Skip to main content

​
Prerequisites

Before you install sd-scripts, make sure your system meets the following requirements:
  • Python 3.10.x — Python 3.11.x and 3.12.x may work but are not officially tested.
  • Git — Required to clone the repository and pull updates.
  • NVIDIA GPU with CUDA support — A CUDA 12.4-capable GPU is recommended. The default PyTorch install targets CUDA 12.4.
Python 3.10.x is the recommended version. Other versions (3.11.x, 3.12.x) are not fully tested and may produce unexpected behavior with some dependencies.

​
Install

​
Windows required dependencies

Download and install the following before proceeding:You also need to allow PowerShell to run virtual environment activation scripts. Open an administrator PowerShell window and run:
Answer A when prompted, then close the administrator window.

​
Installation steps

Open a regular PowerShell terminal and run each command in order:
1

Clone the repository

2

Create and activate a virtual environment

If python is not recognized, try py -m venv venv instead.
3

Install PyTorch with CUDA support

This installs PyTorch 2.6.0 for CUDA 12.4. If you use a different CUDA version, change the index URL accordingly. For CUDA 12.1 use cu121.
4

Install requirements

5

Configure accelerate

When prompted, use these answers:
Answer bf16 to the last question if you want to use bfloat16 mixed precision. If you see ValueError: fp16 mixed precision requires a GPU, answer 0 for the GPU selection question to explicitly select GPU 0.

​
About requirements.txt and PyTorch

requirements.txt does not include PyTorch. The correct PyTorch version depends on your GPU architecture and CUDA version, so you must install it separately before running pip install -r requirements.txt. PyTorch 2.6.0 or later is required. The scripts are tested with PyTorch 2.6.0. For RTX 50 series GPUs, use PyTorch 2.8.0 with CUDA 12.8 or 12.9. The existing requirements.txt is compatible with that version. Key packages installed by requirements.txt include:
  • accelerate — Distributed training and mixed precision launcher
  • transformers, diffusers — Model loading and inference
  • bitsandbytes — 8-bit optimizers for reduced VRAM usage
  • safetensors — Safe model serialization format
  • prodigyopt, lion-pytorch — Alternative optimizers
  • tensorboard — Training metrics visualization

​
Upgrading

When a new release is available, pull the latest code and reinstall requirements:

​
Upgrading PyTorch

To upgrade PyTorch independently, run the same pip install torch==... command from the installation steps above with your target version.

​
xformers (optional)

xformers provides memory-efficient attention implementations that can reduce VRAM usage and speed up training. Install it inside your activated virtual environment:
Change the CUDA version in the URL to match your environment. xformers may not be available for all GPU architectures.