Tungsten Stable Diffusion Template

This repository is for creating your own Tungsten model with a stable diffusion checkpoint.

Using this template, you can create a Stable Diffusion model in tungsten including followings:

• Basic functionalities in automatic1111/stable-diffusion-webui (e.g. prompt syntax)
• ControlNet - reference only, openpose & depth
• LoRA
• Textual inversion
• Default prompt and negative prompt

For Windows

1. Install Docker Desktop for Windows
2. Download tungsten-sd.zip from v0.1.1 and extract its contents
3. Run update.bat
4. Put your SD files into following directories
   • 1 Checkpoint: tungsten-sd/models/Stable-diffusion
   • LoRAs (optional): tungsten-sd/models/Lora
   • embeddings (optional): tungsten-sd/embeddings
   • VAEs (optional): tungsten-sd/models/VAE
5. Run build_and_push.bat and enter the responses (e.g. username and password at tungsten.run).

For Linux

Prerequisites

• Stable diffusion weights
• Python 3.7+
• Docker

Build your Stable Diffusion model

0. Clone this repository

git clone --recursive https://github.com/tungsten-ai/tungsten-sd.git
cd tungsten-sd

1. Install Tungstenkit

First, install Tungstenkit:

pip install tungstenkit

2. Prepare weights

Put your Stable Diffusion model weights to models/Stable-diffusion.

If you want to have your own LoRA and VAE, refer to advanced configuration.

3. Build model

tungsten build . -n tungsten-stable-diffusion

Push your model & run remotely

1. Create a project on Tungsten

Go to tungsten.run and create a project.

2. Push the model to Tungsten

Log in to Tungsten:

tungsten login

Add tag of the model:

# Example: tungsten tag tungsten-stable-diffusion myproject:v1
tungsten tag tungsten-stable-diffusion <YOUR_PROJECT_NAME>:<YOUR_MODEL_VERSION>

Then, push the model to the project:

tungsten push <YOUR_PROJECT_NAME>

3. Run the model on Tungsten

Visit tungsten.run and go to the project page.

Run your model locally

0. Make your GPU visible to docker

Refer to this page: https://tungsten-ai.github.io/docs/running_models/using_gpus/

1. Start demo

tungsten demo tungsten-stable-diffusion

Advanced configuration

LoRA

1. Put your LoRA model weights to models/Lora.
2. Modify StableDiffusion.get_loras function in tungsten_model.py to adjust the lora magnitude.

VAE

Put your VAE model weights to models/VAE.

Embedding

1. Put your embedding files to embeddings.
2. Customize prompt (see Prompt customization)

Prompt customization

Modify following functions in tungsten_model.py:

• StableDiffusion.get_trigger_words - Add trigger words at the start of the prompt.
• StableDiffusion.get_extra_prompt_chunks - Add extra prompt chunks at the end of the prompt.
• StableDiffusion.get_extra_negative_prompt_chunks - Add extra negative prompt chunks at the end of the negative prompt.