Siliconflow free chat

You want an actual language model rewriting your prompts, but you've already spent your VRAM budget on a 12GB checkpoint and a controlnet stack. FreeChat-2lab is the workaround: it phones home to SiliconFlow's free LLM tier instead of loading anything locally. No model download, no extra VRAM, and on the free-tier models, no money. It's a thin node - two inputs, one string out - and for prompt work that's all you need.

FreeChat (2lab) is one half of the comfyUI-siliconflow-api-2lab pack from AI2lab, a small MIT-licensed wrapper around SiliconFlow's OpenAI-compatible API. Its sibling PaidChat points at the big paid models; this node's dropdown is the free tier: Qwen2-7B-Instruct (the default), Qwen2-1.5B, Qwen1.5-7B-Chat, GLM-4-9B-chat, ChatGLM3-6B, and two Yi-1.5 models. "Free" here is the platform's free quota, not the repo's license - the code is MIT, the API tokens are on them, up to whatever their free-tier limits are.

How it works. Run the node and it opens an OpenAI-style chat completion against https://api.siliconflow.cn/v1, sends your prompt as the user turn with a hardcoded "You are a helpful assistant" system message, and returns the reply as a text STRING. The call is synchronous, so the queue blocks while the model thinks - fine for a 7B, slower on the bigger tiers. The API key is loaded from config.json at startup, not per-run, which matters for troubleshooting (below).

The two inputs you'll actually touch:

• prompt - multiline text. What you're asking the model to do. For prompt work, be explicit: "Rewrite this as a detailed Flux prompt: …" gets you much further than "help me."
• model - the free-tier dropdown above. Start on the default Qwen2-7B; it's the best quality-to-latency trade in that list.

There's no temperature, no max tokens, no seed - the node doesn't expose them. If you need knobs, this isn't the node; you wire its output into something downstream instead.

What the text output plugs into. Anything that takes a STRING. The obvious move is feeding it straight into a CLIPTextEncode's text input so the model's answer becomes your generation prompt. Or pipe it to a text-display node so you can see what it wrote before you commit. It's a prompt-expansion machine, essentially - the same pattern as having an LLM chat-turn to conditioning, minus the local encoder.

Install. Via ComfyUI Manager, search for "comfyUI-siliconflow-api-2lab" and install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/AI2lab/comfyUI-siliconflow-api-2lab

Dependency-wise it's one package (openai) and zero model files. Restart ComfyUI. On first import the pack creates config.json in its own folder (from config_template.json) - open it and put a real API key from https://cloud.siliconflow.cn/ in place of the "your key" placeholder.

Where people get burned:

• 401 / auth errors. You left the placeholder key in config.json, or the key is invalid. Get a key, paste it, restart.
• You edited config.json and it still fails. The key is read once when the pack loads. A ComfyUI restart is not optional after editing.
• The pack doesn't load at all. Same file, one step earlier - if config.json is malformed JSON or lacks API_KEY, the import throws and the node vanishes. Check that file first.
• Getting a key is the actual wall. The console the README points to (cloud.siliconflow.cn) wants account verification, and that can demand a Chinese phone number - a real complaint from people outside China. If that blocks you, the international console at siliconflow.com is the path that works.

One more thing worth saying, because after the ComfyUI_LLMVISION malware mess the community is (rightly) wary of LLM nodes that phone home: this one is about 90 lines of MIT code you can read yourself, and its only outbound traffic is the SiliconFlow API with your key. That's about as auditable as a remote-API node gets.