☁️BizyAir Wan2.5 Image To Video

Every other Wan node in this pack - the WanImageToVideo latent-graph node, the Diffusers pipeline, the T2V pipeline - is running an actual open checkpoint (Wan 2.1) on BizyAir's cloud, and you could in principle run that same model yourself if you had the hardware. This node is different in kind, not just in convenience. Wan 2.5 never got open weights. Alibaba shipped it in September 2025 as a cloud-only preview, with an explicit promise that open weights would follow - a promise that never happened, and the same pattern repeated with 2.6 and 2.7 afterward. As of now this API node is the only way to touch Wan 2.5 from ComfyUI, full stop, because there's nothing to download.

That's worth knowing before you go looking for a local alternative - there isn't one, not because you're missing a step, but because Alibaba never published one.

Why you'd actually reach for this over free local Wan 2.2

One real, concrete reason: audio. Base Wan, at every open version, has zero native audio generation - none. It's a documented gap in the open ecosystem, and people route around it with a separate LTX or MMAudio stage bolted on. This closed 2.5 API tier, by contrast, exposes an audio input and an auto_audio toggle directly. If you need image-plus-sound-in, video-plus-sound-out in one call and you're willing to pay for it, this is a real capability gap the open weights genuinely can't close on their own.

Inputs and outputs that matter

• image - required, the starting frame.
• audio (optional) - feed in your own audio track and Wan 2.5 syncs the video to it.
• prompt / negative_prompt (optional) - text guidance.
• resolution - 480P, 720P, or 1080P, defaulting to 1080P.
• duration - 5 or 10 seconds.
• prompt_extend (default true) - per the node's own tooltip, this runs your prompt through a larger model to rewrite and expand it before generation. It helps noticeably on short prompts and costs extra time - leave it on if you're typing a quick one-liner, turn it off if you've already written a detailed prompt and don't want it second-guessed or the added latency.
• auto_audio (default true) - has the model generate its own audio automatically. Per the tooltip, it's lower priority than the audio input: if you supply your own audio track, that wins regardless of this setting.

Outputs: video, actual_prompt (the prompt Wan 2.5 actually used - worth wiring to a text display when prompt_extend is on, so you can see what it rewrote your input into), and bizyair_model_name for logging.

How to install it

• ComfyUI Manager - search "BizyAir", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git, restart.
• Comfy CLI - comfy node install bizyair.

Set your API key first - the in-app login prompt or bizyair.siliconflow.cn. Every call here is metered against your BizyAir credits; there's no free tier to fall back on since the model has no open weights to run locally in the first place.

Common issues & troubleshooting

Generated video doesn't match your prompt closely. Check actual_prompt before assuming the model misread you - if prompt_extend is on, what actually ran through generation may be a meaningfully rewritten version of what you typed.

Your supplied audio doesn't seem to sync, or gets ignored. Confirm auto_audio isn't overriding your intent in the other direction - it only kicks in when you haven't provided an audio input, so this shouldn't normally be the cause, but it's worth double-checking the audio input is actually wired before assuming a model failure.

Slower than expected. prompt_extend adds a real LLM rewriting pass before generation even starts. If speed matters more than the quality bump on short prompts, turn it off.