Wan 3.0 character consistency with a character sheet
Wan 3.0 is Alibaba's newest Wan video model. In reference mode it keeps a character's face, hair, body shape, clothes and accessories close to the images you upload, and it can make up to 30 seconds in one pass. This page shows which turnaround-pack images to upload and in what order, why a first frame and character references can't go in the same request, prompt templates for a 30-second one-take and for a storyboard, how to go past 30 seconds, and what a clip costs.
We have not made an example with Wan 3.0 yet. This page maps the pack to Alibaba Cloud's published API reference, prompt guide and prices; the examples on this site were made with Kling 3.0.
Wan 3.0 (official limits)
Checked 2026-11-22- Announced on the Alibaba Cloud blog on 13 August 2026 and available on Alibaba Cloud Model Studio as wan3.0-video (standard) and wan3.0-video-prime. The API reference still marks the model as a preview.
- 2 to 30 seconds per generation at 30 fps (default 5 s, or -1 to let the model choose). Output 480P, 720P or 1080P (default 1080P); aspect ratio adaptive (default), 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16. Audio is on by default and does not change the price.
- Reference mode: up to 10 reference images, 5 reference videos and 5 audio clips (videos and audio each 15 seconds or less in total), plus one document or web link. Images in JPEG, PNG without transparency, BMP or WEBP, 240 to 8000 px per side, no wider than 8:1, 20 MB or less.
- Images are named in the prompt as Image 1, Image 2… in upload order; videos and audio are numbered separately (Video 1, Audio 1). With two or more images you must give the number.
- First-frame (and first-and-last-frame) mode and reference mode are mutually exclusive. A request that sends first_frame together with reference images fails, and first-frame mode accepts no other inputs.
- Model Studio prices (Singapore, international): $0.05 per second at 480P, $0.10 at 720P, $0.20 at 1080P; the global regions list $0.041256, $0.082513 and $0.165025. Input video seconds are billed too, and input plus output may not exceed 30 seconds.
- wan3.0-video-prime has the same capabilities at higher prices: $0.068 per second at 480P, $0.14 at 720P and $0.28 at 1080P (Singapore). When resources are tight some tasks run at the standard rate and are billed at the standard price.
Sources: Alibaba Cloud: Wan3.0, 30-second AI video generation from any input · Alibaba Cloud Model Studio: Wan3.0 video generation API reference · Alibaba Cloud Model Studio: wan3.0-video model information and pricing · Alibaba Cloud Model Studio: wan3.0 video generation prompt guide · Alibaba Cloud Model Studio: wan3.0-video-prime model information and pricing
From turnaround sheet to Wan 3.0 in five steps
- Get a character sheet: download a pack from the library or a free character. Every pack has the front, side and back as separate files, plus full body, outfit, prop and scene images.
- Pick the mode for this shot. Reference mode keeps the character's identity from several views; first-frame mode locks the opening picture but takes no other images.
- In reference mode, upload the images below in this order and give each one a single job ("Image 2 is her profile"). Alibaba's guide says to separate the purpose of each material so features don't bleed between subjects.
- Write the action as numbered shots with timestamps, 2 to 5 seconds each, joined end to end, then list what must not change ("no costume changes, no scale drift").
- Draft at 480P (a 15-second test is about $0.75), keep the seed if a draft works, then render the same prompt at 720P or 1080P.
Which views to upload, in which order
Seven images cover one character, the outfit, the product and the room, leaving three of the 10 image slots for a second character or a detail crop. Keep the order the same for every shot so Image numbers never change.
- Image 1
02_turnaround_views/view_front.pngFace and front of the outfit - Image 2
02_turnaround_views/view_side.pngProfile and hair from the side - Image 3
02_turnaround_views/view_back.pngHair and outfit from behind - Image 4
01_character_reference/full_body.pngHeight and proportions - Image 5
03_outfits/outfit_work.pngThe job outfit - Image 6
04_props/prop_role.pngThe product or tool - Image 7
05_scenes/scene_workspace.pngThe location
Alibaba's docs do not say whether a whole multi-view sheet works better or worse than split views. Separate files let the prompt point at each view by number, which is what the guide's examples do, so the split files are our default.
Why three separate views beat one good angle
A single image, even a well-chosen three-quarter view, only shows part of the character. When the character turns, the model has to invent the profile and the back, and that is where the face, hair and outfit start to drift. Upload the front, side and back as separate references and give each one a single role in the prompt, as the official guide recommends. If you also have a three-quarter view, add it as an extra reference rather than using it on its own.
Keep one face per image: crop expression grids out of the sheet, and paste the same locked appearance description into every shot's prompt.
Character sheet plus a first frame: choose one mode
Seedance 2.5 lets you add a first frame as one more reference; Wan 3.0 does not. Sending first_frame with reference images returns an error. Two ways around it: use reference mode and describe the opening composition in Shot 1, adding the pack's starter image as one more reference ("Image 8 shows the opening composition") if you want it followed loosely. This is our suggestion, not an official mode. Or use first-frame mode with a starter frame that already shows the character clearly, and accept that the other views are not available, so a turn to the back is invented.
Prompt template: a 30-second one-take
Alibaba's launch post says the 30-second canvas leaves room for continuous camera moves and one-take sequences. Make the character turn at least once while the side and back views are in the request, so the model has a reference for every angle it shows.
A 30-second, 16:9 one-take short about [Name], [one-line story]. Image 1, Image 2 and Image 3 are the only references for [Name]: front, side and back of the same character. Image 4 sets her height and proportions. Image 5 is her outfit, Image 6 is the [product] she holds, Image 7 is the location. Use the references for appearance only, not for pose, background or lighting. Shot 1 [0-5s]: [opening action] Shot 2 [5-12s]: [movement + camera] Shot 3 [12-20s]: [main action, she turns so her side and back are seen] Shot 4 [20-26s]: [reaction or detail] Shot 5 [26-30s]: [closing pose] One continuous shot, no cuts. Style: [art style]. Keep her face, hairstyle, hair color, body shape, outfit and accessories identical throughout. No costume changes, no scale drift, no extra people, no subtitles, no watermark.
A creator on X reported a 30-second Japanese 2D animation that followed one uploaded character image closely. That is a creator report, not something we have reproduced. For anime looks, state the style once and exclude the others in the negative list, for example "no photorealistic live-action, no 3D render".
Character sheet plus a storyboard
Wan 3.0 also reads a multi-grid storyboard as a plot outline. Alibaba's guide says to declare it as shot guidance only, so the grid isn't animated as one picture, then describe each panel in order. Put the character views first and the storyboard after them.
Use Image 4 only as shot guidance; do not generate the video from the storyboard as a single image. Image 1, Image 2 and Image 3 are the front, side and back of [Name]; take her appearance only from them. Panel 1 [0-4s]: [shot size, camera, what Name does] Panel 2 [4-9s]: [..] Panel 3 [9-15s]: [..] Keep [Name] identical in every panel. Style: [art style], 16:9. No subtitles, no [unwanted style].
Longer than 30 seconds
Extension counts the input clip: input plus output must stay within 30 seconds, and the input seconds are billed. So extend a shorter clip, for example 15 seconds plus 15, rather than a 30-second one. Say "extend" and the direction, restate the character's fixed traits, and add the same character images again; the guide recommends duration -1 or the sum of both lengths. For a longer film, cut separate shots that share the same images, order and identity sentence.
Extend Video 1 backward by 15s. [Name] is [fixed traits: face, hair, outfit, accessories], exactly as in Image 1, Image 2 and Image 3. Starting from the last frame, [new action and camera]. Keep the location, lighting and style of Video 1. No costume changes, no extra people.
What one attempt costs on the API
Based on Model Studio's Singapore (international) prices. The global regions are about 18% lower.
- 15-second test at 480P: $0.75
- 15 seconds at 720P: $1.50
- 30 seconds at 720P: $3.00
- 30 seconds at 1080P: $6.00
Reference images are not listed as a separate charge. Extending a 15-second clip by 15 seconds bills 30 seconds, because the input clip counts.
Wan 3.0, Seedance 2.5 or Kling 4.0
All three accept separate front, side and back views and make clips up to 30 seconds. Wan 3.0 takes up to 10 reference images, cannot combine them with a first frame, and lists $0.10 per second at 720P. Seedance 2.5 takes up to 30 images, can take a first frame as one more reference, and its guide lists $0.231 per second at 720p without an input video. Kling 4.0 Omni Reference also takes 10 images. Choose by how many references the shot needs and whether you need a locked first frame, then reuse the same pack files.
Questions
- Can I use a first frame and character references together in Wan 3.0?
- No. The API reference says the two modes are mutually exclusive, and a request with both returns an error. Use reference mode with the opening described in the prompt, or first-frame mode with a starter frame that already shows the character.
- How many character images can Wan 3.0 take?
- Up to 10 reference images per request, numbered Image 1 to Image 10 in upload order. Front, side, back, full body, outfit, prop and scene are seven, so three remain.
- Can Wan 3.0 make a 30-second continuous shot of the same character?
- It can generate up to 30 seconds in one pass. Write "one continuous shot, no cuts", give timed shots of 2 to 5 seconds, and keep the side and back views in the request so the character can turn without the model inventing those angles.
- The outfit or size changes halfway through the clip.
- Repeat the fixed traits in words, not just in images, and add the consistency negatives from Alibaba's guide: "no costume changes, no scale drift, no character position changes". Fewer, slower actions per shot also help.
- Does Wan 3.0 work with anime or 2D characters?
- Alibaba's guide lists virtual characters as a valid subject reference. Name the style once in the prompt and exclude the styles you don't want in the negative list. Creators have posted 2D anime results, but we have not tested it.
- Can Wan 3.0 put my character into a video I already have?
- Yes, through video editing: pass the clip as a reference video (1 to 15 seconds), add the character views as reference images and say in the prompt who to replace and what to keep. The input seconds are billed too. Alibaba also has a dedicated swap model, Wan2.2 Animate Mix, which takes one character image and a clip of up to 30 seconds. Our character swap guide compares both.
Start with a free character
Every character's full 19-image pack is free: eight starter characters need no account, the rest a free sign-up. All nine example ads use pack images only.
Browse free characters