News
Twenty References, Thirty Seconds: What Wan 3.0 Changes for Character Control
ยท RenderBob team
Wan 3.0 accepts up to 20 reference assets and generates up to 30 seconds in one pass, including native audio. A character sheet can travel with the request.
Alibaba's Tongyi Lab launched Wan 3.0 in public beta in early August 2026. Listings give dates between 31 July and 6 August. The headline numbers matter for anyone trying to keep a character recognisable: up to 20 reference assets in one request, and clips of 2 to 30 seconds generated in a single pass at up to 1080p.
What the references can be
Per documentation from model hosts, a task can include up to 10 images, 5 reference videos, and 5 audio clips, with video and audio references each totalling about 15 seconds, plus a document. Alibaba's launch material also describes accepting slides, spreadsheets, and public web pages as inputs. The division of labour is straightforward. Images set identity: a character card, a product shot, a location plate. Video references carry gesture and movement. Audio references carry voice. The model covers text-to-video, first-frame and first-and-last-frame control, reference-driven generation, localised editing, and extension.
Why 30 seconds matters
One analysis points out that a native 30-second take leaves no cut for a character to drift across. Previous Wan reference-to-video releases were limited to around 10 seconds per pass, so a scene needed several generations stitched together, each a chance for a face to change. Wan 3.0 also offers a six-shot director mode that plans several shots from one prompt.
What it means for character creation
A character sheet is now an input, not only a design document. A sensible approach is to build a neutral-light reference set, several angles of the face and costume, add a short clip showing how the character moves, and a clean voice sample, then hold those fixed across every request. Multi-character scenes remain the hard case. Make characters visually distinct, and test two-shots early.
Open questions
Distribution is unclear. Some distributors describe Wan 3.0 as open-weight under Apache 2.0, while Alibaba's own launch describes a public beta on its cloud services. If you plan to run locally, confirm licensing and weight availability from the official source. Quality claims come from the vendor and from aggregators. The Sandbase benchmark of 4 October scored it highly, on a small sample. Billing, reported at roughly $0.05 to $0.20 per second depending on resolution, also changes quickly.
What to watch
Whether references hold up under real production conditions: unusual lighting, profile angles, and scenes with several characters. A system that takes twenty assets is only as good as its worst-chosen reference, so curation becomes the skill.
More from the blog
- Higgsfield Says AI Video Now Runs Past 10 Minutes: What an OpenAI Customer Story Does and Doesn't Show
An OpenAI customer story from 13 October says Higgsfield can now produce AI video longer than 10 minutes using GPT-6 Astra. The claims are the company's, and no independent test has been published.
- Disney Bids $20 Million for a License to Character.AI's IP: What the Stalking-Horse Deal Does and Doesn't Say
Character.AI filed for Chapter 11 on 12 October, and Disney Worldwide Services is the opening bidder at $20 million for a non-exclusive, perpetual licence. Higher bids can still come in.