GALIAIS-Nodes AI Provider

If you're reading this, you've found a ComfyUI pack that wants to hand your prompt to a language model before the sampler ever sees it. Every AI-assisted node in GALIAIS-Nodes - the tag analyzer, the conflict resolver, the style enhancer, all of them - plugs into this one node first. It's not a generator and it doesn't touch pixels. It's a config block that packages your endpoint, key, and model into a little dict that the other AI nodes read, plus a second dict for an embedding model.

The name is honest about the job: this is where you point the pack at an OpenAI-compatible API. That means any provider that speaks the /chat/completions dialect - OpenAI proper, OpenRouter, a local Ollama or LM Studio box, a vLLM server, a Chinese API like DeepSeek or Qwen, whatever. You bring the URL and key, the pack brings the client. There's no model download, no bundled weights, no local VRAM spend. The flip side is that every AI-assisted run is a live API call, so cost and a working key become your problem. That's the tradeoff to accept before you build a workflow around the AI side of this pack.

What you actually set

The required inputs are 服务商URL (provider URL), API密钥 (API key), 模型 (model), and Embedding模型 (embedding model). Note the Chinese labels - this pack's UI is Chinese-first, and "AI服务商" literally means "AI service provider."

• 服务商URL - base URL of the API, like https://api.openai.com/v1. The 接口模式 (API mode) dropdown (自动 / 强制/v1 / 保持原样) decides whether the pack appends /v1 for you. Auto handles most cases; use 保持原样 if your provider already includes it.
• 自动获取模型 (auto-fetch models) - when on, the node calls the provider's /models endpoint at runtime and picks the first available chat model if you left 模型 blank. Handy for Ollama where you don't remember the exact tag name.
• 温度 (temperature, default 0.35) and 最大Token (max tokens, default 1200) - the generation knobs. 0.35 is a good place for structured tag output; don't crank it if you want stable results.
• The rest is resilience: 重试次数 (retries), 重试退避秒 (backoff), 流式失败降级 (stream fallback), and 超时秒 (timeout). If you're on a flaky local server, retries save you.

The outputs include the AI服务商 (the typed provider object), a plain 模型 string, 可用模型JSON (the model list, if you fetched it), 状态JSON (a status dump with the key masked), and the Embedding服务商 object used by the RAG-style tag selection in the character nodes.

The two traps

First, the pack will happily build a "provider" with an empty key and fail silently on the first AI call with a confusing auth error - the health check node exists to catch exactly this, so use it. Second, this is an external-API node by design: it carries a credential and phones home. That's the normal shape of the ecosystem, but it's also the exact shape of the one malware node that got people prosecuted. Install from the official repo, not a sketchy mirror, and don't paste your key into screenshots you share.

Install and gotchas

Install the whole pack once - every node in it shares the same install:

git clone https://github.com/GALIAIS/GALIAIS-Nodes.git ComfyUI/custom_nodes/GALIAIS-Nodes

Restart ComfyUI. ComfyUI Manager also finds it if you search "GALIAIS-Nodes". There are zero pip dependencies and no bundled model files - it's pure Python with its own HTTP client, so nothing heavy to download. The pack is brand new (0 Google impressions, no real community chatter yet), so don't expect tutorial videos; the README's recommended flow puts this node at step 3, right after the Danbooru DB loader.