🏖️Phi3mini 4k

Despite the name, this node doesn't make images. It writes text - specifically, it turns a rough idea into a polished English prompt you can hand to CLIPTextEncode. That's why it keeps appearing in workflows next to SDXL checkpoints: it's LLM-assisted prompting as a node, the practice that went from "open a browser tab with ChatGPT" to "drag a node into the graph" (the KB tracks that shift - dedicated local prompt-writing nodes are now a whole category). No API key, no cloud round-trip, no per-token bill. That's the entire appeal.

It wraps Microsoft's Phi-3-mini-4k-instruct, a 3.8-billion-parameter model under an MIT license. The author's launch pitch - "comparable to GPT-3.5 and Mixtral 8x7B" - is 2024 marketing; it'll give you solid prompt expansions, not a brilliant novelist. Treat it as a handy local helper, which is all you need it to be.

How it works

The node builds a chat-style message list - system instruction plus your prompt - and runs a transformers text-generation pipeline. Generation is capped at 500 new tokens and return_full_text is false, so you get back only the freshly written text, not a transcript. Chinese in, English prompt out (the author's headline feature) is just the default instruction doing its job.

The inputs that matter

Five inputs, and only a few you'll actually touch:

• model / tokenizer - from the 🏖️Phi3mini 4k ModelLoader node. No loader, no text.
• prompt (multiline) - your rough idea. "girl in a rainstorm, cinematic lighting" level of input is fine; the model does the embellishing.
• system_instruction (multiline) - the default is a working Stable Diffusion prompt-writer instruction. This is where you customize the job: captioning an image, translating, forcing a tag style for anime checkpoints, whatever.
• temperature (0–1, default 0) - and here's where people get burned. The node hardcodes do_sample=False, and greedy decoding ignores temperature entirely. The slider is decoration. Don't waste turns twiddling it.

The output and where it goes

One output: text (STRING). Wire it into any text display node - the built-in Show Text, or DisplayText_Zho if you have the author's ComfyUI-Gemini pack. In the bundled CosXL workflow it feeds straight into CLIPTextEncode, so the whole loop is: loader → this node → text encoder → sampler.

A note on which checkpoints to pair it with: it's an LLM writing natural language, and natural language is exactly what LLM-encoded models want. For SDXL-era checkpoints it adds polish and variety to your prompts; for models like Anima or Z-Image it's arguably a more natural fit than tag lists.

Installing and first run

Install with the pack. ComfyUI Manager → search "Phi-3-mini", or:

cd ComfyUI/custom_nodes
git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-Phi-3-mini
cd ComfyUI-Phi-3-mini
pip install -r requirements.txt   # transformers>=4.40.0

then restart ComfyUI. The one genuinely heavy step is the model: the loader auto-downloads it from Hugging Face on first run (several gigabytes, CUDA required - the loader hardcodes device_map="cuda"). The README's upgrade path if transformers is too old: pip uninstall -y transformers && pip install git+https://github.com/huggingface/transformers.

One more quirk: prompt is a text box on the node, not a wire input - you type into it. The whole point of the node is that you hand it a sentence and it hands back a better one; if you want to keep the conversation visible across runs, use the Chat variant instead.