ComfyUI Extension: ComfyUI-Qwen3.5-Uncensored-GGUF

ComfyUI-Qwen3.5-Uncensored-GGUF

GGUF-only build of ComfyUI-Qwen3.5-Uncensored. Uses llama-cpp-python instead of transformers — no transformers dependency, so it won't conflict with older ComfyUI nodes that need transformers <5.

Requirements

• Python: 3.10, 3.11, 3.12, 3.13, or 3.14
• CUDA: 12.4, 12.6, 12.8, or 13.0
• llama-cpp-python: Must be installed from JamePeng's fork (not pip)

Tested with Python 3.13 + CUDA 13.0 + PyTorch 2.10/2.11.

If your Python or CUDA version isn't listed above, the required llama-cpp-python wheels are not available and this node pack won't work.

Why this version?

transformers>=5.2.0 is required by the full version for HuggingFace model support, but it breaks some older ComfyUI custom nodes. This GGUF-only build removes that dependency entirely while keeping all GGUF functionality.

Nodes included

| Node | Description | |------|-------------| | Qwen3.5-Uncensored (GGUF) | Vision-language inference with GGUF models (image + video) | | Qwen3.5-Uncensored Advanced (GGUF) | Same with full parameter control (ctx, batch, gpu layers, etc.) | | Qwen3.5-Uncensored Prompt Enhancer (GGUF) | Text-only prompt enhancement/rewriting | | Story Split Node | Splits WAN 2.2 Story output into 4 separate prompts | | VRAM Cleanup | VRAM cleanup with multiple modes |

Automatic Local Model Discovery

You don't need to edit JSON config files. The nodes automatically scan your local model directories:

• GGUF models: Any .gguf file found under models/LLM/GGUF (or any extra model path you've configured) shows up in the dropdown with a [local] prefix
• mmproj files in the same directory are automatically paired with the model
• Respects ComfyUI's --base-directory, extra_model_paths.yaml, and all registered LLM paths
• Model architecture (Qwen3.5 vs others) is detected from GGUF file metadata, not the model name
• Pre-configured models from gguf_models.json still work with auto-download

REMEMBER: VL models need an mmproj file in the same folder to show up in the VL model list.

---

Installation

cd ComfyUI/custom_nodes/

git clone https://github.com/Deaquay/ComfyUI-Qwen3.5-Uncensored-GGUF.git

cd ComfyUI-Qwen3.5-Uncensored-GGUF

pip install -r requirements.txt

Install llama-cpp-python

You must install from JamePeng's fork, not pip.

Pick the wheel matching your CUDA + Python version.

Example for Python 3.13 + CUDA 13.0 on Linux:

pip install https://github.com/JamePeng/llama-cpp-python/releases/download/v0.3.34-cu130-Basic-linux-20260331/llama_cpp_python-0.3.34+cu130.basic-cp313-cp313-linux_x86_64.whl

Available for CUDA 12.4, 12.6, 12.8, 13.0 and Python 3.10–3.14.

For Windows users with Python 3.13 + CUDA 13.0, use:

pip install https://github.com/JamePeng/llama-cpp-python/releases/download/v0.3.34-cu130-Basic-win-20260331/llama_cpp_python-0.3.34+cu130.basic-cp313-cp313-win_amd64.whl

But better look into the releases of the JamePeng repo for your version.

Adding models

• Drop GGUF files into models/LLM/GGUF/ — they'll be auto-detected
• For VL models, put the mmproj file in the same directory as the model file
• Or add entries to gguf_models.json for auto-download from HuggingFace

What's NOT included (compared to the full version)

• No HuggingFace/transformers model support
• No Flash Attention 2 / SageAttention acceleration
• No BitsAndBytes quantization
• No HF Prompt Enhancer node

If you need those features, use the full version instead.

License

GPL-3.0 — see LICENSE