TranslateGemma

TranslateGemma is Google's open-weight translation family, built on Gemma 3, and this node drops it straight into your ComfyUI graph. Fifty-five languages, no API key, no monthly bill - the whole thing runs on your own GPU. The name is a lie in the best way: it doesn't call any API, and "translate" here covers text and the text you can point at inside an image.

The reason most people reach for it is simpler than the multimodal stuff: prompt translation. If English isn't your first language, you've got two bad options - write weaker English prompts, or tab out to Google Translate every other workflow. This node lets you write the prompt in your native language, pipe it through TranslateGemma, and hand clean English to the text encoder. That's the exact use case the community post about this pack led with: "tired of switching tabs to DeepL or Google Translate just to build your prompts."

How it works

Under the hood it's the official google/translategemma-* repos loaded through transformers (AutoModelForImageTextToText, so it's the same multimodal path whether you feed it a string or an image). First run downloads the model snapshot from Hugging Face into ComfyUI/models/LLM/TranslateGemma/<repo>/ - which is where the friction starts, because these repos are gated. The README walks you through accepting the Gemma terms and running hf auth login, or setting HF_TOKEN, then restarting.

You pick a size with model_size: 4B (the default), 12B, or 27B. This is the trade-off dial - speed and VRAM against translation quality. For source_language it does a best-effort local detect (via langid, which is why langid is in requirements.txt), and the node also normalizes Chinese Simplified/Traditional variants. There's even a chinese_conversion_only mode that converts between Simplified and Traditional script via OpenCC without loading a model at all - handy, deterministic, and free of VRAM cost.

The inputs that actually matter

You can ignore most of the 20-ish widgets on first run. The ones worth touching:

• text / external_text - the string to translate. Connect a prompt from an upstream node to external_text and it overrides the built-in box.
• target_language - where you're going (default English).
• source_language - leave on Auto Detect for text, but images require an explicit source. No OCR pre-pass, so Auto Detect doesn't work there.
• model_size - start with 4B; bump up only if the quality bugs you.
• image - the fun one. Feed an IMAGE and it reads/translates on-screen text, resizing to the model's preferred 896×896 (the letterbox mode is the sensible default).

The single output is translated_text, a STRING - wire it straight into a CLIP text encode or any node that eats strings. That's the whole pipeline: native language in, English prompt out, no browser involved.

Installing it

Easiest via ComfyUI Manager - search TranslateGemma and install. Manual works too:

cd custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-TranslateGemma.git
cd ComfyUI-TranslateGemma
pip install -r requirements.txt

Then restart, accept the Gemma license on Hugging Face, and authenticate (hf auth login or a HF_TOKEN env var) or the download 401s. ComfyUI Desktop/portable users must run that pip install with ComfyUI's own Python environment, not a global one. Quantization (bnb-8bit / bnb-4bit) is a separate optional install and CUDA-only - skip it unless you're chasing the 12B/27B on a consumer card.

Where people get burned

• The gated download. 401/403/gated/forbidden means you haven't accepted the terms or set a token. Not a node bug. Downloads are resumable, and if you're behind a proxy or in a region where Hugging Face is flaky, you can manually copy a complete snapshot into models/LLM/TranslateGemma/<repo>/.
• VRAM reality check. Native, that's ~12 GB for the 4B, ~27 GB for 12B, ~56 GB for 27B. A 4B model eating 12 GB still shocks people - plan around it, or lean on bitsandbytes quantization.
• Images need an explicit source language. Forget this and you'll get garbage or an error, not a surprise.
• max_input_tokens too low can truncate the model's visual tokens on image input and produce unrelated output - the node enforces a safe floor, but don't go tweaking it down for fun.
• Auto Detect can guess Chinese script wrong on short ambiguous text. When it matters, pick the language explicitly.

It also runs remote code with trust_remote_code=False by default and only falls back when required, with an allowlist env var - nice to see after the whole custom-node supply-chain scare, since an LLM node like this is exactly the shape of thing that needs that care. Worth a look if you're a non-English speaker who's been living in browser tabs; it's the one translation node I'd actually reach for.