Features

Meta

• A clean and simple API and implementation, to allow adding new features easily.
• A single, distributable package that can be installed with pip install. That'll install all the dependencies automatically, without requiring you to use conda. (with the exception of torch and torchvision, since they don't host the CUDA versions on PyPI)

Image Generation

• Stable Diffusion 2.1 support.
• txt2img.
• img2img.
• Inpainting.
• 14 Samplers: ddim, plms, heun, euler, euler_a, dpm2, dpm2_a, lms, dpm_solver_stability, dpmpp_2s_a, dpmpp_2m, dpmpp_sde, dpm_fast, dpm_adaptive.

Image Filters

• GFPGAN for face restoration.
• RealESRGAN (and anime) for upscaling.

Models

• Models DB lets you easily download known models with a single line of code.
• Models DB allows automatic detection of the config file to use (e.g. for SD 2.1 and 2.0 "v" type models).
• Fast loading/unloading of VAE models, without needing to reload the entire Stable Diffusion model.
• Custom models for Stable Diffusion in .ckpt or .safetensors format.
• Custom models for VAE in .ckpt or .safetensors format.
• Custom models for Hypernetworks in .pt or .safetensors format.
• API to merge two models (in .ckpt or .safetensors format).

Performance & Security

• VRAM Optimizations to allow generating 512x512 images on GPUs with less than 4gb of VRAM.
• "Balanced" optimizations to achieve high rendering speed with fairly low VRAM usage.
• Support for .safetensors file format, for secure model loading.
• Automatic scanning of Stable Diffusion models before loading them (picklescan).
• Support for running on multiple GPUs in parallel.