Untitled Merger - Advanced Model Merging for Stable Diffusion

A powerful Stable Diffusion checkpoint merging extension with 15 merge algorithms, LoRA support, and intelligent caching.

🎉 Now Works on Forge!

This is an actively maintained fork of groinge/sd-webui-untitledmerger with major updates and Forge compatibility!

✨ What's New in This Fork

• ✅ Forge Compatibility - Works on both SD WebUI Forge AND A1111!
• ✅ 6 New Merge Modes - Added Triple Sum, Quad Sum, Multiply Difference, Sum Twice, Self, and Tensor modes
• ✅ LoRA Merging - Merge LoRAs to checkpoints or combine multiple LoRAs
• ✅ Bug Fixes - Fixed DARE (Power-up) division by zero crash
• ✅ 4-Model Support - Quad Sum mode finally uses all 4 model slots!

---

🚀 Features

Advanced Merge Algorithms (15 Total!)

Basic Modes:

• Weight-Sum - Classic linear interpolation between two models
• Self - Multiply model weights by a scalar value

Multi-Model Modes:

• Triple Sum - Weighted blend of 3 models
• Quad Sum - Weighted blend of 4 models (NEW!)
• Sum Twice - Hierarchical two-stage merging

Difference-Based Modes:

• Add Difference - Add weighted difference: A + (B-C) × α
• Multiply Difference - Multiplicative difference: A × ((B-C) × α + 1) (NEW!)
• Train Difference - Treat difference as fine-tuning relative to base model

Advanced Interpolation:

• Comparative Interp - Adaptive interpolation based on value differences
• Enhanced Manual Interp - Manual threshold control for interpolation
• Enhanced Auto Interp - Automatic threshold calculation

Research-Based Methods:

• Power-up (DARE) - Drop And REscale from research papers (FIXED!)
• Extract - Merge common/uncommon features between models
• Add Dissimilarities - Add dissimilar features to base model

Experimental:

• Tensor - Swap entire tensors by probability instead of blending (NEW!)

LoRA Support

• Merge LoRA to Checkpoint - Permanently bake a LoRA into a checkpoint with adjustable strength
• Merge Multiple LoRAs - Combine 2-3 LoRA files with custom weights

Performance Features

• Calculation Caching - Reuses intermediate calculations to speed up subsequent merges
• Multi-threading - Configurable worker threads for parallel operations
• Memory Efficient - Smart caching with configurable size limits (0-16GB)

---

📋 Requirements

• Stable Diffusion WebUI (A1111) OR Stable Diffusion WebUI Forge
• Only supports .safetensors format checkpoints
• Python 3.10+ with torch, safetensors, scipy (usually pre-installed)

---

🎯 Compatibility

Platform	Status	Tested Version
SD WebUI Forge	✅ Working	Latest (2024)
SD WebUI A1111	✅ Working	1.7.0 - 1.9.3+
SD 1.5 Models	✅ Fully Supported	All variants
SDXL Models	✅ Generally Works	May have edge cases

---

🙏 Credits & Inspiration

This extension builds upon the excellent work of:

• Original Extension: groinge/sd-webui-untitledmerger
• Supermerger: hako-mikan/sd-webui-supermerger - UI patterns and merge algorithms
• DARE/Power-up: martyn/safetensors-merge-supermario - Research implementation
• MergeLM: yule-BUAA/MergeLM - Theoretical foundations

---

🛠️ Installation

1. Navigate to your WebUI extensions folder:

   cd stable-diffusion-webui/extensions/
   # OR for Forge:
   cd stable-diffusion-webui-forge/extensions/

2. Clone this repository:

   git clone https://github.com/YOUR-USERNAME/sd-webui-untitledmerger.git

3. Restart your WebUI

4. Look for the "Untitled merger" tab!

---

📖 Usage

Basic Checkpoint Merging

1. Select your merge mode from the dropdown
2. Choose base models (2-4 depending on mode)
3. Adjust sliders for merge weights
4. Configure save options (fp16/bf16, autosave, etc.)
5. Click Merge!

Advanced Features

• Custom Sliders - Per-block weight control for fine-grained merging
• Include/Exclude Filters - Regex-based layer targeting
• Supermerger Adjust - Fine-tune detail, contrast, brightness, and color
• YAML Weight Editor - Batch operations via configuration files

LoRA Merging

• LoRA → Checkpoint: Bake a LoRA into a model permanently
• LoRA → LoRA: Combine multiple LoRAs with custom ratios

---

🐛 Known Issues & Limitations

• Image generation tab intentionally removed (not needed for merging workflow)
• LoRA merging uses heuristic key matching (may need adjustment for exotic LoRAs)
• SDXL merging works but less tested than SD 1.5

---

📝 To-Do List

• XYZ plotting for parameter exploration
• State saving and merge history
• Improve LoRA key matching for better compatibility
• Block weight presets compatible with Supermerger format
• Merge history logger

---

📜 License

Same as original - check LICENSE file for details.

---

💬 Support

For issues, feature requests, or questions:

• Open an issue on GitHub
• Check existing issues for solutions

---

Happy Merging! 🎨✨