🌐 Ollama Cloud Translator

The main event in the Comfy_olima pack: a node that takes Korean (or Japanese, Chinese, English) text and turns it into a prompt-ready English string, using an LLM through Ollama's OpenAI-compatible API. Its author is Korean and the README is Korean-first, which tells you exactly who this was built for - people who think about the image in their native language and can't be bothered to hand-write comma-separated English tags. Default style is sd_prompt, and that's the whole pitch: not a plain translator, a prompt engineer.

This is the "LLM-in-the-graph" pattern from the ComfyUI ecosystem, pointed at text instead of the sampler. You type "겨울 숲속의 여성, 은은한 햇살" and get back something like a woman in a winter forest, soft sunlight, cinematic lighting, detailed, masterpiece. Wire that into a CLIP Text Encode node and it behaves exactly like a prompt you wrote by hand - because it is one.

How it works

It's a thin HTTP client, not a model loader. It POSTs to {base_url}/chat/completions with a system prompt built from your style choice plus your text as the user message. The endpoint defaults to https://ollama.com/v1 (Ollama Cloud, needs an API key from ollama.com/settings/keys), or point it at http://localhost:11434/v1 for your own Ollama server, where no key is needed. The code even auto-fixes sloppy URLs - strips a trailing /chat/completions you pasted in, appends /v1 if missing - which is a small kindness you'll appreciate.

Because the URL is a free text field, the same node works against any OpenAI-compatible chat endpoint, not just Ollama. And it retries sensibly: transient failures (429, timeout) get retried up to max_retries times, while a 4xx auth error breaks immediately so you don't hammer a bad key.

The inputs that matter

Most of the fields you can safely leave alone. These are the ones worth touching:

• text - what you want translated, multiline.
• model - dropdown of Ollama Cloud models (deepseek-v4-flash:cloud is the default). Note this list is hardcoded in the pack; if your model isn't there, use the custom_model field instead.
• target_language - English / Japanese / Chinese / Korean.
• translation_style - the five styles, and this is where the node earns its keep:
  • sd_prompt (default, recommended): outputs a comma-separated English SD/FLUX prompt with quality and style keywords. This is the one you want 90% of the time.
  • preserve_tags: translates the prose but leaves booru tags, (word:1.2) weights, <lora:name:1.0> and artist refs untouched - for when you're editing an existing prompt instead of starting from scratch.
  • natural, literal, concise: plain translations of decreasing warmth.
• api_key - required in cloud mode.

Optional but handy: temperature (0.3 default - good for translation; crank it if you want the LLM to get creative with tag suggestions), max_tokens (512 default), seed (bumping it forces a re-run and dodges ComfyUI's cache), and custom_model which overrides the dropdown entirely.

Outputs

Two strings: translated_text - the good stuff, wire it into the text input of a CLIP Text Encode node - and original_text, which just echoes your input, useful for comparison or if you want to chain the translation into another node while keeping the source around.

Install & gotchas

Install via ComfyUI Manager (search Comfy_olima) or:

cd ComfyUI/custom_nodes
git clone https://github.com/utahime12/Comfy_olima.git

Then restart ComfyUI. The pack's only dependency is requests, which you already have - no model downloads, no heavy installs.

Two things that bite people:

• In cloud mode the API key is mandatory. Leave it blank and the node won't crash - it returns an [Error] ... string as its output, which then flows straight into your CLIP encoder if you're not looking. Any output starting with [Error] is a message, not a translation; read it.
• The dropdown is a snapshot. Model names drift. If you get a 404-ish error for a model, run the pack's Ollama Cloud Model List node to see what your server actually has, then put that name in custom_model.

One honest caveat, same as every API LLM node: your text and key leave the machine and go to ollama.com, and cloud APIs have their own content filtering. If that matters to you, flip base_url to local and run the same node for free on your own Ollama - that's the mode a lot of people end up on. The translator stays identical; only the server changes.