ThinkingLLM Prompt Enhancer (GGUF)

The HF Prompt Enhancer is convenient but it drags a vision model in to do a text job. This GGUF version is the lean alternative: a text-only Qwen3 quant that exists to expand and refine prompts, nothing else. The default is a Q4_K_M Josiefied Qwen3-4B at ~2.4 GB - about a third of the download of the base node's default VL model - so this is the one to drop into a video-prompt loop where you're enhancing a dozen prompts in a row and don't want to load a 7 GB vision model for any of them.

No image input, no audio input, no vision handlers needed. It's text in, ENHANCED_OUTPUT out, with RAW_TRACE alongside. This is also the node where the pack's uncensored-model catalog shows its hand: the defaults are abliterated and "gabliterated" Qwen3 quants, and the style presets include the LTX 2.3 and Wan 2.2 NSFW scene/timeline templates. If that's not your lane, pick a base Qwen3 quant and a Refine style and it behaves like a polite copy editor.

What to set

• preset_system_prompt - the style list: 📝 Enhance (default), Refine, Creative Rewrite, Detailed Visual, the LTX/Wan video templates, and ✍️ Custom Only for full control via custom_system_prompt.
• prompt_text - the rough prompt to enhance; blank means the selected preset instruction becomes the base.
• custom_system_prompt - required with Custom Only, otherwise prepended to the style.
• english_output - a small but handy toggle: ask the model to return the final prompt in English even if the source prompt isn't.
• temperature / top_p / repetition_penalty - 0.7 / 0.9 / 1.1 defaults; the repetition penalty above 1.0 matters more here than on a single-shot answer, because long generated prompts love to loop.
• max_tokens - 1024 default; bump it when good detail gets cut off.

Rounding out the list: model_name, device (auto/cuda/cpu/mps - auto prefers GPU), keep_model_loaded, seed, keep_last_prompt, stream_tokens_to_terminal, enable_thinking, auto_finalization_retry, and hf_token. The GGUF list marks [local] files it auto-discovers in models/LLM/GGUF/, skipping anything named mmproj.

Installing

Manager (search ThinkingLLM) or:

cd ComfyUI/custom_nodes
git clone https://github.com/goodguy1963/ComfyUI-ThinkingLLM.git
cd ComfyUI-ThinkingLLM
pip install -r requirements.txt

Because this is a GGUF node it still needs a working llama-cpp-python, though not the vision-flavored build - the backend check here runs without requiring the multimodal handlers. Linux auto-installs a matching wheel; Windows users should still verify their llama.cpp build isn't silently CPU-only, since a text-only model on CPU is workable but slow.

Gotchas

The fixed-seed cache returns a saved prompt for identical inputs, and enabling terminal streaming forces a fresh run - that's by design, so you can tell whether you're seeing a cached result or a new one. The catalog defaults are abliterated quants, which is fine and worth knowing, but don't assume every model in the list behaves identically on the thinking toggle - it's advisory on non-Qwen GGUFs. And for the love of clean workflows: the hf_token field is real, and it does not get written to RAW_TRACE or cache keys, but it can be serialized into your saved workflow file - clear it before you share.