Simple Flux.1 Merger for ComfyUI
A custom ComfyUI node set for merging Flux.1-based models with intuitive control. This extension provides both simplified group merging and expert per-layer control, including support for advanced difference-based merge modes. Basically the code was generated by ChatGPT.
Nodes (2)
Simple Flux.1 Merger for ComfyUI
A custom ComfyUI node set for merging Flux.1-based models with intuitive control.
This extension provides both simplified group merging and expert per-layer control, including support for advanced difference-based merge modes.
Basically the code was generated by ChatGPT.
📦 Included Nodes
1. SimplifiedFlux1Merge
- Intuitive UI with 9 group-based sliders for merging double and single transformer blocks.
- Preserves control over top-level layers like
img_in,txt_in,final_layer. - Designed for quick and effective model blending.
2. ExpertFlux1Merge
- Fine-grained control for all 19
double_blocks, 38single_blocks, and all top-level blocks (img_in,txt_in, etc.). - Ideal for advanced users seeking precise model tuning.
- Includes interpolation-safe FP16/BF16 handling and memory optimization.
🧠 Supported Merge Modes
| Mode | Description |
|-------------------|-----------------------------------------------------------------------------|
| standard | Linear interpolation between A and B (A * w + B * (1-w)) |
| add_difference | Adds difference between B and A to A (A + w * (B - A)) |
| train_difference| Adds training diff from C to A (A + w * (B - C)), mimicking LoRA learning |
🛠 Installation
Clone or download this repository into your ComfyUI custom_nodes/ directory:
git clone https://github.com/yourname/comfyui-simplified-flux1.git
🚀 Usage
Launch ComfyUI.
Add either:
Simplified Flux.1 Merge node
Expert Flux.1 Merge node
Connect 3 Flux.1-compatible models to model_a, model_b, and model_c.
Adjust sliders to define merge ratios.
Select merge mode and execute.
📌 Notes
Compatible with Flux.1 and Schnell variants (if matching architecture).
Requires sufficient GPU memory; optimized for BF16.
train_difference mode requires model C to match A/B in structure.
📜 License
MIT License
🙏 Credits
Based on Flux.1 architecture by Black Forest Labs.
Inspired by SuperMerger.
Developed for ComfyUI advanced model merging workflows.