ThinkingLLM Prompt Enhancer

Prompt enhancement is one of those workflows that felt like a gimmick in the SD 1.5 era and became genuinely useful once 2026's models started using LLM text encoders. Those models want natural-language descriptions, not keyword soup - and a local LLM is exactly what expands "girl by a window, rain" into a full scene description the encoder can chew on. This node is ThinkingLLM's HF version of that: you feed it a rough prompt_text, it returns a cleaned, expanded ENHANCED_OUTPUT string you wire straight into your CLIP/text-encode stage.

It's the same Transformers path as the rest of the pack, so install friction is low compared to the GGUF nodes. The model catalog is the full Qwen/Qwen-VL/Gemma list, though you're not using the vision here - this node takes text in and gives text out.

What to set

• enhancement_style - the real decision. 📝 Enhance is the general-purpose default, and there are Refine, Creative Rewrite, and Detailed Visual styles plus a pile of LTX 2.3 and Wan 2.2 scene/timeline templates that are clearly aimed at video-prompt builders. ✍️ Custom Only hands full control to custom_system_prompt.
• prompt_text - your rough prompt. Leave it blank and the selected preset instruction becomes the base prompt.
• custom_system_prompt - required with Custom Only; otherwise it's prepended to the selected style.
• quantization - None (FP16) for best quality if memory allows, 8-bit for 8–16 GB GPUs, 4-bit for 6 GB or less (slower). FP16 is the default and right for most people.
• temperature / top_p / repetition_penalty - 0.7 / 0.9 / 1.1 defaults are fine; nudge temperature up for wilder rewrites, and repetition_penalty above 1.0 already helps keep long prompts from looping.
• max_tokens - 1024 default; raise it only when the model truncates useful detail.

Also present, in case you've read the rest of this pack: attention_mode, use_torch_compile, device, keep_model_loaded, seed, keep_last_prompt, stream_tokens_to_terminal, enable_thinking, and hf_token for gated models. Outputs are ENHANCED_OUTPUT and RAW_TRACE - the cleaned prompt on the left, the raw stream (including hidden reasoning) on the right.

Installing

ComfyUI Manager → search ThinkingLLM → install, or:

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

Requires transformers>=5.2.0; restart after install. First run downloads the chosen weights into models/LLM/Qwen-VL.

Gotchas

The fixed-seed prompt cache applies here too: identical inputs reuse the saved enhanced prompt, and enabling stream_tokens_to_terminal deliberately bypasses that cache for a fresh run. Keep in mind this node's purpose is text, so don't expect it to look at your image - for image-grounded enhancement you'd describe the image first with a vision node, then enhance. And if you're only ever enhancing prompts, the GGUF version of this node uses a small text-only quant that's a lot lighter than pulling a vision model for the job.