torch numpy safetensors>=0.4.0 gguf packaging transformers diffusers huggingface_hub accelerate optimum sentencepiece tqdm pyyaml torchaudio lycoris
Advanced raw patching toolkit for ComfyUI.
Install via ComfyUI Manager or clone manually:
cd ComfyUI/custom_nodes
git clone https://github.com/TripleHeadedMonkey/ComfyUI-Zlycoris.git
Restart ComfyUI.
Zlycoris – ComfyUI Nodes for Transformer Merging & LoRA Automation
Zlycoris is a collection of ComfyUI nodes focused on two main areas:
Transformer model merging (including Qwen 3 4B support)
Advanced LoRA loading and prompt-driven automation utilities
This extension is designed to give you structured control over model merging and dynamic LoRA behavior inside ComfyUI workflows.
Features
🧠 Transformer Model Merging
Includes nodes for merging full transformer models using structured weight merging techniques.
Highlights
TIES-based transformer merging
Qwen 3 4B-specific merging nodes
GGUF raw loading support
Dequantization utilities
Advanced weight operations
Use Case
Blend multiple Qwen models into a hybrid model
Combine instruction-tuned and creative variants
Experiment with task arithmetic on transformer checkpoints
Create custom merged models inside ComfyUI
These nodes operate at the full model weight level — not LoRA merging.
🎛 Universal LoRA / LyCORIS Loader
Supports multiple LoRA-style formats through a unified loader.
Supported Types
LoRA
LyCORIS variants
LoHa
LoKr
DoRA (if applicable)
Use Case
Load LoRAs directly into your diffusion pipeline
Adjust strength dynamically
Stack multiple adapters
Build structured LoRA systems
🔍 Keyword Matching Gate
A utility node that detects keywords in prompts and outputs boolean signals.
Use Case
Enable specific LoRAs only when keywords appear
Automatically activate lighting, style, or character adapters
Build smart prompt-reactive workflows
Example:
If prompt contains “cinematic” → enable lighting LoRA
If prompt contains “anime” → disable photoreal LoRA
🔄 Primitive to String Utilities
Converts numeric or widget values into strings for dynamic parameter control.
Use Case
Dynamically control LoRA strength
Build string-based automation systems
Drive behavior from sliders or external values
Useful when combining user inputs with prompt-aware logic.
⚙️ Conditional & Advanced Conditioning Nodes
Includes tools for:
Conditional routing
Conditioning scaling (e.g., CondMul)
Advanced conditioning manipulation (ZCondAdv)
Use Case
Multiply or scale conditioning signals
Create branch-based LoRA stacking systems
Build structured style control systems
📦 GGUF & Dequant Utilities
Utilities for:
Loading raw GGUF transformer models
Dequantizing weights
Preparing models for merging
Use Case
Work with quantized Qwen models
Experiment with merging lower-memory checkpoints
Prototype transformer blends inside ComfyUI
Example Workflow Concepts
These nodes are modular and designed to work together. Here are simple examples of how they might be used:
1️⃣ Prompt-Driven LoRA Activation
Feed prompt into Keyword Match Gate
Use output to enable/disable specific LoRAs
Adjust strength with slider or numeric input
Result: One workflow that automatically adapts to prompt content.
2️⃣ Dynamic LoRA Strength Control
Use slider → Primitive to String
Combine with keyword detection
Adjust LoRA intensity based on words like:
“slightly”
“very”
“extremely”
Result: Prompt-aware LoRA strength scaling.
3️⃣ Transformer Personality Mixer (Qwen)
Load multiple Qwen models
Merge with TIES node
Adjust merge ratios with sliders
Result: Custom blended LLM personality or task behavior.
4️⃣ Structured LoRA Stacking System
Create separate branches:
Character LoRAs
Lighting LoRAs
Texture LoRAs
Style LoRAs
Control each branch with:
Keyword gates
Sliders
Conditional routing
Result: Clean, modular LoRA architecture instead of uncontrolled stacking.
Design Philosophy
Zlycoris focuses on:
Structured model merging
Modular LoRA control
Prompt-aware automation
Workflow-driven experimentation
It is designed for users who want more control than simple “load LoRA and set strength” systems.
Intended Users
This extension is best suited for:
Advanced ComfyUI users
Model experimenters
Users blending Qwen models
Creators building automated prompt-reactive pipelines