This extension for AUTOMATIC1111's Stable Diffusion Web UI enables fine-grained control over model merging by allowing users to specify different weighting ratios for each of the 25 U-Net blocks. It targets users who want to experiment with and optimize the results of merging different Stable Diffusion models, offering a more nuanced approach than standard weighted merging.

How It Works

The extension implements the "Merge Block Weighted" concept, allowing users to assign distinct weights to the input, middle, and output blocks of the U-Net architecture. This granular control is achieved through a GUI where users can select two models (model_A and model_B), define an output name, and then adjust sliders or input text values for each of the 25 U-Net blocks. This approach allows for precise tuning of how features from each base model are incorporated into the merged model, potentially leading to improved or novel results compared to uniform weighting.

Quick Start & Requirements

• Install via the Extensions tab in AUTOMATIC1111's Web UI by providing the repository URL.
• Requires AUTOMATIC1111's Stable Diffusion Web UI.
• Official documentation and examples are available within the README.

Highlighted Details

• Supports saving merged models as half-precision (float16) or safetensors.
• Includes functionality to skip or reset CLIP position_ids, addressing potential issues with prompt embedding compatibility in merged models.
• Allows saving merge logs for reproducibility and analysis.
• Offers various preset weighting functions (e.g., Cosine, Smoothstep) and the ability to create custom presets via TSV files.

Maintenance & Community

The project is an extension for a popular UI, indicating a potentially active user base. Specific contributor or community links (like Discord/Slack) are not explicitly mentioned in the README.

Licensing & Compatibility

The README does not explicitly state a license. Given its nature as an extension for AUTOMATIC1111's Web UI, it is likely intended for non-commercial use or requires careful review for commercial applications, depending on the underlying licenses of the models being merged and the UI itself.

Limitations & Caveats

The project is an extension and relies on the stability and compatibility of the AUTOMATIC1111 Stable Diffusion Web UI. The effectiveness of specific weighting schemes is experimental and may require significant user effort to tune.