MinusZone - deprecated - CLIPTextEncode(LLama3)

The display name says it plainly: deprecated - CLIPTextEncode(LLama3). This was the pack's original Llama-3-specific prompt beautifier, and the author has since replaced it with the universal CLIPTextEncode(LLamaCPP Universal) + model-config nodes. It still works, but it's the v1 of an idea - understand it for what it teaches, then use the newer node.

What it does is the same core trick as the flagship: you type a rough idea into text, a local Llama-3 GGUF (default Meta-Llama-3-8B-Instruct.Q4_K_M.gguf) expands it into a structured SD prompt, and with a clip wired in you get conditioning out. The difference from the modern node is that the model is baked into a preset dropdown - 14 quantizations of Meta-Llama-3-8B-Instruct from Q2_K to Q5_K_M - plus a download_source selector (none / modelscope / hf-mirror.com) so the pack can fetch whichever quant you pick. That's it. The model is not swappable to Qwen or Phi; you get Llama-3 variants or nothing.

The inputs

Same shape as the flagship: llama_cpp_model (the quant dropdown, [downloaded] marks what you already have), download_source, 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.

Why it got deprecated

The design constraint - one hardcoded model family - is the whole story. Once the author added ModelConfigManualSelect/DownloaderSelect and the universal encode node, any model could be used, and this node became redundant. The useful bits (style presets, A1111-style encoding) all live on in the modern node. There's no feature here you can't get with the universal node plus a LLamaCPPModelConfig - and the universal node gives you translate, format, and text_raw outputs on top. If you're starting a workflow today, start there.

Installing

Part of ComfyUI-Prompt-MZ (it ships in the v1 subpackage, always loaded):

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

Restart. Depends on llama-cpp-python + diskcache; first run downloads the chosen GGUF into ComfyUI/models/gguf.

Troubleshooting

Because it's deprecated, it's also the place you'll hit the pack's older bugs first. The README FAQ is all llama-cpp-python friction: LLAMA_SPLIT_MODE_LAYER missing → upgrade llama-cpp-python; LLama.dll won't load → CUDA 12.1 PyTorch; pip wheel build failing → network, proxy up or install the package manually. One more honest note: an 8B Q4 model with the default temperature-heavy settings can ramble, and the style_presets quality text (masterpiece, best quality) is SD 1.5-era vocabulary - largely inert on newer LLM-encoder checkpoints, so lean on none + your own customize_instruct if that's your stack. If a preset model name shows [downloaded], the pack found it in ComfyUI/models/gguf and won't re-fetch - a file there with a matching name is picked up automatically.