WAN
The prompt-understanding line — strongest at turning a still into motion
WAN is the family to reach for when you have a first frame and a description of what should happen next. The 2.7 line reads long prompts more literally than anything else in the catalogue, holds physics together across a 15-second clip, and ships an image model that renders bodies and composition with noticeably fewer limb artifacts than the older general workhorse.
Alibaba · Tongyi Lab
Strengths and limits
Key Features
- Best-in-catalogue prompt adherence: a long, specific instruction comes back as the shot you described rather than an average of it.
- Animates a supplied first frame faithfully, so a character you already generated stays the same character.
- Uncensored across the line, with a LoRA track for action presets you drive from a single reference photo.
- Both halves of the family are available — WAN image models for stills, WAN video models for motion — so a look carries from one to the other.
Where it falls down
- Weak at text inside the image. For a poster, a sign or a caption, use Nano Banana instead.
- Uncensored is not the same as explicit-from-scratch: the line transforms references accurately but will not invent full NSFW anatomy from a text prompt alone.
- The 2.2 and 2.5 generations are kept for LoRA compatibility and price, not because they match 2.7 on quality.
Model
Capability tables are read from the live model catalogue. Where a table and a paragraph disagree, the table is right.
| Model | Tools | Quality | Duration | Resolution | Audio | Filters | Credits |
|---|---|---|---|---|---|---|---|
| WAN 2.7 Image ProWAN_2_7_IMAGE_PRO | Reference Generator | high-quality | — | 2K | Filters: Minimal | 2 credits | |
| WAN 2.5wan@2.5 | Image-to-Video | WITH AUDIO & UNCENSORED | 5–10s | 480p · 720p · 1080p | Filters: Minimal | 5 credits |
Tools
Start from an image
Frequently Asked Questions
Start at 2.7 — it is the current line and reads prompts best. Drop to 2.6 Flash when you are iterating and want the cheapest usable draft, and to 2.2 + LoRA when you want a trained action preset rather than a written description.
The 2.7 line goes to 15 seconds. The exact set of durations and resolutions each version supports is shown in the capability table above and enforced by the tool — a duration a model cannot honour is struck through rather than hidden.