Alibaba Wan 2.7 Text To Video (Official Edition)

Wan 2.7 is Alibaba's newest video model, and here's the thing you need to know up front: it's not open weights. The community grumbled about exactly that when it launched in 2026 - "released or provided as a service only? That's not the same thing" was the general mood. But that's precisely what this node is for. It hands you Wan 2.7's text-to-video through an API, inside your normal ComfyUI graph, with zero local VRAM involved. If you've got a decent prompt but a decidedly not-decent GPU, this is the honest way to use the model.

It's part of the BizyAirPlus pack, which is SiliconFlow's cloud-execution layer for ComfyUI: you build and edit workflows locally, flip the BizyAirPlus button ON in the action bar, and the graph runs on their servers instead of your machine.

How it works

The node is a thin client. You give it a prompt, it uploads your request to the BizyAir cloud, the server runs Wan 2.7, and the result comes back as a video output you can wire straight into a SaveVideo node. There's no model download, no diffusers install, no CUDA. Your machine only has to edit the graph and display the result. Alongside video, every run also emits request_id (useful if you need to chase a failed job with support) and urls (the hosted file links, handy if you want to grab the raw file).

Cost is the real gotcha, and the author is upfront about it in the tooltips: resolution and duration "directly impact costs." This is a pay-per-generation node, not a free local one.

The inputs that matter

You set surprisingly few things for a first good run:

• prompt - up to 5000 characters, and it supports controlling single- or multi-shot narration in plain language. That's Wan 2.7's party trick: describe a scene change mid-sentence and it'll cut.
• resolution - defaults to 720P. Higher tiers cost more. Start at 720P to learn the API before spending.
• ratio - defaults to 16:9, combined with resolution to set the output pixel size.
• duration - 2–15 seconds, default 5. It "directly affects cost," so keep it short while you iterate.
• prompt_extend - off by default. Flipping it on runs your prompt through a large language model first; it visibly improves short prompts but slows the job. Leave it off until your prompt feels weak.
• watermark - adds a fixed "AI generation" mark in the lower-right corner. Off by default; flip it on if you need provenance.

Beyond that: negative_prompt (≤500 chars), seed (-1 = server-random), and an audio input where you can attach a WAV/MP3 (2–30s, ≤15MB) that the model uses as a motion/audio drive source - or it auto-generates background music and sound effects when you leave it empty. Two safety valves worth knowing: skip_error lets the run continue with a placeholder instead of killing your workflow, and api_config overrides the base URL/API key for that single request.

Installation

Install the whole pack once via ComfyUI Manager (search "BizyAirPlus") or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAirPlus.git
cd BizyAirPlus
python -m pip install -r requirements.txt

Restart ComfyUI, then grab an API key at bizyair.ai. Set it through the prompt that appears when you turn BizyAirPlus ON, under Settings > BizyAirPlus > API Key, or via the BIZYAIR_API_KEY environment variable.

Common issues

• Button missing after install - the dependencies were probably installed with the wrong Python. Reinstall requirements.txt using the same interpreter ComfyUI runs, then restart.
• "API Key missing" - set it via the settings or env var before starting ComfyUI.
• Job fails mid-run - check the key is valid and the network is up; detailed errors land in the ComfyUI console.
• It's hijacking your local workflows - the auto-installed pack turns its cloud mode on via that action-bar button, and more than one person has had a local run stall because BizyAir was sitting ON. Click it so it reads OFF whenever you want to run locally.