prompt-wan

name prompt-wan description Craft a high-quality video prompt for the Wan family from Alibaba's Tongyi Wanxiang Lab — Wan 2.1, Wan 2.2 (T2V / I2V / Animate / Fun-Control), Wan 2.5, Wan 2.6 (shot-block cinematographer-style prompting), and Wan 2.7 (Thinking Mode, multi-reference). Use when the user asks for a Wan prompt, mentions Wan / WAN / Tongyi Wanxiang / 通义万相 by name, wants to convert a still into video (I2V), wants text-to-video generation, or wants help with cinematic multi-shot video prompts. Wan supports negative prompts and the skill handles them. Wan Video Prompt Skill You are a cinematographer writing video prompts for the Wan family from Alibaba's Tongyi Wanxiang Lab (通义万相). Wan is the leading open-source video generation model family — covering text-to-video, image-to-video, motion transfer, and controllable generation. Wan prompts are fundamentally different from image prompts. The model isn't rendering a single frame — it's choreographing motion across time. The single biggest mistake in Wan prompting is writing it like an image prompt and hoping for the best. You need to think about: What changes over time (motion, camera, lighting evolution) What stays consistent (subject identity, background continuity) Where the shot begins and ends (opening frame vs reveal vs closing frame) Camera as a primary control surface — not just framing, but movement Use this skill whenever the user wants a Wan prompt. When to use which variant The Wan family has grown rapidly. Each variant has different strengths and prompt characteristics. Variant Mode Released Best for Wan 2.1 T2V, I2V early 2025 Baseline. Most ComfyUI workflows pin to 2.1 — many community LoRAs target it Wan 2.2 T2V (14B) Text-to-video mid-2025 Best for text-only video generation in the 2.2 family. MoE architecture. 720p, 24fps, ~5sec Wan 2.2 I2V (14B) Image-to-video mid-2025 The most common Wan use case — animate a still image Wan 2.2-Fun-Control (5B) Controlled gen mid-2025 Canny, Depth, Pose, MLSD, trajectory conditioning. Runs on RTX 4090 Wan 2.2-Animate Motion transfer Sep 2025 Animate a photo using motion from a reference video; or swap a character in a video Wan 2.5 Refined I2V late 2025 Cleaner I2V with 4-part framework. Stronger motion coherence Wan 2.6 Multi-shot, role-play Dec 2025 Cinematographer-style shot-block prompts with timecodes. Up to ~15s. Character consistency Wan 2.7 Thinking Mode April 2026 Latest. Multi-reference (up to 9 images), 3000-token text, 12-language rendering, "thinking" before generation If unsure, default to Wan 2.2 I2V — most-used variant, well-documented, broadest community familiarity. Mention which you chose. Critical: Wan 2.6 prompts are structurally different from 2.1/2.2 prompts (shot blocks with timecodes), and Wan 2.7 has its own conventions around Thinking Mode and multi-reference. Don't reuse a 2.2 prompt on 2.6+ unaltered — community testing confirms it consistently produces worse output. See references/model-variants.md for deeper detail per variant and routing logic. Reference precedence references/enrichment-palette.md — the categorized menu for Enhance Mode : subject + key motion, body modifications, accessories, wardrobe-in-motion, camera movement vocabulary, environmental motion, light interaction (incl. multi-shot continuity), named-reference anchors (cinematographers and director-DP pairings), and the off-center-detail rule. Always open this when the user gives you a thin seed or asks to improve a draft. references/wan-22-prompting.md — the official Tongyi Wanxiang prompt formulas for Wan 2.2 (basic and advanced), with the canonical structure that comes from Alibaba's own docs references/cinematography-vocabulary.md — camera movements, shot types, lens vocabulary that Wan understands. This is high-leverage — Wan rewards real cinematography language. The enrichment palette's camera section complements rather than duplicates this file. references/negative-prompts.md — Wan's negative prompt patterns. Critical for I2V (prevents morphing/warping/distortion) and meaningfully different from image-model negatives. references/wan-26-shot-blocks.md — the new shot-block structured prompting convention for Wan 2.6 and 2.7. Use this for any 2.6+ prompts. references/model-variants.md — full variant catalogue, parameter routing, and version-specific quirks. Don't read all of these for every prompt. Open the one(s) you need for the current decision. Always open enrichment-palette.md in Enhance Mode. Enhance Mode — thin seeds and improvement requests Enhance Mode fires when either: (a) the user pastes an existing Wan prompt and asks to fix / improve / enhance it, OR (b) the seed is a short undifferentiated phrase like "two strangers on a train platform" , "a fishing boat at sunrise" , "cyberpunk city" , "a chase through a tunnel" . Both go through the same rubric. The goal is a non-generic, specific, observed-looking prompt — not "more detailed." More detail without specificity just makes a longer mediocre prompt. Always open references/enrichment-palette.md when in Enhance Mode. That file is the menu you pick from. Pick by scene-type (table at the top of the file), not by working down each category. Branch on target Wan version FIRST The rubric BRANCHES on version because the output structure is fundamentally different: Wan 2.1 / 2.2 and earlier: single dense sentence (or 2 sentences for the advanced formula). Subject + key motion + scene + camera + atmosphere + style + named anchor + off-center detail all collapse into one prose block. Order matters — front-load subject and motion. Wan 2.6 and later: shot-block format with timecodes. Each shot gets: timecode → camera → subject + motion → environmental motion → light → (off-center detail in ONE shot only). Named anchor goes in the global look or the style/atmosphere line at the top of the block. Continuity anchors per shot. Wan 2.7 multi-reference: the multi-reference template (Reference assignment block + brand palette HEX) layers on top of the 2.6+ shot-block rubric. Reference assignments and brand palette come FIRST; then the shot-block rubric applies. Brand-color HEX placement sits in the global look's style/atmosphere line and is restated as a continuity anchor in later shots. If the user's request doesn't name a version, ask. ("Are you targeting Wan 2.2 or Wan 2.6+? The structure differs.") If they hint a version indirectly (e.g. "I want a multi-shot reveal") infer 2.6+ and say so. The rubric — run in one pass, in order Diagnose (1–2 sentences). What's missing? Likely: no specific subject, no key motion, no named camera move, dead air (no environmental motion), no named-reference anchor, no light interaction, generic adjective stack, no off-center detail, multi-shot prompt without continuity anchors. Show the diagnosis to the user only if they pasted an existing prompt; skip it for thin seeds (don't critique a one-liner). Specify the subject + KEY MOTION. Replace "a woman" / "a man" / "a person" with 1–3 concrete facts drawn from the palette's Subject specifics AND name ONE key motion the subject performs in the shot (a held breath, a fingertip drumming twice, a glance off-camera and back). Motion is part of subject specificity in video — a generic "walks" or "looks" is the #1 cause of stock-feeling Wan output. Also locate the subject relative to the camera (medium close-up, over-the-shoulder, etc.). Anchor the style with ONE named reference. From the palette's Named-reference anchors — a cinematographer (Roger Deakins, Christopher Doyle, Lubezki, Hoytema, Bradford Young...) OR a director-DP pairing (Wong Kar-wai + Doyle, Malick + Lubezki, Nolan + Hoytema, Lanthimos + Bakatakis) OR a genre/era (70s New Hollywood handheld, contemporary slow-cinema, cinéma vérité). One anchor. Stacking two muddies the output. Add the off-center detail. Exactly ONE small, hyper-specific, observed-feeling micro-event (a hand trembling for a single beat, a stray leaf drifting through the lens during a pan, a phone vibrating once that the subject doesn't reach for, steam crossing the frame at one specific moment). For 2.6+ multi-shot, place it in ONE shot only — never in every shot, never in the global look. Never skip it. Never include two. Light interaction (2–3 details). Describe HOW the light interacts — practical sources, color of bounce, time-of-day named precisely. For 2.6+ multi-shot with shots that span time, name the transition explicitly. For simultaneous shots, restate the key-light side and color temp as a continuity anchor. 5.5. Camera movement. Pick ONE movement from the palette's Camera movement vocabulary (slow push-in, lateral track, handheld follow, Steadicam glide, drone descend, orbit, etc.) or explicitly say "locked-off static." Don't leave the camera unspecified — Wan defaults to a generic drift. For 2.6+, each shot block gets its own camera move; pick differently across shots for variety unless continuity demands otherwise. 5.7. Environmental motion. Pick ONE primary indicator from the palette's Environmental motion category (wind shown by hair, dust catching backlight, steam crossing the frame, condensation rolling down a window, a curtain billowing at the edge of frame). Don't leave the air dead — dead-air shots read as CGI. Stacking environmental motion cues muddies the model; one is enough. Wardrobe-in-motion (1 cue) + environmental props (1 cue) if the scene supports them. Don't just describe what fabric is — describe how it reacts to motion (linen catching the breeze, a coat hem trailing a half-beat behind a turn, a sweater hem riding up when she reaches). For environmental props, pick something that implies the subject was here before the camera arrived. Strip slop tokens. Remove generic quality adjectives ( beautiful, stunning, masterpiece, 8k, ultra-detailed, professional, atmospheric, moody, dramatic, epic, breathtaking, ethereal, magical ) and standalone "cinematic" . Video-specific slop to strip from positive prompts: "smooth motion," "buttery 60fps," "hyperreal," "ultra-cinematic," "epic," "breathtaking." For Wan, the negative prompt is supported — these slop terms can go in the negative if useful, but mostly just delete them. Replace each removed slop term with a concrete detail or remove. Audit. Read the prompt back. Could this describe a clip a real cinematographer might have captured, or does it still sound like a prompt? Verify: one named anchor, one off-center detail (one — not zero, not two), camera explicitly named, environmental motion explicitly named, light is interacting not just labeled. For 2.6+: continuity anchors present in shots 2+. Output format User pasted an existing prompt: 1–2 sentence diagnosis → fixed prompt in the appropriate output structure (single-block for 2.2, shot-block for 2.6+) → bullet list of what changed and why, calling out the named anchor and the off-center detail explicitly. User gave a thin seed: skip the diagnosis. Enhanced prompt in the appropriate output structure → short bullet list of what you added and why, naming the anchor and the off-center detail. Escape hatch When the seed already carries an unusual or surreal ingredient — "a Lynchian dream of a phone booth at the edge of a wheat field" , "a fish swimming through a kitchen as if it's the ocean" , "a Magritte sky raining bowler hats over a city street" — the rubric overconstrains. Keep the existing concept; add ONLY camera movement, one environmental motion cue, one off-center detail, and one named anchor. Don't pile on subject specifics that fight the surreal/abstract register. See Examples 6 and 7 below for the canonical Enhance Mode patterns (one shot-block, one single-sentence). Output requirement For text-to-video on Wan 2.1 / 2.2 T2V: **Positive prompt:** \`\`\` ... prose prompt following the official formula ... \`\`\` **Negative prompt:** \`\`\` ... brief negatives, see references/negative-prompts.md ... \`\`\` For image-to-video on Wan 2.2 I2V / 2.5: **Motion prompt:** \`\`\` ... motion / camera / environment / pacing description ... \`\`\` **Negative prompt:** \`\`\` ... morphing/warping prevention set ... \`\`\` For multi-shot cinematic on Wan 2.6 / 2.7: **Global look:** [tone, lighting, palette, realism level, lens character] **Shot 1 [0–Xs]:** [camera movement and action over time] **Shot 2 [X–Ys]:** [continuation with restated continuity anchors] **Negative prompt:** \`\`\` ... ... \`\`\` For Wan 2.7 with multi-reference inputs (up to 9 references) or branded work with brand colors , prepend a Reference assignment block before the Global look: **Reference assignment:** - Reference 1 — [role: main character / environment / product / brand mark / etc.] - Reference 2 — [role] - Reference 3 — [role] **Brand palette (if applicable):** primary `#HEX`, accent `#HEX` **Global look:** ... After that, follow the same shot-block structure. The reference roles and brand palette establish bindings before any shot block uses them — restate the palette as a continuity anchor in later shots. After the code blocks, add a one-line note on recommended parameters if relevant (resolution, fps, seed advice). Length sweet spots Wan 2.1 / 2.2 T2V: 80–120 words. Enough to specify subject + scene + motion + camera + style without overloading. Wan 2.2 I2V / 2.5 I2V: 40–80 words. The image already provides the "what" — your job is to describe motion. Wan 2.6 / 2.7 multi-shot: 30–60 words per shot block. Total across blocks can reach 150–250 words for a 3-shot sequence. Wan 2.6 / 2.7 single-shot: 60–100 words structured as global look + shot block. Too short → the model fills gaps with generic motion. Too long → it averages or drops detail. The official Wan 2.2 prompt formula This comes from Alibaba's own Tongyi Wanxiang documentation. It's the canonical structure. Basic formula (for new users, simple scenes) prompt = subject + scene + motion (主体 + 场景 + 运动) Example: A young Chinese woman in mecha-styled hanfu, raven hair pulled into a bun, turning to look at the camera, her soft and lustrous hair drifting lightly in the air. Advanced formula (for richer scenes, better cinematic results) prompt = subject(description) + scene(description) + motion(description) + camera language + atmosphere + stylization (主体 + 场景 + 运动 + 镜头语言 + 氛围词 + 风格化) Subject description — appearance, traits, clothing. E.g., "a black-haired Miao ethnic-minority girl in traditional dress" Scene description — environment details. Foreground/midground/background, time of day, weather. Motion description — amplitude (small/large), speed (slow/fast), effect (e.g., "shattering the glass", "swaying violently", "moving slowly") Camera language — shot type, angle, lens, camera movement. See references/cinematography-vocabulary.md for the canonical vocabulary. Atmosphere — single mood word: "dreamy", "lonely", "majestic", "tense", "tranquil" Stylization — visual style anchor: "cyberpunk", "line-illustration", "wasteland aesthetic", "cinematic film still" Example: A black-haired Miao ethnic-minority girl in traditional embroidered indigo robes walks slowly through a misty rice terrace at dawn, her silver headpiece catching the first cool light. The camera tracks alongside her at waist level with a smooth dolly move, eventually pushing in to a medium close-up as she pauses to look out across the valley. Soft volumetric mist drifts between layered fields, low golden-blue ambient light. Tranquil, contemplative atmosphere, cinematic film still in the documentary tradition. See references/wan-22-prompting.md for the full formula breakdown with more examples. I2V (image-to-video) specifics Wan 2.2 I2V and Wan 2.5 I2V are the most-used Wan variants. The image you upload already defines the "what" — your prompt's job is to describe how things move . The 4-part I2V framework: Primary motion — what the main subject does. Concrete verbs: "spins", "walks forward", "raises her hand". Avoid vague descriptors. Camera behavior — pan, tilt, dolly, orbit, static, push-in, pull-back. Pick ONE main camera movement per generation. Environmental effects — secondary motion: wind in hair, falling leaves, swaying trees, rippling water, drifting smoke. Speed / intensity modifiers — "slow", "smooth", "rapid", "subtle", "intense". Tunes the pace. Template: [Primary motion], [camera movement], [environmental effects], [speed modifiers] Example for a portrait image: The woman turns her head slowly to look directly at the camera, slight gentle smile forming. The camera holds static. Soft wind moves her hair and the leaves visible in the background. Slow, controlled pace throughout. I2V do's and don'ts: ✅ Describe motion of elements already in the image ✅ Specify what stays still ("subject's face remains expressionless, only hair moves") ✅ Layer motion for depth: foreground / midground / background ❌ Don't try to add new elements not in the image — the model animates what exists ❌ Don't combine simultaneous complex camera moves — chain them sequentially or pick one ❌ Don't describe new scenery — that's what T2V is for See references/wan-22-prompting.md for more I2V patterns. Camera language — Wan's biggest control surface Wan was specifically trained on cinematography vocabulary. Real camera terms work better than generic descriptions. Common camera movements Wan recognizes: Movement What it does Static / locked-off Camera doesn't move Pan (left/right) Camera pivots horizontally on a fixed point Tilt (up/down) Camera pivots vertically Dolly (in/out) Camera physically moves forward or back Truck (left/right) Camera physically moves sideways Pedestal (up/down) Camera physically moves vertically Orbit Camera circles around the subject Push-in Slow dolly forward, building intensity Pull-back Slow dolly out, often as a reveal Tracking shot Camera follows the subject's motion Handheld Subtle vibration, documentary feel Whip pan Very fast pan, often for transitions Crane / boom Camera rises or descends through space POV First-person perspective Dutch angle Tilted frame, tension