Wan 3.0 makes recurring AI characters real
Wan 3.0 just shipped cross-session Identity Lock — save your AI character once, and the same face returns next week. Recurring characters are finally real.
Every AI character you built last month is dead. Not because the model deleted them — because the next generation session gave you a slightly different face, and the recurring character you were building isn't quite the same person anymore.
The Wan 3.0 highlights that matter for AI creators
Cross-session Identity Lock — save a character once in Wan 3.0, get the exact same face back next week, next month, next campaign.
Native 4K, 30-second single-pass clips with synchronized dialogue, SFX, ambient, and music — all generated in the same run, no post-sync.
AI Director stitches up to 6 shots per pass with per-shot camera control, and you can chain generations for anything longer.
Why Wan 3.0 finally solves the recurring-character problem
The whole reason AI creators can't grow an audience around a digital persona is that the face drifts. You nail a look on Monday, and Wednesday's generation quietly widens the jaw or shifts the eye color. Your "same character" is a rotating cast.
Alibaba's Tongyi Lab shipped Wan 3.0 — its latest video model — with a feature they call Identity Lock. Save the character's visual profile after the first generation, and calling that profile in a later session produces the same character in a new scene, without re-describing them. That's the cross-session part, and it's the one that changes everything about serialised AI content.
This week YouArt brought Wan 3.0 into its browser platform, joining the Alibaba Cloud Model Studio API that shipped a few weeks earlier. In short: you don't need to code against DashScope to try it.
Receipts: Alibaba Cloud Wan 3.0 API reference · Alibaba Cloud technical writeup · Atlas Cloud 49-run bench vs. Seedance 2.0 · YouArt launch note (Aug 20).
How to make a recurring AI character in Wan 3.0
The following images were generated using Nano Banana 2:
Lock the face on the first run. Start with a text-to-video or image-to-video prompt that over-specifies the character — age, ethnicity, hair, eye color, wardrobe, distinguishing marks. When the first clip comes back with the face you want, save that generation as a named Identity Lock profile inside Wan 3.0. This profile — not your prompt — is what future sessions will call.
Storyboard six shots at a time, not one. AI Director runs up to six shots per generation with independent camera moves. Instead of prompting shot-by-shot, write a mini-scene — close-up smile → over-the-shoulder in the café → wide walking shot at night — and let the model hold your Identity Locked character across all six.
Generate audio in the same pass. Wan 3.0 renders dialogue, SFX, ambient, and music as native tracks — phoneme-level lip sync across 12 languages. Write the line into the prompt ("she says: '…'") instead of generating a silent clip and dubbing later. That single change kills the round-trip that eats most creator hours.
Chain runs for anything longer than 30 seconds. A single Wan 3.0 generation caps at 30 seconds. Use Video Continuation — a follow-on prompt that carries your character, environment, and lighting forward — to build a 90-second piece out of three chained runs. Trim generously: Atlas Cloud's 49-run bench flagged uncommanded camera cuts as the model's most reproducible defect, so plan for a scissors pass.
Example A: Mia on a rainy street
Prompt
Identity Lock: Mia. Medium shot: Mia walks slowly down a wet neon-lit city street at blue hour, holding a small paper coffee cup. Coral cropped jacket, white tee. Bright split lighting — cool blue key from the left, warm coral rim from the neon signage on the right. She glances slightly off-camera and smiles. Ambient: light rain, distant traffic. No dialogue. 6 seconds, 1080P.
Example B: Same Mia, next morning
Prompt
Identity Lock: Mia. Kitchen close-up: Mia at a marble island, coral cropped jacket with sleeves rolled up, chopping fresh herbs on a wooden board. Warm morning sunlight through a large window on the right. She hums quietly. Ambient: knife on board, distant birds. She says, in Japanese: 'ちょっと待ってて.' 8 seconds, 1080P.
Same face. Same eyes. Different scene, different wardrobe context, different language, generated on a different day. That's the whole trick — and until Wan 3.0, it wasn't reliably possible in a single-model workflow.
Why Identity Lock beats the 4K headline
Almost every write-up on Wan 3.0 leads with 4K or the 30-second cap. Both are useful — the resolution kills the upscaler softness, the length fits a full short-form beat — but they're incremental. Every model is going to hit those numbers eventually. Cross-session Identity Lock is the one that reshapes what a solo AI creator can build: a recurring character with a real audience, showing up in a new scene every week. That's the format Instagram, TikTok, and YouTube reward. Resolution is table stakes; a consistent face is a channel.
The bottom line
The recurring AI character is the next audience-building unit for creators, and Wan 3.0 is the first model that treats identity as a saved asset, not a per-session negotiation. If you're serious about building a digital persona your audience actually recognises next week, the workflow starts the same place it always does: with a face you actually own. Lock it in — start your AI avatar here — and everything downstream, from Wan 3.0's Identity Lock profile to the next model that ships, has something real to hold onto.
FAQs
What is Wan 3.0?
Alibaba Tongyi Lab's latest AI video model. It generates native 4K clips up to 30 seconds long in a single pass, with synchronised dialogue, SFX, ambient, and music, and it holds character identity across separate generation sessions.
How is Wan 3.0's Identity Lock different from Wan 2.7's character consistency?
Wan 2.7 held a character within a single clip. Wan 3.0 holds it across sessions — you save a profile once, and future generations on different days call that same profile. Same face, new scene.
How much does Wan 3.0 cost to run?
On the Alibaba Cloud Model Studio API, output is billed per second: $0.05/s at 480P, $0.10/s at 720P, $0.20/s at 1080P. A 30-second 1080P clip lands at $6.00. Browser platforms like YouArt bundle it into their credit plans instead.
What's the catch?
Atlas Cloud's same-seed bench found Wan 3.0 sometimes inserts uncommanded camera cuts inside a requested continuous move, especially on 30-second runs. Plan a trim pass. Audio texture and on-screen text rendering are also still improving.
Can I use Wan 3.0 for my avatar without touching an API?
Yes — YouArt shipped browser access on August 20, and Wan 3.0 is available through several other hosted platforms alongside the direct Alibaba Cloud API. Pick whichever pricing model fits your credit budget.