Updated 21 September 2026 · wan3.0-video · vs Seedance notes
Wan 3.0 prompt guide: Alibaba shot language vs Seedance after Sora
Sora liked one physics paragraph. Alibaba’s published Wan 3.0 examples use Shot N [start–end] plus native audio. ByteDance Seedance on this site wants FORMAT / TIMELINE / CAMERA. Same story, two grammars. This page is the Wan column and the vs-Seedance notes — not a ranking and not invented search volumes.
OpenAI listed no successor. If you already sit on Alibaba Cloud Model Studio, wan3.0-video (and wan3.0-video-prime) is an all-in-one name — the docs say it routes from input.media types plus prompt intent. If you already sit on Ark / Seedance, stay on that Seedance remap. Switching vendors just to “replace Sora” is optional; switching sentence shape is not.
Unofficial craft notes · official capabilities cited
Independent page. Not affiliated with Alibaba, Wan, Tongyi, DashScope, ByteDance, Seedance, or OpenAI. Capability bullets below follow Alibaba’s English Wan 3.0 guide (last updated on that page 26 Aug 2026; re-checked here 21 Sep 2026). Billing is regional — see sources, verify the console. Cartoon hero is original art, not official Wan branding.
All-in-one model IDs: wan3.0-video / wan3.0-video-prime (no separate t2v/i2v model name for this series).
Up to 30 seconds per generation; output MP4 at 30 fps.
Native dialogue, BGM, and sound effects in the same pass.
Up to 20 multimodal refs per request (images, videos, audio, documents, web pages — type caps below).
First-frame / first-last-frame control; video editing and extension via prompt intent + reference_video.
Parameters the samples use: resolution (480P / 720P / 1080P in examples), ratio (including adaptive), duration (or -1 smart duration), prompt_extend.
Vs wan2.7 (same FAQ): unified name; 30s vs wan2.7’s 10s cap; new reference mode; adaptive ratio; editing/extension.
Media types (cannot mix first/last frame with reference/file/link in one request):
Constraints from the official Wan 3.0 guide. File-size and duration caps can move — verify on the API reference.
type
Role
Published cap
first_frame / last_frame
Exact opening / ending still
Max 1 each
reference_image
Identity / product / scene
Max 10; prompt says “Image 1”…
reference_video
Motion / edit / extend source
Max 5; total duration ≤ 15s
reference_audio
Voice / bed
Max 5; total duration ≤ 15s
file or link
Doc or public page
Max 1; mutually exclusive
Wan 3.0 vs Seedance (prompt grammar)
Both can do timestamped multi-beat clips. They do not share labels. Seedance copy-paste on this site: remap pack. Picker: Seedance vs Kling vs Veo.
Craft comparison. Not a quality score. Seedance billing is tokens (Volcengine); Wan 3.0 is per generated second on Model Studio — different meters, see sources.
Wan 3.0 (Alibaba docs)
Seedance 2.5 (this lab’s remap)
Shot list
Shot 1 [0-2s] … in one prompt (official sample uses four shots to 10s, duration 20)
TIMELINE: 0–5s … 5–12s … plus FORMAT / CAMERA / CONSTRAINTS
Identity
“Image 1”, “Video 1”, “Audio 1” matching input.media order by type
@Image1 face/outfit roles; cameo IDs still do not port
Length habit
Docs: up to 30s native multi-shot with audio
Directed beats; long Sora films still split better than one 30s dump
Audio
Native dialogue / BGM / SFX; still say “no licensed song” if you do not want one
Singapore (international deployment row on that page) — 480P 0.37471 CNY/s; 720P 0.74942 CNY/s; 1080P 1.49884 CNY/s.
USD rows, if your console shows them, belong to that console — this page will not invent a CNY→USD rate or a “% cheaper than Sora” figure. Seedance token rates stay on the leaving-Sora cost page.
Copy-ready Wan 3.0 prompts (unofficial, follow official shape)
Text-to-video · official shot pattern
Sora mini-film → Shot timestamps, not one physics blob
Shape matches Alibaba’s garden sample: Shot N [0-2s]. Swap the story; keep the timestamps honest vs duration.
wan3.0-video prompt
Copied
Photoreal dusk in a narrow Tokyo alley, practical neon, wet cobbles, no on-screen text. Shot 1 [0-3s] Medium-wide: a woman in a wool coat reads a paper ticket beside a ramen cart; steam drifts. Shot 2 [3-7s] She looks up as a train passes overhead; camera begins a slow dolly-in on the same axis, no orbit. Shot 3 [7-10s] Medium close-up of her face, neon in puddles; hold. Diegetic audio only: steam hiss, train rumble, light rain. No music, no voiceover.
API sample parameters: resolution 480P or 720P while testing, ratio 16:9 or adaptive, duration matching the last timestamp, prompt_extend true if you want the official expander. Seedance version of this beat: dolly remap.
Reference · Image 1 / Video 1
Sora cameo → numbered media, counted by type
Official FAQ: first reference_image is “Image 1”; first reference_video is “Video 1”, even if images sit first in the array.
Reference prompt
Copied
The person in Image 1 (face, hair, jacket only) walks a sunflower path. Motion follows Video 1’s walk cycle; do not copy Video 1’s background. Camera trucks sideways at waist height. Wind in the stalks, footsteps, no music. Keep identity locked; do not invent a different person. 8 seconds, photoreal, 16:9.
Vs Seedance · same rainy-exit beat
Write Wan shots; do not paste FORMAT blocks into DashScope
Wan
Copied
Shot 1 [0-4s] Doorway of [VENUE] at night, rain, two partners under a clear umbrella, guests as silhouettes. Shot 2 [4-8s] Dolly with them toward a waiting car; petals stick to the umbrella; headlights bloom. Camera: one forward move, no orbit. Audio: rain on umbrella, wet footsteps, a car door. No commercial song.
Seedance (do not send this to Wan)
Copied
FORMAT: 16:9, 8 seconds, rainy night.
SUBJECT + ACTION: Two partners leave [VENUE] under a clear umbrella; petals; they reach a car.
CAMERA: Slow dolly forward with them. No orbit.
AUDIO: Rain on umbrella, footsteps, car door. No music.
CONSTRAINTS: Guests unreadable. Same wardrobe throughout.
Ready-made ceremony beats: /ai-wedding-video-prompts/
