This extension provides Dreambooth training capabilities for Stable Diffusion WebUI, targeting users who want to fine-tune models with custom concepts. It offers optimizations for lower VRAM GPUs and supports training multiple concepts simultaneously, aiming to simplify and enhance the Dreambooth workflow.

How It Works

The extension is a port of Shivam Shriao's Diffusers Repo, incorporating optimizations for lower VRAM GPUs and features from Koyha SS. It allows for training on custom datasets with configurable parameters for learning rate, batching, image processing, and regularization images. Advanced options include gradient checkpointing, mixed precision, and various attention mechanisms (xformers, flash_attention) to manage VRAM usage and training speed.

Quick Start & Requirements

• Installation: Via the "Extensions" tab in Stable-Diffusion-WebUI, then "Load from:" and install "Dreambooth".
• Prerequisites: Requires diffusers version >= 0.10.0 (SD-WebUI uses 0.3.0).
• Setup: Restart Stable-Diffusion-WebUI after installation. Use REQS_FILE=.\extensions\sd_dreambooth_extension\requirements.txt or DREAMBOOTH_SKIP_INSTALL=True environment flags for specific installation scenarios.
• Documentation: Dreambooth Usage Tutorial

Highlighted Details

• Optimized for lower VRAM GPUs.
• Supports training multiple concepts simultaneously.
• Includes a "Performance Wizard" (WIP) for automatic parameter tuning.
• Offers options for Lora model merging and checkpoint generation.
• Supports Imagic training for faster, single-image fine-tuning.

Maintenance & Community

• Based on work by ShivamShrirao and BMaltais.
• Active development and community support via Discord.
• Discord

Licensing & Compatibility

• License details are not explicitly stated in the README, but it builds upon Huggingface Diffusers and Stable Diffusion WebUI, which typically use permissive licenses. Compatibility with commercial use depends on the underlying licenses of its dependencies.

Limitations & Caveats

• The "Performance Wizard" is marked as Work In Progress (WIP).
• Training text encoders may require >= 12GB VRAM.
• Some features like inpainting training are listed as "Coming soon".