Seed-Nodes: SiliconFlowVLAPI

SiliconFlowVLAPI is the sibling of the pack's QwenVLAPI node, and the differences matter. Both stick a vision-language model inside your ComfyUI graph. But where QwenVLAPI talks directly to Alibaba's DashScope, this one talks to SiliconFlow - a Chinese GPU-cloud API provider that hosts a whole menu of open models, Qwen-VL among them - and it's OpenAI-compatible under the hood. If you want to caption, describe, or question images with a hosted VL model, this is the more flexible door into that world.

How it works

Your image is base64-encoded as a WEBP, bundled with a prompt and a detail setting, and sent to SiliconFlow's /v1/chat/completions endpoint - the same shape as any OpenAI-compatible API. The model's reply comes back as the response string. The provider does all the heavy lifting, so there's no local GPU requirement and no model download; you just need an API key.

Same three limitations as its sibling: it's a paid cloud call, it uses only the first frame of your image batch, and a missing or bad api_key doesn't raise an error - it returns an error string as the output. Both of those are "read the output before you panic" moments.

The inputs that matter

• api_key - your SiliconFlow key from the provider's console. Non-negotiable.
• prompt - the question/instruction. Default is "描述这张图片的内容" ("describe the content of this image").
• model - the pick of hosted models, and here's where this node earns its keep. The list runs from Qwen/Qwen2.5-VL-32B-Instruct and the bigger 72B, through Qwen/QVQ-72B-Preview (a reasoning model - slower, more thoughtful), the older Qwen2-VL 72B, the smaller Pro/ 7B variants for speed, down to deepseek-ai/deepseek-vl2. Default is the 32B, a sensible quality/speed balance.
• detail - low, high, or auto (default low). Controls the resolution at which SiliconFlow processes your image - high for fine detail like reading text, low to save tokens and money.

Optional image is what gets described. Output is a single response string for whatever text-consuming node comes next.

How to install it

It's part of ComfyUI-Seed-Nodes, so the one-line pack install covers it:

cd ComfyUI/custom_nodes
git clone https://github.com/Aerse/ComfyUI-Seed-Nodes
# restart ComfyUI

Or ComfyUI Manager → search "ComfyUI-Seed-Nodes". Nothing else to download - the models live on SiliconFlow's servers, and the only thing you must bring is that API key.

Common issues

• Error string as the response - missing or invalid key, or a network/API failure surfaced into the output. Check the key, check the request.
• Too slow / too pricey - you're on the 72B or QVQ reasoning model with detail on high. Drop to a Pro/ 7B model and low for bulk captioning.
• Image not included - like its sibling, the node tolerates running with no image attached; that's usually a wiring mistake.

Between this and QwenVLAPI, pick by provider preference and model menu: this one's the better buy if you want the reasoning model or the smaller fast variants, and SiliconFlow's hosted selection makes it a solid pick for batch captioning on a budget.