☁️BizyAir Wan2.5 Text To Video

There's a specific reason this node exists as an API call instead of a local checkpoint loader, and it's worth knowing before you go hunting for Wan2.5 weights anywhere else: there aren't any. Alibaba's Wan 2.2 is, as of this writing, the last version of Wan that shipped as open weights. Wan 2.5 launched in September 2025 as a cloud-only preview, with an "open weight will release soon" note attached - and 2.6 and 2.7 followed the same pattern with the same promise, none of them kept. If you want Wan past 2.2, API access through something like this node is currently the only door in.

That context matters for how you should think about this node: it's not a lightweight cloud alternative to a local option you could otherwise run - for Wan 2.5 specifically, this genuinely is the only way to use it.

What it does

Text-to-video generation, with the option to attach audio. You write a prompt describing a scene, set your output size and length, and get a video clip back.

Inputs that matter

• prompt - required, multiline. The only required field on the node.
• audio - optional. Feed in an audio clip to sync with the generated video, rather than relying on the model to invent its own.
• negative_prompt - optional, standard usage.
• size - a big list of preset resolutions, from portrait and square options up through 1920*1080 (the default) and 1080*1920, plus a few more beyond the sampled list. Pick based on your target aspect ratio and platform - vertical for social, 16:9 landscape for the default widescreen option.
• duration - 5 or 10 seconds, default 5.
• prompt_extend - boolean, default on. Straight from the node's own tooltip: this uses a large model to intelligently rewrite your prompt before generation. It noticeably helps short prompts, at the cost of extra processing time. If you've already written a detailed, well-structured prompt, turning this off will save time without losing much quality; if you're typing something quick and loose, leave it on.
• auto_audio - boolean, default on. Per the tooltip: lets the model automatically generate matching audio for the video, and this is lower priority than the audio input - meaning if you've supplied your own audio clip, it wins; auto_audio only kicks in when you haven't.

Outputs

• video - the generated clip.
• actual_prompt - the prompt actually used for generation. If prompt_extend rewrote your input, this is where you can see what it turned it into - genuinely useful for understanding why a generation came out different from what you typed, or for reusing a good rewritten prompt directly next time.
• bizyair_model_name - internal plumbing, maps back to wan2.5-t2v-preview.

How to install it

Ships with the full BizyAir pack.

ComfyUI Manager: search "BizyAir," Install, restart.

git clone:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir.git

Requires a BizyAir API key set before it'll run, same as every cloud node in the pack - and since this is metered video-generation API usage rather than local compute, keep an eye on your account's credit balance, especially if you're testing at 10-second duration or looping over multiple prompts.

Common issues

Don't go looking for a local Wan2.5 checkpoint. If you're used to Wan 2.1/2.2 workflows where you load weights with a UNet loader, this node breaks that pattern on purpose - there's genuinely nothing to download, because Alibaba never released one.

prompt_extend changes what actually generates. If your output doesn't match your prompt's intent, check actual_prompt before assuming the model ignored you - the rewrite step may have shifted the meaning in a way you didn't expect. Turn prompt_extend off if you want exact control over the prompt that reaches the model.

Audio priority is easy to get backwards. If you supplied an audio input and are seeing model-generated audio instead, double check auto_audio isn't somehow overriding it, or that your audio input actually connected - the tooltip states the input takes priority, but a disconnected or empty audio input will fall through to auto_audio behavior.

Generation isn't instant. As with any API video node, expect the request to take real time server-side, longer for 10-second duration. That's normal.