sd-webui-additional-networks  by kohya-ss

WebUI extension for Stable Diffusion with LoRA networks

created 2 years ago
1,831 stars

Top 24.1% on sourcepulse

GitHubView on GitHub
Project Summary

This extension for AUTOMATIC1111's Stable Diffusion web UI enables on-the-fly application of multiple LoRA models to generated images without requiring model merging. It targets users who want to fine-tune image generation by layering stylistic or character-specific LoRAs, offering advanced control over their application.

How It Works

The extension injects LoRA models directly into the Stable Diffusion pipeline during inference. It supports models trained with sd-scripts, allowing up to five LoRAs to be applied sequentially with adjustable weights. A key feature is experimental masked LoRA application, where an RGB mask image controls the spatial influence of individual LoRAs on specific image regions, mapping R, G, and B channels to LoRAs 1-3 respectively.

Quick Start & Requirements

  • Installation: Install via the "Install from URL" tab in the Stable Diffusion web UI's "Extensions" tab, using the repository URL. Restart the web UI after installation.
  • Model Placement: Place LoRA models (.pt, .ckpt, .safetensors) in the sd-webui-additional-networks/models/LoRA folder.
  • Usage: Access the "Additional Networks" panel in the web UI, select LoRA modules, models, and weights.
  • Dependencies: Requires AUTOMATIC1111's Stable Diffusion web UI.

Highlighted Details

  • Supports applying up to 5 LoRA models simultaneously with individual weight control.
  • Experimental masked LoRA feature allows region-specific application using an RGB mask.
  • Integrates with X/Y plot for testing different LoRA combinations and weights.
  • Compatible with ControlNet.

Maintenance & Community

The project is actively maintained by kohya-ss, a prominent figure in the Stable Diffusion training community. Recent updates address model loading issues, X/Y plot compatibility, and specific LoRA module compatibility.

Licensing & Compatibility

The repository does not explicitly state a license in the README. Compatibility with commercial use or closed-source projects is not specified.

Limitations & Caveats

The masked LoRA feature is experimental, with limitations on the number of maskable models (1-3) and potential resolution issues in deeper U-Net layers. Masked LoRA does not affect Text Encoder modules. The README notes that compatibility with future LoCon versions is not guaranteed.

Health Check
Last commit

1 year ago

Responsiveness

1 day

Pull Requests (30d)
0
Issues (30d)
1
Star History
18 stars in the last 90 days

Explore Similar Projects

Feedback? Help us improve.