Qwen Translator (KO→EN)

If you write prompts in Korean and generate with English-prompt models, this is the node that kills the copy-paste-into-DeepL step. You type a Korean sentence into a box, and out comes a finished, flowing English prompt ready to feed your checkpoint - no API call, no key, no internet connection required. The whole thing runs on whatever GPU you already own.

It's the second half of the comfyui_qwen_translator_node pack. The first half, Load Qwen Model, loads Qwen3-8B into VRAM and hands you a qwen_model handle; this node takes that handle and does the actual translation. It's the same "drop an LLM into the graph as a tool" pattern that's become completely mainstream in ComfyUI - a local worker that runs before generation, not the text encoder baked into your checkpoint. The trade-off is the same one everyone hits with local 8B models: the ceiling is "good enough," not "frontier API." It fixes the blank-page problem and it does it privately and free per call.

How it works

Reading the source, the flow is: it builds a chat template from your system_prompt and korean_text, and notably passes enable_thinking=False - Qwen3's default reasoning mode is switched off, so you get a direct answer instead of a chain-of-thought essay. Generation runs at a fixed temperature of 0.1 with repetition_penalty of 1.1, tuned for deterministic, clean output. Only the newly generated tokens get decoded back to text, so you don't get the input echoed back at you. It's uncomplicated, and for a translation node that's a feature.

The inputs that matter

• qwen_model - the handle from Load Qwen Model. Wire the two nodes together and you're done.
• system_prompt - the author's default is the terse "Translate the following Korean text into an English prompt suitable for AI image generation. Output only the English prompt." But the example workflow ships a much richer version: a full "expert AI image generation prompt writer" persona with step-by-step instructions, style rules, and worked Korean→English examples. That one makes a visible difference in output quality, so steal it.
• korean_text - your raw Korean, multiline.
• max_new_tokens - caps the answer length; default 200, range 32–1024 in steps of 8. Bump it toward 400–512 if your prompts come back truncated.

Output: english_prompt, a plain STRING. Wire it into your positive prompt text (a CLIP Text Encode, a text display node, or this pack's own Show Text node) and the workflow just runs.

One nice property hiding behind the name: KO→EN is really the system prompt's job, not a hard-coded language pair. Qwen3-8B handles plenty of languages, so you can retask this node by rewriting system_prompt - say, Japanese→English or English→a tag-style prompt.

Installing it

Same pack, same steps as its sibling: install via ComfyUI Manager (search "comfyui_qwen_translator_node"), or git clone https://github.com/hansol-icelab/comfyui_qwen_translator_node into custom_nodes, then pip install -r requirements.txt and restart. The real install is the model:

cd ComfyUI
hf download Qwen/Qwen3-8B --local-dir models/text_encoders/Qwen3-8B

Roughly 16GB, and it must land in models/text_encoders/ with a config.json inside or the loader won't list it. The example workflow in example_workflow/ is a clean starting point - just hit Load in ComfyUI.

Common issues

• "No models found" / empty dropdown on the loader - the model isn't where the loader scans. Get it into models/text_encoders/ first.
• Translation comes back cut off - you're hitting the max_new_tokens cap. Raise it (it tops out at 1024).
• The model is eating your VRAM while you generate. There's no unload node in the pack, so after the translator finishes, Qwen stays resident next to your checkpoint. On a 12GB card run the loader at 8bit or 4bit and treat a reload as the way to free memory.
• Output sounds generic. This is the honest ceiling of a local 8B. The richer example system prompt helps a lot; chaining it as a focused first step rather than asking for one mega-rewrite is the community-tested fix.

It's a niche tool - you have to write Korean prompts and be on a GPU that can host 8B. If that's you, this is the cleanest way to make it a native part of the graph.