Translator

The name sounds like it phones home to Google Translate. It doesn't. UL_Translator runs a translation model locally - no API, no key, no network call once the model's downloaded - and hands back the translated string. It's the working half of the pack's translation pair: UL_TranslatorLoader loads the model, this node feeds text through it.

Why it's here: AnyText's world is often Chinese, but the SD 1.5 checkpoints under it respond best to English prompts. Rather than translating your prompt in a browser tab and pasting it back, you drop this node in the graph and the translation happens inline, every run, offline.

How it works

You connect a loaded TRANSLATE_MODEL, type or wire in a string, and set the source and target languages. The node runs the text through the model and returns the result as a STRING you can route into a prompt input. Both language dropdowns list 22 options - Chinese (Simplified and Traditional), Yue/Cantonese, Tibetan, Japanese, Korean, Spanish, Russian, Arabic, French, German and more - though which pairs actually work depends on the model you loaded. The lightweight default T5 covers English/Russian/Chinese; if you picked an NLLB model in the loader, the broader list comes alive.

The inputs and outputs that matter

• model (TRANSLATE_MODEL) - from UL_TranslatorLoader. Required; this node has no model of its own.
• string (multiline STRING) - the text to translate. Wire it from a prompt node or type directly.
• src_language / tgt_language - from and to. Set these to a pair your loaded model actually supports; a mismatch is the usual cause of junk output.
• max_new_tokens (default 512) - the length cap on the translation. Long prompts can get truncated if this is too low; bump it for paragraph-length text.
• use_fast (default off) - use the fast tokenizer. Leave off unless a specific model complains.
• keep_model_loaded / keep_model_device - hold the translator in memory / on device between runs. On is convenient; off frees VRAM if you're tight.

Output is a single string (STRING) - the translated text.

How to install it

Ships with the pack. ComfyUI Manager, search ComfyUI_Anytext, or:

cd ComfyUI/custom_nodes
git clone https://github.com/zmwv823/ComfyUI_Anytext

then restart. The translation model itself is loaded by UL_TranslatorLoader (into ComfyUI/models/prompt_generator); this node needs no separate download.

Common issues & troubleshooting

Output is garbage or unchanged. Almost always a language-pair problem. The 22-entry dropdowns are the same regardless of model, but a given model only handles some pairs - the small T5 does EN/RU/ZH and not much else. Match src_language/tgt_language to what you loaded, or load an NLLB model for wider coverage.

Long prompt gets cut off. Raise max_new_tokens. The 512 default is fine for a sentence, thin for a paragraph.

Nothing runs. Check the model wire from UL_TranslatorLoader - without a loaded model, this node has nothing to translate with.

Do I even need translation? If AnyText encodes your Chinese directly (via the miaobi_clip option on the loader), you can skip translating the rendered text - you only want to translate the scene prompt into English for the base checkpoint's benefit. Keep the two separate in your head: the text you're drawing versus the prompt describing the picture around it.