AI Video Β· Alibaba Tongyi Wanxiang
Wan 2.6: Star in Your Own Multi-Shot Short Drama, Voice Included
This is the video model Alibaba's Tongyi Wanxiang team released on December 16, 2025, and it was pitched with a simple promise: you can be in the video. Upload a short clip of yourself, a friend or an original character, and the model borrows that look and voice to act out a brand-new scene. Add a one-line idea and it plans several shots, keeps the same faces from cut to cut, and delivers dialogue, sound effects and lip sync in one pass. Below you will find what changed since Wan 2.5, where short-drama makers and online sellers are using it, five fun ideas, step-by-step instructions and prompt templates you can paste right away.
- Reference-video role-play
- Multi-shot from one prompt
- Dialogue, sound and lip sync
From Wan 2.5 to Wan 2.6: What You Can Do Now
Wan 2.5 already made sound part of the package, but each clip was one shot, ten seconds at most, and the character only existed inside that single generation. The new release turns those limits into features. Clips now run to 15 seconds, one prompt can be split into several camera setups, and a reference video lets the same person appear again and again across different scenes. For anyone telling a story in installments, that consistency matters more than any jump in sharpness.
Alibaba ships the family as three modes in its Model Studio: text-to-video, image-to-video and reference-to-video (often shortened to R2V). Output comes in 720P or 1080P at 24 frames per second, in landscape 16:9 or vertical 9:16, with 5, 10 or 15 second lengths.
| Wan 2.5 | New version | |
|---|---|---|
| Longest clip | 10 seconds | 15 seconds |
| Shots per clip | One continuous shot | Several shots planned from one prompt |
| Starring a real or invented character | Only via a start image | Reference video copies look and voice |
| Two characters in one scene | Hard to control | Two references, each with its own voice |
| Sound | Audio supported | Dialogue, effects and lip sync in one pass |
Role-Play With a Reference Video: Put Anyone in the Scene
Role-play is the headline feature, and Alibaba called it a first for a video model in China. You record or pick a short clip, a few seconds of someone talking and moving is enough, and the model picks up their appearance and the tone of their voice. Then you write the scene: walking into a cafΓ©, arguing on a rooftop, presenting a product. The character performs it, and the voice you hear matches the reference instead of a stock narrator.
It also works with more than one reference. Two clips give you a two-person dialogue where each speaker keeps their own voice, and a person plus an object or pet gives you those 'me and my dog on the beach' moments. In prompts the references are usually called character1 and character2, which makes it easy to say who does what.
- Solo: a creator appears as a detective, a chef or a news anchor without filming the scene.
- Duo: two friends or two original characters trade lines, each with their own voice.
- Person and object: a host and a product, or an owner and a pet, share the frame naturally.
One Idea, Several Shots: How Multi-Shot Storytelling Works
Ask for 'a reunion at a train station' and the model does not just render one wide view. It expands the idea into a small script, for example a wide establishing shot, a close-up on a face and an over-the-shoulder reveal, then cuts between them while keeping the same people, outfits and setting. If you want control, write the shots yourself: one line per shot, with the framing and camera move named.
That is why the release caught on with vertical short-drama studios. A 15-second clip with three shots is exactly the size of a conflict, a twist and a cliffhanger, and the faces do not drift between cuts the way stitched clips often do.

Where It Is Taking Off: Short Dramas, Shop Videos and Ads
The model was positioned for professional film and advertising work, and the clearest early uses come from teams that need many short, consistent videos every week.
- Vertical short dramas and comic adaptations: one character reference plus a three-shot prompt delivers a conflict-and-twist beat in 9:16, ready for episode after episode.
- E-commerce and live-selling clips: a host holds the product, names a benefit and smiles at camera, with the product shown from several angles in the same clip.
- Brand campaigns: the same spokesperson or mascot stays recognizable across dozens of ads and languages.
- Vlogs and social skits: creators feed in their own clip and star in sketches they never had to shoot.
- Training and pre-visualization: teachers turn a script into a talking explainer, and directors preview a scene before a real shoot.
Five Fun Ideas People Keep Making
Once the basics click, a few formats keep coming back on social feeds because they are quick to make and easy to share.
- Me and my pet: you and a golden retriever play fetch at sunset, in slow motion with barking and waves.
- Two-voice dialogue: two references argue, flirt or make up, and each line comes in the right voice.
- One-line micro movie: a single sentence becomes a three-shot mini film with music and ambience.
- Talking photo: an old portrait or a selfie turns to camera and says a line you wrote.
- Series with one lead: the same character returns every day in a new location, building a running story.
How to Make a Wan 2.6 Video in Five Steps
The flow is the same whether you use Alibaba's Wanxiang website or the Model Studio API; only the buttons look different.
- Pick a mode: text-to-video for a fresh idea, image-to-video to animate a photo, or reference-to-video to cast a character.
- Add your material: a start image, or a reference clip of about 2 to 10 seconds per character.
- Write the shots: one line per shot, name the framing (wide, medium, close-up) and the camera move, and put dialogue in quotes with the speaker named.
- Choose the format: 720P or 1080P, 5, 10 or 15 seconds, and 16:9 or 9:16.
- Generate, review and download; if a hand or face looks off, shorten the action or split it into another shot.
Wan 2.6 Prompt Templates You Can Copy
These templates follow the shot-by-shot style the model reads best. Swap the people, places and lines, and keep the structure.
Three-shot short drama
Shot 1: wide shot, rainy night street, a young woman in a red coat stops under a neon sign. Shot 2: close-up, she turns and says "You finally came." Shot 3: over-the-shoulder, a man in a black jacket steps out of the shadows. Cinematic lighting, 9:16, consistent characters.
Product pitch
Medium shot, a female host holds a glass serum bottle toward the camera, smiles and says "Three drops, every night." Soft studio light, product label readable, clean beige background.
Reference role-play
character1 walks into a cozy cafΓ©, waves at character2, they sit down and laugh; warm afternoon light, handheld camera, natural ambient sound.
Person and pet
character1 plays fetch with a golden retriever on a beach at sunset, slow-motion jump, waves and barking sounds.
Summer lookbook (18+)
Shot 1: wide shot, a confident adult woman (25+) in a white bikini walks along a tropical pool deck. Shot 2: medium shot, she adjusts her sunglasses and smiles at the camera. Shot 3: close-up, water ripples, summer music. Fashion-editorial style, 9:16.
Swimwear runway (18+)
Wide shot, adult models in colorful swimsuits walk a seaside runway at golden hour, camera tracks alongside, crowd cheering and upbeat music, then a close-up of the lead model turning to camera.
Tip: name the speaker before each quoted line, and keep complex body movement to one shot.
Wan 3.0, Kling 3.0 or This Release: Picking the Right Tool
Alibaba has since shipped Wan 2.7 and Wan 3.0, yet this version is still on sale in Model Studio, and for many jobs it remains the sensible choice. If your clips are 15 seconds or shorter and you care most about a recurring character with a recognizable voice, it covers the job. Reach for Wan 3.0 when you need a single take of up to 30 seconds or want to feed in documents and web pages. Kling 3.0 from Kuaishou is the closest rival for multi-shot clips with native sound.
| This release | Wan 3.0 | Kling 3.0 | |
|---|---|---|---|
| Best for | Role-play series, shop clips, short drama beats | Longer single takes, document or web input | Multi-shot scenes with sound |
| Longest clip | 15 seconds | Up to 30 seconds | Up to 15 seconds |
| Open weights | No | No | No |
What Creators Are Saying
Early feedback from Chinese short-video makers and international testers follows a clear pattern. The praise centers on how well characters hold up across cuts and on lip sync for Mandarin dialogue, which used to be a weak point for imported tools. Product sellers like that a host can actually speak the line instead of being dubbed afterwards.
The complaints are just as consistent: complex body movement such as dancing or fighting can still bend limbs, the last few seconds of a 15-second clip sometimes drift away from the reference, and the official moderation is strict enough to reject swimwear or romance prompts that feel harmless. Most users work around the first two by keeping each action inside one shot.
Open Source, Download and ComfyUI: The Straight Answer
Alibaba has not released the weights. You use it through the Wanxiang website or the Model Studio API, and any ComfyUI node labelled with this version simply calls that hosted API rather than running the model on your own GPU. Files sold as a download or an 'uncensored' build of this version are usually renamed copies of older open models.
The newest Wan video model you can actually download is Wan 2.2, published under Apache-2.0 on GitHub and Hugging Face with text-to-video, image-to-video and a smaller 5B model, and it has a large community of fine-tunes and LoRAs.
Wan 2.6 Uncensored Video Generator
On Alibaba's own platforms every prompt and upload goes through content moderation, so nudity and explicit scenes are blocked. Spicy AI offers the same image-to-video model with fewer limits for adult (18+) fiction: upload a start photo, describe the motion and the lines, and pick 5, 10 or 15 seconds. Popular ideas include a bikini pool-deck lookbook, a lingerie-style studio shoot and a beach holiday vlog with a fictional adult character.

All characters must be adults. Do not upload real people without their permission.
Open image to videoUsing the Official Release vs Generating on Spicy AI
| Official service | Spicy AI online | |
|---|---|---|
| Content filtering | Moderated on Alibaba Cloud and Wanxiang | Adult (18+) content allowed |
| Open weights | No, closed service | Not applicable, runs online |
| Model actually used | Text, image and reference modes | Image-to-video from a start photo |
| Price | Pay per second of video on Model Studio | Free to start, credits for more |
- Official: reference-video role-play and text-to-video on the Wanxiang website and API.
- Spicy AI: upload your own photo and animate it in the browser.
Frequently Asked Questions
Who made Wan 2.6 and when was it released?
Alibaba's Tongyi Wanxiang team released it on December 16, 2025, on Alibaba Cloud Model Studio and the Wanxiang website.
How long can a video be?
Each clip runs 5, 10 or 15 seconds in 720P or 1080P, at 24 frames per second.
What is reference-to-video role-play?
You upload a short reference clip and the model copies that person's look and voice to act out a new scene from your prompt. It supports solo scenes, two-person dialogue and a person together with an object or pet.
Can it make videos with dialogue?
Yes. Audio is generated together with the picture, including speech with lip sync, sound effects and ambience.
Is it open source, and can I download it?
No official weights have been published, so it only runs on Alibaba's services. Wan 2.2 is the latest open-weight Wan video model you can run locally.
Should I pick this version or Wan 3.0?
Choose Wan 3.0 for single takes up to 30 seconds or document and web input. For clips of 15 seconds or less, especially role-play series and product videos, the older release is still a good fit.
Is there an uncensored version?
There is no official or trustworthy community uncensored build, because the weights were never released. Downloads with that name are usually older open models.
Sources
More AI Video Models
- Uncensored Wan
- Uncensored Wan 3.0
- Uncensored Kling
- Uncensored Kling 2.6
- Uncensored Kling O1
- Uncensored Sora
- Uncensored Veo
- Uncensored Hailuo AI
- Uncensored MiniMax H3 Max
- Uncensored Seedance
- Uncensored Seedance 2.5
- Uncensored Seedance 2.0
- Uncensored Seedance 2.0 Mini
- Uncensored HunyuanVideo
- Uncensored LTX Video
- Uncensored Runway
- Uncensored Pika
- Uncensored Luma Dream Machine
- Uncensored Vidu
- Uncensored PixVerse
- Uncensored EdenLive
- Uncensored Grok Imagine
- Uncensored AI Avatars
- Uncensored HappyHorse 1.0
- Uncensored MiniMax H3 FL2VA Ref2VA
- Uncensored Kling 4.0
- Uncensored Kling 3.0
- Uncensored MiniMax H3 RefMods
- Uncensored Kandinsky 6.0
- Uncensored Looking Ahead
- Uncensored Kandinsky S2V 1.0 Pro
- Uncensored TaoMate-H3
- Uncensored SANA-Video 2.0
- Uncensored MiniMax H3 Remix
- Uncensored Sora 2
- Uncensored LTX-2.5 Slow Motion
- Uncensored VDN MiniMax H3
Wan and Tongyi Wanxiang are trademarks of Alibaba Group. All content must depict adults (18+).
