Gemma4 Translate

If you write prompts in anything but English, you already know the dance: compose in your language, alt-tab to DeepL or Google Translate, paste the result into a CLIP Text Encode, repeat every time you tweak a word. Gemma4 Translate removes the tab-switching. It's a fully local translation node that runs Google's gemma-4-E2B-it model right inside ComfyUI - no API, no key, no data leaving your machine. The name is a lie only in the sense that there's no backend to pay for; it's pure transformers on your own GPU.

It's a fork of the popular ComfyUI-TranslateGemma, reworked for Gemma 4 E2B and for people who want to run it offline. The headline feature most people miss: it's multimodal. The E2B variant reads images too, so you can feed it a screenshot, a poster, or a manga page and get the visible text back - translated or not.

How it works

The model is fixed: google/gemma-4-E2B-it, an image-to-text vision-language model loaded through AutoModelForImageTextToText plus AutoProcessor. The node is genuinely offline-capable - there's a resolution order for where the weights come from:

1. local_model_dir (you can type a path on the node, which gets saved into the workflow - a real improvement over the parent pack's env-var-only approach)
2. the GEMMA4_E2B_LOCAL_DIR environment variable
3. your Hugging Face cache
4. a download from the Hub, dumped into local_model_dir

If offline_mode is on and nothing local exists, it just errors instead of phoning home.

The other knob that matters is load_strategy. Default is reload_every_run, which unloads the model and clears CUDA cache after every execution - the VRAM-conscious choice. Flip to keep_loaded and it stays resident between runs, which makes batch translating a lot faster at the cost of VRAM.

The inputs that matter

• source_language / target_language - pick from ~35 languages, with Auto Detect as the source default. Target defaults to English.
• input_text - the text to translate. If you leave it empty, you need an image.
• task_mode - this is the interesting one. translate for plain text, ocr_translate to read text off an image and translate it, ocr_only to just extract text, and translate_and_explain which adds a short note on ambiguous bits.
• soft_token_budget - a budget for how much of the context window the image gets. Crank it up (560/1120) when the image has small or dense text you need OCR'd; 70 is fine for a quick caption.
• quantization - none, bnb-8bit, or bnb-4bit. The 4-bit option is how you squeeze the model onto a mid-range card.

Both outputs are plain strings: translated_text (wire it into a Show Text node, or into any prompt input) and info (a diagnostic string showing the repo, load path, and settings used - handy when you're debugging why a run was slow).

Installing it

ComfyUI Manager is the easy route - search for "Gemma4Translate" or the pack title comfyui-gemma4translate-e2b. Or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/jkmq-f/ComfyUI-Gemma4Translate-E2B
cd ComfyUI-Gemma4Translate-E2B
pip install -r requirements.txt

Then restart ComfyUI. Heads-up: the dependencies are chunky and pinned. It wants transformers==5.5.0, accelerate, sentencepiece, langcodes, and bitsandbytes (Linux + CUDA only - 8-bit/4-bit quant simply won't work on Windows). The README deliberately does not touch your Torch install, expecting the version ComfyUI bundles (2.5.x-ish). Don't go "fixing" that.

Where people get burned

• The first run downloads the model. It's a multi-GB fetch from Hugging Face, so don't panic when the queue sits there. If you're on a metered connection or repeating a run, set offline_mode + local_model_dir to skip the HTTP check every time.
• CUDA memory flakiness. The README suggests launching ComfyUI with python main.py --disable-cuda-malloc if you hit instability around memory allocation. It's the first thing to try before blaming the node.
• It re-runs every queue. The class sets IS_CHANGED to nan, so ComfyUI treats it as always dirty - every queue run re-translates, and with reload_every_run that means a reload per run. keep_loaded is your friend for batch jobs.
• Transformers version fights. If another custom node drags transformers elsewhere, the README's fix is pip install --upgrade --force-reinstall "transformers==5.5.0".

Honest take: if you only translate a prompt occasionally, alt-tabbing to Google Translate is still fine and this is overkill. This node earns its keep when you translate a lot, want it private and offline, or need in-workflow OCR-translation of images - that last one is genuinely hard to do any other way without leaving ComfyUI.