Wan Video (API)
Wan is Alibaba Tongyi Lab's video generation family. This page hosts the API-only Wan releases — the versions Civitai runs as a hosted service rather than as downloadable weights.
Choose a version from the selector above. Each version's own description covers what it adds and the resolutions, durations and inputs it accepts.
Which version should I use?
- 3.0 — longest clips and the strongest subject consistency. Start here unless you need something specific from an older release.
- 2.7 — the previous generation, kept available for reproducing earlier results.
Shared across versions
- Text-to-video, and image-to-video from a first frame with an optional last frame for loops and transitions
- Prompt and negative prompt, plus a seed for reproducible results
- Resolution, aspect ratio and duration controls — the available values differ per version
- Billed in Buzz per second of output, so higher resolutions and longer clips cost more
No downloads
These are API models. Alibaba has not published open weights for them — the open Wan line stops at Wan 2.2, which has its own page. There are no files to download here, and LoRAs cannot be applied.
Credit
Wan is created and operated by Alibaba Tongyi Lab, and mirrored on Civitai by CivitaiOfficial. Use is subject to the Wan / Tongyi Wanxiang Terms of Service.
Description
Wan 2.7 is the previous generation of Alibaba Tongyi Lab's video model. It stays available for reproducing earlier results; new work is better started on 3.0.
What you can do here
- Text-to-video at 16:9, 4:3, 1:1, 3:4 or 9:16
- Image-to-video from a first frame, with an optional last frame
- Reference-to-video from up to 5 reference images
- Video editing from an existing clip
- 720p or 1080p output, 2–15 seconds (2–10 for reference and edit)
Comments (30)
Does it still use High and Low? Hopefully it is only one contained situation
Its not open source so its not matter since its cannot be local.
It doesn't matter since no one is able to see what is under the hood. This isn't open-source.
Wan has abandoned us. Don't support it. They have released several models since 2.2 and none of them are open source.
Also, the High Low might have been a good thing, breaking up the model to fit more on consumer hardware. Obviously 1 thing is easier, but if we what better, high quality, more complex and capable, we may need more steps like what they did with High Low and not less steps. We'll see, but if higher end local AI means 'one contained solution' which would likely be limited, for example a 15 billion parameter model. I'd be plenty fine with having a 5 step model that's in essence then a 75 billion parameter model, if that means significantly better AI can be run on local hardware. Offloading and reloading as necessary. We don't know what the future holds, but if you want local AI that competes with paid AI that's now pulling way ahead, I wouldn't write off multiple steps instead of wanting 1 step. If 3 step High, Mid, Low means better quality and longer length, I'll take it. If 5 step means even better, I'll take it. Let's see what happens, but local AI is falling way behind paid Kling, Seedance, Veo, etc. for the past year+ and I'm not sure if that will ever change. Our local hardware won't compete with billion dollar equipment and much larger, much more intelligent models that wouldn't come anywhere near fitting into our VRAM or RAM, so be ready to embrace non-1 step solutions most likely. That's the only way you'll be able to fit a higher parameter model into VRAM.
@civitai7_ i agree with every word
Let's hope they make 2.5 or 2.6 open source. For science.
awesome great news
When i can't test it localy it is only one big meh.
It can be perfect, beautifull, amazing, but when it's locked i can say it is golden poop.
Want it on comfy not here
If WAN is not open source anymore, I don't care about any future release.
Bingo. Very disheartened. I thought they learned their lesson. Cause no way does WAN2.5 have more users than 2.2. 2.7 being closed to the public means dark times ahead. Pony died, hopefully Illustrious and NOob AI's next models won't be closed source onsite only too. :/
LTX2.3 better than Wan2.7, and Seedance 2.0 is best of the best so far i know.
how is it better the face emotions look blank
@PastellPastellPastellI Wan is limited, sure it has better face quality but no way you can make 30+ seconds videos at 50fps and ltx2.3 does exactly what prompt you give. Also the face can be fixed with lora's so yes, Ltx2.3 is wayyyy better than wan.
LTX2.3 is bad at human anatomy. LoRAs don't even fix that 99% of the time. It's great at adding audio to WAN videos but other just frustrating to use. I'm not happy about WAN2.5 and WAN2.7 not being publicly available.
And LTX2.3 does not do exactly what the prompt you feed to it. If you write is novel of text it will but most people don't want to have to write out novels worth of data just to get a model to function.
@DaddyWolfgang same
I agree with @DaddyWolfgang. The only advantage I see to LTXV2.3 is the audio and speech, otherwise Wan2.2 is better at everything else concerning video. Note: You can run 30sec+ vids on Wan2.2 using SVI and a ForLoop System. My personnal best is a 1 min long video with this system.
It's now avaliable for on-site generation, and happily produces nude ladies with good sound.
As someone without a local set-up, the lack of local generation and LORA type support is annoying, but not a deal breaker in and of itself.
Jesus fucking Christ. On-site only? Again? Too bad. Another dead in the water model released.
"check out this cool toy you can't have"
-closed source models
Generation only? no Ally of mine posts closed source. ironic name @theally
Not local?... Not interested.
the example video above doesn't play, only sound is heard.
No local dl ??? who cares...
Did this break? I can't generate any videos based on images anymore. It tries to process for 1-2 seconds and then fails with no explanation.
Not bad, expensive