Wan Text to Video

There's a small irony in this node's existence. Wan 2.1 and 2.2 are the most important open-weights video models in the ecosystem - the Apache 2.0 base that half the local video stack is built on. Then Alibaba took the numbered series closed: Wan 2.5 was previewed with a promise of open weights that never came, and 2.6 shipped commercial-only in December 2025. This node is how you use that closed Wan - the API-only Wan 2.6 - from inside ComfyUI, billed through your Comfy account.

So the first question to answer is which Wan you want. If you have the GPU for it, Wan 2.2 runs locally, free, forever, with the entire LoRA ecosystem behind it - that's almost certainly the better option. This node is for when you want Wan 2.6's capabilities (longer clips, cleaner adherence) without the hardware, or when you're already in a paid-workflow context and want everything in one graph.

The inputs

• model - wan2.6-t2v (default) or wan2.5-t2v-preview. Wan 2.6 is the current one; 2.5 is the earlier cloud preview.
• prompt - the interesting bit: it supports English and Chinese prompts natively. That's rare and genuinely useful if you're localizing.
• negative_prompt - what to avoid.
• size - a combo of resolution tier (480p/720p/1080p) × aspect ratio (1:1, 16:9, 9:16, 4:3, 3:4), with explicit pixel dims in the label. The 15-second duration is only on Wan 2.6, and 480p isn't offered on 2.6, so some combinations just won't combine.
• duration - 5, 10, or 15 seconds (15 only for 2.6).
• audio - this is the sleeper feature. Wan's API accepts a reference audio clip (must contain a clear, loud voice, no background music) and the video follows it. Want lip-sync or a voice-over that matches mouth movement? Feed an audio file here.
• generate_audio - if you don't provide audio, this tells the API to synthesize it.
• prompt_extend - AI-assisted prompt expansion, on by default.
• shot_type - single or multi: one continuous shot, or multiple shots with cuts. Only applies when prompt_extend is on.
• watermark - whether to stamp an AI-generated watermark on the result. Off by default.

Mechanism and gotchas

Standard API-node plumbing: ComfyUI POSTs to Alibaba's video-synthesis endpoint through the Comfy proxy, polls the task, and downloads the finished video. Needs a Comfy account, credits, internet - no local model. The output is a VIDEO object; save it with a video node downstream or it's gone when the queue clears.

The practical gotcha is pricing. Video APIs bill per second and resolution tier, and 1080p 15-second clips on Wan 2.6 are not a "try it on a whim" expense - the node's price badge makes it visible before you queue. Iterate on short 5-second 720p clips, lock the seed when you find one, and only then spend the credits on the long expensive pass. And remember the emotional context here: Wan 2.7 came out in April 2026 to a community that mostly shrugged, because the closed Wan had lost the thing that made Wan matter - the weights.