Simple Google Translate Text

The name is the whole story: one node, a text field, two language dropdowns, and it hands you back a translation - right inside the graph, at queue time, with no API key and no account. It sits in the same "small utilities that make the graph usable" territory as the text helpers scattered through packs like WAS Node Suite and Efficiency Nodes, except this one does one job and disappears.

Why would you care? Because the best prompt for your checkpoint might not be in the language you think in. The whole reason this category exists is the community's discovery that text encoders are opinionated about languages: Z-Image's Qwen3 encoder measurably boosts Chinese prompts, Danbooru-tag models want their tags in English, and most 2026 LLM-encoded checkpoints read English prose best. People have been wiring translators into their pipelines since the A1111 days, and this is the ComfyUI-native version of that trick.

How it works

Under the hood it's the deep-translator library's GoogleTranslator backend - an unofficial, free interface to Google's public translate endpoint. No Google Cloud account, no billing, no local model. The source is about as thin as a node gets: every time you queue, it constructs a fresh GoogleTranslator(source=..., target=...) and makes a live HTTP request. That has two consequences worth internalizing. First, it needs internet access at queue time, not just at startup. Second, nothing is cached - re-run the workflow and you pay the round-trip again.

One detail the source makes clear that the README doesn't: if the translation call throws, the node doesn't error out. It returns the string [Translation error] <message> as its output. If you feed that into a positive prompt, your next batch silently conditions on garbage. Worth knowing.

The inputs and outputs

Only three inputs, and you'll set two of them once and forget them:

• source_lang - the original language, default auto. Auto-detection is genuinely good, so you can leave it unless it keeps guessing wrong.
• target_lang - the output language, default en. Both dropdowns use ISO 639-1 codes (en, ja, zh-CN, ko, ...), 133-134 of them, not friendly names. Yes, "which one is Danish?" is a real problem. It's the standard tradeoff for a minimal node.
• text - the content to translate. Multiline, so paste your whole prompt or a batch of captions in.

The single output, translated (STRING), wires straight into a CLIP Text Encode node's text input in place of your prompt - or anywhere else a string goes.

Installing it

Via ComfyUI Manager: search for ComfyUI_SimpleGoogleTranslate and hit install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/KAVVATARE/ComfyUI_SimpleGoogleTranslate

Then restart ComfyUI. The only dependency is deep-translator>=1.11.4, which auto-installs from the requirements.txt on first load. No model files, no weights, nothing heavy.

One trap: the README's own clone command uses ComfyUI-SimpleGoogleTranslate with dashes, which is a dead link (404). The underscore version above is the real repo. Use it.

Where people get burned

The unofficial endpoint is the whole risk surface. Google changes their translate service without warning, deep-translator falls behind, and your node starts emitting [Translation error] ... strings. It happens; the README says as much. Watch for it, and if you rely on translation for paid work, this is exactly the spot where the community recommends a proper API (DeepL or Google Cloud) instead of the free scraping backend. Also keep expectations honest: this translates words, it doesn't understand prompting. Machine-translating a whole Danbooru-tag prompt often garbles the tags that model would have understood perfectly - it shines for natural-language prose prompts, less for tag soup. For that use case, it's the easiest free option in the graph, and it does what it says.