Correction, 2026-10-01: this page listed base Wan 3.0's reference-to-video mode (34, 68 or 135 credits for 5 seconds) next to the canvas prices without saying where it runs. It is available only through the 8frame MCP connector, and it starts from a reference clip, not reference images. On the canvas, base Wan 3.0 takes text or one start image, and Wan 3.0 Prime is the Wan that takes a reference clip. We also sent readers to Kling v3 for a specific face held across shots, but on 8frame Kling v3 takes one start image and no reference images, so the page now names the models that take reference images and no longer promises those models "the cleanest possible detail", which we never measured. A second pass the same day found more. We called Wan 3.0 the only Wan that runs past 15 seconds, but Wan 3.0 Prime also runs 2 to 30 seconds, and we said most video models on 8frame stop at 15 seconds without naming Seedance 2.5, which runs 4 to 30. We did not say that on 8frame Wan 2.2 runs only from a start image or an input video (no text-to-video there), or that the Wan node opens at 1080p, its most expensive tier. We removed the 10- and 15-second Wan 2.5 prices, because we have not confirmed that those settings return clips of that length. And two statements we never measured, about what vague prompts do on every model and about which subjects expose a model's weaknesses, are now advice.
Wan AI is Alibaba's video generation model series. Wan 2.2 is the last release with downloadable weights (Apache 2.0); Wan 2.5 and Wan 3.0 are available through APIs only. On 8frame, Wan 3.0 is the default Wan: 24, 48 or 95 credits for 5 seconds at 480p, 720p or 1080p, with single takes from 2 to 30 seconds. The Wan node opens at 1080p, so set the resolution before you generate.
Correction, 2026-09-30: the first version of this page described two prompts as tested output with generation times and results we cannot reproduce, gave Wan 2.5 a 7-second clip, a length it does not offer, and didn't cover Wan 3.0, which is now the default Wan on 8frame. This version uses registry prices and the published model card.
The Wan versions on 8frame
| Version | 5 seconds | Resolution | Inputs on the canvas | Weights |
|---|---|---|---|---|
| Wan 3.0 | 24 / 48 / 95 | 480p / 720p / 1080p (opens at 1080p) | Text or one start image | API only |
| Wan 3.0 Prime | 46 / 95 / 189 | 480p / 720p / 1080p | Text, one start image or a reference clip | API only; audio by default |
| Wan 2.5 | 102 | 1080p only | Text or one start image | API only |
| Wan 2.2 | 50 | 720p only | One start image or an input video (no text alone) | Apache 2.0 |
Wan 3.0 and Wan 3.0 Prime bill per second for any length from 2 to 30 seconds: 30 seconds of base Wan 3.0 costs 142, 284 or 567 at 480p, 720p or 1080p. Base Wan 3.0 also has a reference-to-video mode (34, 68 or 135 for 5 seconds), but only through the 8frame MCP connector, where it starts from a reference clip. On the canvas, base Wan 3.0 takes text or one start image, and Wan 3.0 Prime takes a reference clip at its usual prices. A credit is about $0.01 at pack rate; output carries no watermark on any plan.
How Wan AI works
Wan 2.2, the release with a public model card, is a video diffusion model: you describe a scene, subject, camera move and light, and the model generates the clip by progressively denoising from noise, conditioned on your prompt and, if you give one, a start image.
For Wan 2.2, the weights are public, so you can run it on your own hardware. The Wan2.2-TI2V-5B model card lists text-to-video and image-to-video at 720p and 24fps, a 24 GB VRAM minimum and a 5-second clip in under 9 minutes on a single consumer GPU. Wan 2.5 and 3.0 have no published weights; you reach them through a hosted API. On 8frame all of them run on managed infrastructure: add a Wan node, pick a version, write the prompt and generate. On Wan 3.0 and Prime, also set the length and the resolution. Wan 2.5 is fixed at 1080p and Wan 2.2 at 720p, and on 8frame Wan 2.2 needs a start image or an input video.
Be explicit in the prompt: name a light direction, a colour temperature and a camera move. If you leave them out, the model chooses for you, so check the light and the background on a cheap 480p draft before you scale up.
When you use Wan AI
Budget iteration. Wan 3.0 at 480p (24 credits for 5 seconds) is one of the cheaper ways on 8frame to check a composition before spending 448 on an 8-second Veo 3.1 Standard clip with sound. Switch the node from its 1080p default first.
Long single takes. Wan 3.0 generates up to 30 seconds in one take: slow pans, ambient loops, establishing shots. Wan 3.0 Prime also runs 2 to 30 seconds, and Seedance 2.5 runs 4 to 30.
Self-hosting. Wan 2.2 is the option if the model has to run on your own hardware, for cost at volume or data-privacy reasons. The hosted Wan 2.2 on 8frame starts from a start image or an input video; for a text prompt there, use Wan 3.0.
Abstract and texture b-roll. Water, sky, fog and transition overlays are cheap to draft on Wan 3.0 at 480p or 720p. Whatever the subject, look closely at faces, hands and fine architecture before you commit a budget model to them.
Move to another model when the clip needs synchronized sound in the same pass (base Wan 3.0 exposes no audio setting; Kling v3 and Veo 3.1 have a sound switch, and Wan 3.0 Prime generates sound by default) or a specific face carried from reference images. Wan 3.0 and Kling v3 take one start image and no reference images on the canvas; Seedance 2.5 takes up to 9, Grok Imagine's reference mode 1 to 10, Gemini Omni Flash 1 to 7 and Veo 3.1 multi-reference 1 to 3. Limits and rules are in reference image limits by model.
Examples to try
Abstract water texture:
Extreme close-up of rippling water surface with light from above refracting into shifting gold caustic patterns. Camera static, straight down. Warm 4500K light source directly overhead. Gentle continuous ripple motion, loop-friendly.
Nature establishing shot:
Tall pine forest on a misty morning. Dense fog sits low between tree trunks. Dappled early light filtering through the canopy from directly above, cool 4000K. Camera very slowly pans right.
Use Wan 3.0, and set 16:9, 720p and the length on the node (it opens at 1080p). What to check: caustics that shift rather than flicker, a loop point you can hide in the edit, and fog that moves as a volume rather than a flat overlay. On Wan 3.0 at 720p, each 5-second take is 48 credits.
FAQ
Is Wan AI open source?
Wan 2.2 is: its weights are published under Apache 2.0. Wan 2.5 and Wan 3.0 have no public weights as of our check of Wan-AI's Hugging Face models on 2026-10-01. More in is Wan 3.0 open source.
Which Wan should I use on 8frame?
Wan 3.0 for most work: it's the cheapest Wan at 480p and 720p, it takes text or a start image, and it and Wan 3.0 Prime are the only Wans that run past 15 seconds. At 1080p, Wan 3.0 is 95 for 5 seconds against Wan 2.5's 102. Pick Wan 2.2 (50 for 5 seconds at 720p) when you want the hosted version of the model you could also run yourself; on 8frame it needs a start image or an input video.
Is Wan free?
The Wan 2.2 weights are free to download under Apache 2.0; you supply the GPU. On 8frame, Wan generation is paid: the Starter plan is $19/month for 1,000 credits, about 20 Wan 3.0 clips of 5 seconds at 720p.
Related concepts
Wan prompts for b-roll covers prompt formulas and failure modes. Hailuo H3 vs Wan 3.0 compares the two budget families. What is b-roll? covers the concept Wan is most often used for.
Sources
- Wan-AI repositories, newest first (no Wan 2.5 or 3.0 weights; the newest uploads are Wan 2.2 Animate, 2026-08-06): Hugging Face API, author Wan-AI, checked 2026-10-01
- Wan2.2-TI2V-5B model card (Apache 2.0, video diffusion model, text-to-video and image-to-video at 720p and 24fps, 24 GB VRAM minimum, a 5-second 720p clip in under 9 minutes on one consumer GPU): huggingface.co/Wan-AI/Wan2.2-TI2V-5B, checked 2026-10-01
- Wan 3.0 and Wan 3.0 Prime in Alibaba's own model list: Alibaba Cloud Model Studio, newly released models, checked 2026-10-01
- 8frame Wan prices, resolutions, lengths and the 1080p default, plus Seedance 2.5 lengths and the Veo 3.1 Standard price: the 8frame tool registry and content fact sheet, read 2026-10-01
- Wan 2.2 input requirement on 8frame (start image or input video) and Wan 2.5 inputs: 8frame video generation code, read 2026-10-01
- Wan 3.0 inputs on the canvas and reference-to-video through the MCP connector: 8frame tool configuration and MCP connector tool definitions, read 2026-10-01
- Kling v3 start-image-only input and reference image limits for Seedance 2.5, Grok Imagine, Gemini Omni Flash and Veo 3.1: 8frame tool configuration and per-model generation code, read 2026-10-01
Run Wan without a GPU, and set the resolution on the node before you generate.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.