越光 API 连接 / YueGuang API Connector

Paying full price for every LLM API you touch is a hobby this node quietly kills. 越光 API 连接 / YueGuang API Connector is a connector for YueGuang (Nebula), a Chinese LLM aggregator that fronts OpenAI, Anthropic, DeepSeek and Kimi behind one OpenAI-compatible endpoint (llm.ai-nebula.com/v1/chat/completions). One API key, 25 models, and - the genuinely nice touch - every dropdown option has its price baked into the label, like deepseek-v4-flash [入$0.1/M 出$0.3/M] (in/out per million tokens). The default model is the cheapest thing on the list, deliberately, so a mistriggered node costs you pocket change.

It's a sibling of the pack's ZenMux node with the same interface, so you can swap aggregators by rewiring one node. If you want the same shape with 130+ models and vendor-prefixed ids, that's the ZenMux article.

What it does

Text generation and vision (up to six images) through one chat-completions call. The model registry is built into the pack (yueguang/model_registry.py) rather than fetched live - YueGuang has no enumerable models endpoint, so the list comes from the official spec docs and updates mean editing that table. No network round-trip for the list, and no risk of an empty dropdown because a fetch failed.

Two engineering habits carry over from the author's API-node work and both save you real pain. First, base_url defaults to a value without :// - the ComfyUI frontend eats the protocol fragment, so the backend appends it. Don't type https://, or you'll wonder why the URL comes out mangled. Second, adaptive retry: some newer models deprecate temperature or demand max_completion_tokens instead of max_tokens. When the API answers 400, the node strips or renames the offending parameter and retries automatically - normal requests cost nothing extra.

Inputs and outputs

• api_key - your YueGuang key (sk-...). The workflow saves it into the JSON, so clear it before sharing.
• model - 25 price-tagged choices; search gpt / claude / deepseek / kimi to filter. Note ids have no vendor prefix here (gpt-4o, not openai/gpt-4o) - same-vendor options cluster together in the sort.
• system_prompt / user_prompt - role and the actual ask. Output-format rules ("only return JSON") belong in the system prompt.
• seed - mostly a "force re-run" lever; most of these models won't truly reproduce anyway.

Optional: temperature, top_p (leave at 1.0 and drive with temperature), max_tokens, image_1…image_6 (vision models only, resized to image_max_size and base64'd), detail (low/auto/high), image_max_size (default 1024 - smaller is cheaper), proxy_url (127.0.0.1:7890 style, no protocol), usd_to_cny (only affects the cost display).

Outputs: text, model_id, and usage_stats - a five-line cost report: token counts, output characters, vendor, model with unit prices, and the dollar/CNY conversion.

Gotchas

If a vision model ignores your images, you picked a text-only model - the dropdown doesn't gate it for you. If a call errors and you don't see a retry note, check the ComfyUI console; the parameter-fix path logs there. And if your usage_stats looks wrong, check usd_to_cny against today's rate - it's display math only, never the actual bill.