wan-2-6-prompting

Wan 2.6 → two modes, two prompt shapes

Wan 2.6 generates from text or animates a still. Each mode wants a differently shaped prompt. Getting the shape wrong is the usual reason a clip comes back generic.

Reach it with find_model for text_to_video or image_to_video, then generate_video / animate_image. Those forward the prompt (and the image), a negative prompt (not animate_image), the duration, the aspect ratio and the resolution. The shot-mode switch, prompt expansion and an audio file exist only on the route's own node: find it with search_nodes, read its fields with get_node_info, and run it with run_node. Each run is a paid generation, so confirm the scope with the user before rendering several.

fal's guide caps main prompts at 800 characters. No shipped route states the limit, but write as if it holds.

Text to video

Two components: a global style line, then shots with timing brackets.

A cinematic journey through ancient ruins at sunset. Photoreal, 4K, film grain.

Shot 1 [0-3s] Wide establishing shot of stone pillars with sunlight streaming through.
Shot 2 [3-7s] Camera tracks forward through an archway revealing a hidden chamber.
Shot 3 [7-10s] Close-up of ancient inscriptions as dust particles float in light beams.

The shot mode is a switch, and its name and default differ by provider: AtlasCloud's shot_type defaults to multi, and Replicate's multi_shots defaults to false, so set it true there before sending bracketed shots. Give each shot a timing indicator, a camera action (push, pull, pan, orbit, track) and its scene elements — subject position, lighting change, environmental detail. Keep continuity by referencing the same characters, locations and objects across shots; unrelated shots produce disjointed results.

AtlasCloud's durations are 5, 10 or 15 seconds, at 720p or 1080p, in 16:9, 9:16, 1:1, 4:3 or 3:4. Write to the ratio you picked: wide establishing shots and horizontal movement for 16:9, tighter framing and vertical composition for 9:16, centred subjects for 1:1.

Image to video

Describe the temporal change, not the image. The model already has the frame; words spent re-describing it are words not spent on motion.

Continue from first frame. Gentle camera push toward the mountain peak as
clouds drift overhead. Light changes from morning to golden hour. Cinematic and
serene movement.

Spend the prompt on camera motion, lighting shifts, and environmental animation — water, clouds, foliage, smoke. The same shot switch applies: AtlasCloud defaults to multi, Replicate to multi_shots false. Set single shot for a plain animated still.

Images animate better when they are high resolution, have clear depth of field, carry atmospheric elements, and are uncluttered with a well-defined subject. Busy compositions with competing elements produce inconsistent motion. fal's guide gives input images 360 to 2000 px per side, up to 100 MB.

Controls

Negative prompt — fal's guide caps it at 500 characters. Prioritise the artifacts you actually see: "low quality, blurry, distorted faces, unnatural movement, text, watermarks, shaky camera".

Motion intensity is a wording choice, not a parameter. Minimal: "subtle camera drift", "gentle movement". Moderate: "smooth camera track", "flowing motion". Dramatic: "dynamic sweeping motion", "rapid camera movement".

Audio input (audio on the route's node) takes WAV or MP3, 3 to 30 seconds, up to 15 MB, per fal's guide. Audio longer than the video is truncated; audio shorter leaves the rest silent. For dialogue, put speaker cues in the prompt.

Prompt expansion (enable_prompt_expansion) is on by default on every Wan 2.6 video route and runs an LLM over the prompt before generation. It adds detail without eating the 800-character budget, and it works best when you have given it style references, visual descriptors and precise terminology to expand from.

Symptoms

What went wrong	What to change
Incoherent narrative	Break the scene into specific shots with timing brackets
Inconsistent subjects	Repeat the character description across every shot
Unnatural motion	Name the camera movement explicitly: push, pan, orbit
Flat visual quality	Add quality descriptors to the global style line
An element ignored	Move it earlier in the shot description

Iterate one variable at a time, and start in text-to-video — the shot and timing format transfers to image-to-video without any asset dependency. Check the render with analyze_video, detect_video_scenes and understand_video.

Adapted from fal's Wan 2.6 prompt guide: https://fal.ai/learn/devs/wan-2-6-prompt-guide-mastering-all-three-generation-modes