Prompt textarea kept in sync with tag chips (comma-separated)
Tag chips with hover popup: adjust weight and delete without retyping
Double-click a chip to edit its text
Press-and-drag to reorder tags (dragging starts immediately on movement)
Single click a chip to disable / re-enable it — disabled tags stay in the list (grayed out) but are excluded from the prompt and outputs
Translated tag chips with per-language cache; untranslated chips shown dashed
Auto-translate toggle: missing translations are filled automatically after edits
Chinese (CSV) language option (default): translates tags by looking them up in the dataset CSVs — instant, no model, and switching to it unloads the translation model; tags with no dataset entry show red "untranslatable"
Underscore → space toggle: when on, every _ in the prompt (typed, pasted or inserted from suggestions) becomes a space
8 model-backed target languages: Chinese / Japanese / Korean / English / French / German / Spanish / Russian (loaded on demand, as before)
Lazy-loaded Hy-MT2-1.8B-Q4_K_M GGUF, full GPU offload (n_gpu_layers=-1), batched request with per-item fallback and translation cache

Outputs

| Output | Content | |---|---| | Prompt | Serialized prompt with weights, e.g. (white dress:1.2), long hair | | Translated Prompt | Comma-separated translations | | Tags JSON | Full tag list (text / translation / weight / disabled) as JSON |

Install

Via ComfyUI-Manager (recommended)

Custom Nodes Manager → search KKMLMF or Prompt Tag Editor → Install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/kkmlimingfeng/ComfyUI-KKMLMF-Nodes

Restart ComfyUI.

Translation model

The node works without the model; only the translate button needs it.

On first use, Hy-MT2-1.8B-Q4_K_M.gguf (~1.1 GB) is downloaded automatically to ComfyUI/models/text_translation/Hy-MT2-1.8B-GGUF/ — from Hugging Face, with hf-mirror.com as a fallback. The editor status line shows download progress while it runs.

Manual placement also works — put the GGUF here (any *.gguf in that folder is accepted, Q4_K_M preferred):

ComfyUI/
  models/
    text_translation/
      Hy-MT2-1.8B-GGUF/
        Hy-MT2-1.8B-Q4_K_M.gguf

Hugging Face: https://huggingface.co/tencent/Hy-MT2-1.8B-GGUF
China mirror: https://hf-mirror.com/tencent/Hy-MT2-1.8B-GGUF

requirements.txt intentionally lists no hard dependency: llama-cpp-python is imported lazily so the editor works even without it.

GPU offload needs a CUDA-enabled llama-cpp-python build. Recommended: grab a prebuilt wheel matching your CUDA toolkit (cu124 / cu128 / ...) and Python version (cp310 / cp311 / cp312 / ...) from https://github.com/JamePeng/llama-cpp-python/releases, then:

pip install llama_cpp_python-<version>+cu128-cp312-cp312-win_amd64.whl

Alternatives:

# CUDA build via the official wheel index
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu128

# CPU-only build (works everywhere, slower)
pip install llama-cpp-python

Usage

Add the node from KKMLMF-Nodes → Prompt Tag Editor (or type prompt directly).
Edit tags as chips; type in the textarea to add tags.
Hover a chip → popup with title / weight / delete.
Toggle 自动 (auto) to translate automatically after edits, or press the ↻ button.
Connect the outputs to downstream nodes.

License

MIT