🌐 Gemini Banana AI Text Translation

The display name says "Gemini Banana AI Text Translation," but the default engine doesn't touch Gemini at all - it's google-translate, and it doesn't need your key. That's the first thing to know about this node: it's a multi-engine translation box with a free tier that actually works out of the box, which is rarer than it should be in the API-node world.

Where does a translation node fit in a ComfyUI workflow? Same place every LLM-ish node fits - as plumbing upstream of generation. You're working in Chinese, your Flux prompt wants English, or you pulled a prompt from a Japanese source and need it rephrased before it hits the text encoder. Wire text in, get translated text out, done. For a pack that's mostly image generation, this is the quiet utility that keeps a pipeline running.

How it works

The translation_engine dropdown is the whole node, really. It lists google-translate (default, free), youdao-translate, libre-translate, mymemory-translate, baidu-api, deepl-free, gemini-ai, openai-gpt, and auto-best. The free ones scrape or hit public endpoints - google-translate calls the translate_a/single endpoint directly, so no credentials. The LLM engines (gemini-ai, openai-gpt) need keys from Translation_config.json or the node's api_key field, and give you the context-aware, tone-controlled result that free engines can't. auto-best tries the best available engine and falls back. There's also automatic proxy detection - it checks env vars and common local proxy ports (Clash, Shadowsocks, V2Ray) so it works in regions where the free endpoints need a hop.

The inputs that matter

• text - what you're translating.
• source_language / target_language - 25 options each, auto detection included. Default is auto → zh-CN.
• translation_engine - the free defaults need no key; switch to gemini-ai only if you want quality and have a key.
• quality_mode - standard / high / creative / formal / casual. Only the LLM engines really honor this; the free engines translate flat.
• preserve_formatting and context_aware - keep line breaks, and add context (via context_info) for better technical translation.
• context_info / custom_instructions (optional) - feed it domain context or override the translation style.

Outputs are translated_text, plus detected_language and translation_info if you want to know what the engine decided.

Install

ComfyUI Manager → search ComfyUI_LLM_Banana, or:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_LLM_Banana
cd ComfyUI_LLM_Banana
pip install -r requirements.txt

restart, then find it under the Ken-Chen category.

Gotchas

The free endpoints are exactly as reliable as free endpoints usually are - rate limits and occasional timeouts, especially mymemory and the scraped DeepL methods. If you're building a workflow that must translate every run, budget for a real key and set the engine to gemini-ai or openai-gpt. And if the node keeps failing and you're behind a proxy, check that the automatic proxy detection picked up your port - that's the silent killer for this node outside mainland networks.