comfyui_toriigate

ComfyUI custom nodes for Minthy/ToriiGate-0.5, an image captioning model for anime-style and digital art.

Original Model: Minthy/ToriiGate-0.5 GGUF Models: DraconicDragon/ToriiGate-0.5-GGUF

---

Installation

Clone this repository into your ComfyUI/custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/litch230/comfyui_toriigate.git

If you plan to use the local Transformers node instead of the API, install the requirements:

# Standard Python / venv
pip install -r ComfyUI/custom_nodes/comfyui_toriigate/requirements.txt

# ComfyUI Portable (Windows)
python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\comfyui_toriigate\requirements.txt

---

Usage Methods

This node pack provides two different ways to run ToriiGate.

1. Llama.cpp API (Recommended)

Connects to an external llama-server.exe running a GGUF version of the model.

• Node: ToriiGate Llama.cpp Vision Generate
• Model format: .gguf only
• Pros: Much faster inference, lower VRAM usage, bypasses python dependency issues.
• Cons: You need to run the llama-server manually before starting ComfyUI.

2. Transformers / PyTorch (Local)

Runs the model natively inside ComfyUI using the transformers library.

• Node: ToriiGate Captioner
• Model format: Original unquantized HuggingFace model format (cannot run GGUF).
• Pros: No external server needed.
• Cons: Slower, uses more VRAM.

---

Running the Llama.cpp Server

For the best performance on NVIDIA GPUs, follow these steps to run the API server:

1. Download the latest release from the llama.cpp releases page.

   • Choose the win-cuda zip matching your CUDA version (e.g. llama-b4109-bin-win-cuda-12.4-x64.zip).
   • If you have an AMD GPU, use win-vulkan or win-rocm.

2. Download the cudart zip file from the same release page (e.g. cudart-llama-bin-win-cuda-12.4-x64.zip).

   • Extract the cudart files directly into the same folder where you extracted llama-server.exe. Without this, inference on NVIDIA GPUs will be very slow.

3. Open CMD in that folder and run the server with the following optimizations. (Note: The -m and --mmproj flags are optional if you want to preload local files; if omitted, the ComfyUI node will automatically instruct the server to download and load the correct GGUF model):

llama-server.exe [-m <model.gguf>] [--mmproj <mmproj.gguf>] -b 2048 -ub 1024 -fa on -fit on -fitt 1024 -ngl 999

Command breakdown:

• -m: (Optional) Path to your downloaded .gguf language model.
• --mmproj: (Optional) Path to your downloaded .gguf vision projector model (required if preloading a vision model locally).
• -b 2048 -ub 1024: Batch sizes.
• -fa on: Enables Flash Attention.
• -fit on -fitt 1024: Speeds up image processing.
• -ngl 999: Offloads all layers to the GPU.

Once the server is running on http://127.0.0.1:8080, you can generate captions using the ToriiGate Llama.cpp Vision Generate node in ComfyUI.

---

Node Reference

• ToriiGate Grounding Builder: Compiles tags, characters, and descriptions into the final prompt string needed by the generator nodes.
• ToriiGate Captioner: Local PyTorch generator. Takes the image and the prompt string.
• ToriiGate Llama.cpp Vision Generate: API generator. Connects to llama-server and passes the image and prompt string.
• ToriiGate Llama.cpp Text Generate: Text-only node for chatting with the model.