical lighting, warm colors, medium close-up shot. An elderly watch repairer with a grey beard and wire glasses sits at a cluttered workbench under a brass desk lamp. He lifts a tiny gear with tweezers, squints at it, then sets it gently into an open pocket watch. Clocks on the shelves behind him are softly out of focus. Fixed camera.
Writing motion and camera movement
The official rewriter spells out how each action unfolds, step by step, and adds small background movement such as drifting clouds when a scene would otherwise sit still. Speed and size words help: slowly, suddenly, a slight turn of the head.
Write camera moves as sentences: the camera pushes in, pulls back to reveal the room, moves left or stays fixed. When you do, drop the separate angle tag, as the rewriter does. Keep orbits under 45 degrees; the guide warns that wider arcs can distort the space.
Sunrise time, daylight, extreme wide shot, warm colors. A winding river cuts through a valley of birch trees in yellow autumn leaves. Low fog lifts off the water and drifts downstream. A line of geese crosses the frame from right to left. The camera flies forward high above the river at a slow, steady pace. Clear deep-blue sky.
Image-to-video and first and last frames
With a start image, the picture already fixes the subject, setting and style, so Alibaba's formula shrinks to motion plus camera movement. Describe what changes, not what is visible; the official image-to-video rewriter stays under 100 words and keeps your camera directions. An empty prompt also works when prompt extension lets a vision model write one. In the open-weights code, size sets the pixel area and the image sets the aspect ratio.
For first and last frames (wan2.2-kf2v-flash on Model Studio), the video follows the first frame's aspect ratio. When the two frames differ a lot, the docs suggest describing the transition, such as a camera move.
The barista tilts a steel milk pitcher and pours a leaf pattern into the coffee. Steam rises from the machine behind her. She sets the pitcher down, looks up and smiles at someone off camera. The camera slowly moves right, then holds still.
Settings that change results
These parameters shape the output as much as the wording of your wan 2.2 prompts.
- Length: hosted Wan 2.2 models return fixed five-second clips. The open-weights A14B models default to 81 frames at 16 fps and TI2V-5B to 121 frames at 24 fps; frame counts must be 4n+1.
- Resolution: open-weights A14B takes 1280*720 or 832*480 in either orientation, and TI2V-5B uses 1280*704 or 704*1280. Hosted wan2.2-t2v-plus offers 480P and 1080P sizes.
- Negative prompt: left blank, the open-weights code applies a built-in list against oversaturated color, overexposure, static frames, blur, captions, extra fingers and crowded backgrounds. Model Studio accepts up to 500 characters.
- Prompt rewriting: prompt_extend is on by default on Model Studio, and the actual_prompt field shows the rewritten text. Turn it off when your wording must stay exact.
- Seed: a fixed seed makes runs more repeatable, though not identical.
Sound, speech and later Wan versions
Wan 2.2 text-to-video and image-to-video output is silent, so dialogue written into wan 2.2 prompts produces no audio. Speech uses a separate model, Wan2.2-S2V-14B, which takes a reference image, an audio clip and an optional text prompt; clip length follows the audio.
Later hosted versions differ. From Wan 2.5, prompts can describe voices, sound effects and background music, with a 1,500-character limit on 2.5 and 2.6. Wan 2.6 and later accept numbered shots with time ranges such as 'Shot 1 [0-3 s]'. Adjust prompts when moving between versions.
Common mistakes in Wan 2.2 prompts
Most weak results trace back to a few habits.
- Mood words such as 'full of hope': the official rewriter leaves these out. Show the light, expression or gesture instead.
- Several scenes in one clip: each generation is one continuous shot, so split sequences.
- Long choreography: keep actions simple and short.
- Exact on-screen text: lettering renders approximately, so add it in editing if it must be right.
- Real people by name: results are inconsistent and often refused. Describe the appearance instead.
- Blown-out skies: the rewriter turns sky descriptions into a clear deep-blue sky to avoid overexposure.