ComfyUI-RT-LTX2-RareTools

ComfyUI-RT-RareTools

Advanced custom nodes suite for ComfyUI featuring LTX-Video 2 / LTX-2.5, Qwen Image 2.1, MiniMax H3, and real-time Interactive Visual A/B Image and Video Comparison.

Tutorials and walkthroughs: YouTube @raretutor

---

Key Features

• RT Image Compare: Real-time canvas-based A/B image comparison with zero-drift slide wipe, smooth mouse wheel zoom (1x to 30x centered on cursor), pan support, click toggle, GPU difference blend, and fullscreen lightbox inspection.
• RT Video Compare: Synchronized dual-video comparator with wipe slider, interactive timeline scrubber, frame-by-frame stepping, and playback speed controls.
• Prompt Enhancers (Qwen Image 2.1 & LTX-2.5): High-performance local vision/LLM prompt enhancement powered by llama-cpp-python (v0.4.1) with hardware CUDA / Metal acceleration.
• LTX2 Video Optimization: Tiled VAE encoder/decoder, streaming decoder, STG guidance, self-refining patch, long video scheduler, and video-only LoRA loader.
• MiniMax H3 Suite: Advanced sigma scheduling, AutoSigma, split sigmas, and universal LoRA tooling.

---

Installation Guide

Step 1: Clone the Repository

Open your terminal or command prompt inside your ComfyUI custom_nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/monnky/ComfyUI-RT-RareTools.git
cd ComfyUI-RT-RareTools

---

Step 2: Install Dependencies by Operating System

This node suite utilizes pre-compiled llama-cpp-python (v0.4.1) wheels with CUDA / Metal acceleration for fast local GGUF model execution, alongside standard media processing libraries (imageio, imageio-ffmpeg, json-repair).

Windows

Method A: One-Click Installer (Recommended for ComfyUI Portable)

If you are using the official ComfyUI Windows Portable package:

1. Navigate to the ComfyUI/custom_nodes/ComfyUI-RT-RareTools folder.
2. Double-click Llama-CPP-Python_Windows(No_Deps).bat.
   • It automatically locates your python_embeded executable.
   • It installs required packages from requirements.txt.
   • It detects your GPU CUDA version and downloads the matching pre-compiled v0.4.1 wheel without modifying your existing PyTorch installation (--no-deps).

Method B: Manual Installation (Command Prompt / PowerShell)

Open a terminal in the repository directory:

:: For ComfyUI Portable:
..\..\..\python_embeded\python.exe install.py

:: For standard Python / venv:
python install.py

To force-reinstall or upgrade llama-cpp-python to v0.4.1:

..\..\..\python_embeded\python.exe install.py --force

Pre-compiled Windows CUDA wheels are hosted on GitHub by JamePeng:

• CUDA 13.1: v0.4.1-cu131-win-20260926
• CUDA 13.0: v0.4.1-cu130-win-20260926
• CUDA 12.8: v0.4.1-cu128-win-20260926
• CUDA 12.6: v0.4.1-cu126-win-20260926
• CUDA 12.4: v0.4.1-cu124-win-20260926

---

Linux

Method A: Auto-Installer Script

Open a terminal in the repository directory:

chmod +x install.sh
./install.sh

Method B: Manual Virtual Environment Setup

Activate your ComfyUI virtual environment or Conda environment, then run:

# Activate your environment
source /path/to/comfyui_env/bin/activate
# Or: conda activate comfyui

# Install dependencies and resolve matching CUDA wheel
python install.py

To force-reinstall or upgrade:

python install.py --force

Supported Linux CUDA builds:

• CUDA 13.1: v0.4.1-cu131-linux-20260926
• CUDA 12.8: v0.4.1-cu128-linux-20260926
• CUDA 12.6: v0.4.1-cu126-linux-20260926
• CUDA 12.4: v0.4.1-cu124-linux-20260926

---

macOS (Apple Silicon M-Series & Intel)

Method A: Auto-Installer Script

Open Terminal in the repository directory:

chmod +x install.sh
./install.sh

Method B: Manual Setup

# Activate your ComfyUI virtualenv
source /path/to/comfyui_env/bin/activate

# Run installer
python install.py

On Apple Silicon (M1, M2, M3, M4), the installer automatically downloads the pre-compiled Metal-accelerated build (v0.4.1-Metal-macos-20260926), allowing GGUF models to run on Apple GPU / Unified Memory.

---

Dependencies Summary

All base dependencies are listed in requirements.txt:

imageio
imageio-ffmpeg
json-repair

llama-cpp-python (v0.4.1) is installed with --no-deps to preserve your environment's PyTorch, TorchVision, and NumPy versions.

---

Recommended GGUF Models for Prompt Enhancers

Place your downloaded .gguf model files into:

• LLM Models: ComfyUI/models/LLM/ (or ComfyUI/models/prompt_generator/)
• Vision Projector Models: ComfyUI/models/mmproj/

Suggested Models:

1. Gemma 4 / Gemma 3:
   • Gemma-4-E4B-it-GGUF (Unsloth)
   • Vision projector: mmproj-BF16.gguf
   • Gemma-3-12B-it-GGUF (Unsloth)
2. Qwen 2.5 / Qwen 2.5-VL:
   • Qwen2.5-VL-7B-Instruct-GGUF / Qwen2.5-VL-3B-Instruct-GGUF
   • Matching vision projector: mmproj-*.gguf

---

Troubleshooting & FAQ

• RuntimeError: CUDA version mismatch: Run python install.py --force (or Llama-CPP-Python_Windows(No_Deps).bat) to let the installer re-query your active PyTorch CUDA version and download the corresponding pre-compiled wheel.
• Why is --no-deps used for llama-cpp-python?: llama-cpp-python wheels from PyPI can pull in conflicting NumPy or standard packages. Using --no-deps guarantees that your ComfyUI environment remains intact and stable.
• Node visual widgets not showing up: Ensure you refresh your browser (Ctrl+F5) after restarting ComfyUI so the new frontend extensions in web/ are loaded by LiteGraph.

---