MinusZone - deprecated - CLIPTextEncode(Phi3)

Another deprecated - prefix, another v1 prompt beautifier that the pack has moved past. CLIPTextEncode(Phi3) was added in April 2024 as the lightweight sibling to the Llama-3 node - Microsoft's Phi-3-mini (3.8B) instead of an 8B - and it's been superseded by the universal CLIPTextEncode(LLamaCPP Universal) plus model-config nodes, exactly like its Llama-3 cousin.

It does the pack's signature trick with one fixed model: type a rough idea into text, and a local Phi-3-mini GGUF (Phi-3-mini-4k-instruct-q4.gguf is the sole preset) expands it into a structured prompt, with optional clip wiring for conditioning out. The whole appeal was size: 3.8B Q4 is ~2.5GB and runs acceptably on CPU, which made it the "I don't have an 8B's worth of patience or VRAM" option of its day.

The inputs

llama_cpp_model (a single preset), download_source (none / modelscope / hf-mirror.com), prompt_version (v1, fixed), style_presets (none / high_quality / photography / illustration), text, keep_device, seed. Optional: clip, llama_cpp_options, customize_instruct. Outputs: text and conditioning. It's the Phi-3 mirror image of the Llama-3 node - same UI skeleton, different model.

Why it's deprecated and what to use instead

The hardcoded single-model design is the whole reason it's retired: the universal node + a LLamaCPPModelConfig will happily load the exact same Phi-3 GGUF (or any other model) with more controls. And honestly, the modern path is better here than it was for the Llama-3 node, because "small model for prompt beautification" is a real niche - a 0.5B Qwen or Phi-class model is plenty to expand a prompt sentence, and it's way faster than an 8B. The universal node lets you chase that; this node just points at one preset. If a legacy workflow imports with this node, it still runs - but don't add it to anything new.

Installing

It ships in the pack's v1 subpackage, always registered:

cd ComfyUI/custom_nodes
git clone https://github.com/MinusZoneAI/ComfyUI-Prompt-MZ

Restart. Requires llama-cpp-python + diskcache; first run pulls the ~2.5GB GGUF into ComfyUI/models/gguf from your chosen download_source.

Troubleshooting

Same trio of llama-cpp-python issues as every local-LLM node in this pack, all documented in the README FAQ: missing LLAMA_SPLIT_MODE_LAYER → upgrade llama-cpp-python; LLama.dll failing to load → switch PyTorch to CUDA 12.1; a pip wheel build returning nonzero exit → network problem, use a proxy or install llama_cpp_python manually. One Phi-3-specific expectation to set: at 3.8B Q4 it's a weaker reasoner than the 8B models, so prompts come out more generic, and the default style_presets quality vocabulary (masterpiece, best quality) is SD 1.5-era - inert on LLM-encoder checkpoints, so none is your friend there. For a faster modern equivalent, skip this node and run the universal encode node with a small Qwen GGUF from the downloader.