☁️BizyAir Wan2.6 Text To Video

What it is

If you know Wan as the open video model you download and run locally, this node is not that - and that's the whole story worth understanding before you use it. Alibaba open-weighted Wan 2.1 and 2.2, and then stopped. Wan 2.5, 2.6, and 2.7 exist only as paid cloud APIs, with no weights ever released - Alibaba even ran a "coming soon" disclaimer on 2.5's preview that never resolved into a release. The community called the shift at 2.6 directly: "WAN 2.6 has been released, but it's a commercial version. Does this mean the era of open-source WAN models is over?" This node is that commercial version, wrapped for ComfyUI through BizyAir.

So reach for this specifically when you want Wan 2.6's improvements over the last open release (2.2) and you're fine paying per generation for it - not as a drop-in replacement for a local Wan workflow, because there's no local path here at all.

How it works

Submit a prompt (and optionally a negative prompt, audio, and a few structural options) and BizyAir routes the job to Alibaba's Wan 2.6 endpoint. One notable feature baked into this specific node: prompt_extend, which runs your prompt through a large language model to rewrite and flesh it out before generation - genuinely useful if you're feeding it short prompts, since the tooltip notes the effect is much more pronounced on brief input, at the cost of added generation time.

Inputs and outputs that matter

prompt is the only required field. Everything else is optional but shapes the result meaningfully: negative_prompt for what to avoid, size (a dropdown of specific pixel dimensions from 1280×720 up through 1920×1080 and beyond, defaulting to 1920×1080), and duration (5, 10, or 15 seconds). prompt_extend (on by default) triggers the LLM rewrite described above; when it's active, shot_type (single or multi) becomes meaningful - it decides whether the rewritten prompt aims for one continuous camera move or a sequence of cuts, and the tooltip is explicit that it only takes effect when prompt_extend is true. auto_audio (on by default) lets the model generate its own audio track, but it's explicitly lower priority than the audio input - if you attach your own audio, that takes precedence over auto-generation.

Outputs: video is your clip, actual_prompt returns the prompt the model actually used (useful for seeing what prompt_extend did to your original text), and bizyair_model_name is an internal routing tag for logging.

Installing it

Install via ComfyUI Manager (search "BizyAir") or git clone https://github.com/siliconflow/BizyAir.git into custom_nodes, then restart. A BizyAir API key is mandatory before any node in this pack, this one included, will run - set it up through the login prompt on first launch.

Common issues

If you came here expecting to find Wan 2.6 model files somewhere to download, stop looking - they don't exist and never will per Alibaba's own pattern with 2.5 and 2.7, so any site offering "local Wan 2.6" downloads is not legitimate. If shot_type doesn't seem to be doing anything, check whether prompt_extend is actually on - it's a silent no-op otherwise. And check the actual_prompt output if a generation comes out unexpected; with prompt_extend enabled, the model isn't literally using the text you typed, and comparing the two is the fastest way to see what changed.