☁️BizyAir SiliconCloud VLM API

What it is

This is a general-purpose vision-language model node - you hand it an image and a question, it hands back a text answer. SiliconCloud is BizyAir's parent company's own hosted-inference platform (BizyAir is siliconflow's ComfyUI-facing product; SiliconCloud is the broader API), so this node is essentially "ask a VLM about this picture" wired straight into your workflow instead of a separate API call you'd otherwise have to script yourself.

It's the flexible, unopinionated cousin of the pack's other vision nodes (like BizyAirJoyCaption3, which is purpose-built for training captions). Reach for this one when you need something more general: describing an image for review, extracting specific details on demand, or building a VLM-driven branch into a larger automated workflow.

How it works

You supply a system prompt (the model's persona/instructions), a user prompt (your actual question about the image), and the image itself. The model reads the pixels and answers in plain text, same as any chat-style VLM API. The defaults ship in Chinese - "You are an AI assistant that can analyze images. Please observe the image carefully and provide detailed, accurate descriptions based on the user's question" is the system prompt's gist, and the user prompt default asks it to describe the image and call out anything unusual - but there's nothing stopping you from writing your own prompts in English or any other language the underlying model handles.

Inputs and outputs that matter

model is a dropdown, but note the brief for this node shows it with zero populated choices - in practice this list is filled from what's available on your BizyAir account, so don't be surprised if it's empty until you're authenticated. system_prompt and user_prompt are both multiline text fields carrying the defaults described above; rewrite both for your use case. images is the picture you're asking about. max_tokens (default 512, up to 100,000) caps the answer length - bump it up if you want a genuinely long description and not just a caption. temperature (0–2, default 0.7) controls how much the model varies its wording - lower it toward 0 for consistent, literal answers; raise it if you want more creative phrasing. detail (auto/low/high) trades off how closely the model examines the image against speed and cost - high for fine detail work, low when you just need a rough read.

The single output is STRING - the model's text answer, ready to feed into a prompt field, a save-text node, or anything else downstream that consumes plain text.

Installing it

Install through ComfyUI Manager (search "BizyAir") or clone manually with git clone https://github.com/siliconflow/BizyAir.git into custom_nodes, then restart. As with every node in this pack, you need a BizyAir API key set up first - the in-app login prompt on first run handles registration.

Common issues

If model shows no options, that's almost always an authentication problem - get your API key configured and restart, since the list is populated from your account rather than hardcoded. Vague or generic answers usually trace back to the default prompts still being active; write a specific user_prompt rather than relying on "describe this image," which is deliberately broad. And because this is a metered cloud call like everything else in the pack, a workflow that runs this node in a large batch loop will draw down your BizyAir credits fast - worth testing on a single image before scaling up.