Fig.1 Generated Images (1024×1024) using this code: (a) Futuristic Forest Details, (b) Cosmic Constellation Map, (c) Paint Strokes of a Virtual Reality Art (d) Augmented Human Eye, (e) Architectural Marvels in a Smart City, (f) Macro Shot of a Nano-Insect, (g) Interactive Globe Details, (h) Clockwork of a Time Machine
▶ Generated image files (1024×1024) and a pre-upscaled image file (512×512) are located in this directory.
Fig.2 Comparison of Resolution Before and After Processing: (left) 512×512, (right) 1024×1024
This project aims to generate high-resolution 1024x1024 pixel images using the Stable Diffusion model 1.4/1.5 and an advanced upsampler prototype. It is specifically designed to meet the requirements of strategically important clients who demand higher image resolutions than what is currently available in the standard Stable Diffusion packages.
Currently, a researcher at Stability AI has released it on Colab. However, it only works in a notebook environment. Therefore, I have prepared a program that can generate 1024x1024 pixel images in a Linux environment by providing prompts through command line arguments
▼ Related Tweet
- Generate 1024x1024 pixel images from given prompts.
- Utilizes Stable Diffusion 1.4/1.5 model.
- Runs locally on Linux.
- Python 3.8 or higher
- Version 0.0.15 of the k-diffusion package
- Version 4.31.0 of the transformers package
- Docker (for DevContainer support)
-
Clone this repository.
git clone https://github.com/hogaku/StableDiffusion-Upscaler-CLI.git
-
Open the project folder in Visual Studio Code.
cd StableDiffusion-Upscaler-CLI code .
-
When Visual Studio Code prompts you to reopen the folder in a DevContainer, click "Reopen in Container". Alternatively, you can press F1 and select the "Remote-Containers: Reopen Folder in Container" command. This will build the DevContainer defined in the .devcontainer folder, installing all the required dependencies automatically.
- Clone this repository.
git clone https://github.com/hogaku/StableDiffusion-Upscaler-CLI.git
- Open the project folder in Visual Studio Code.
cd StableDiffusion-Upscaler-CLI code .
- Install required packages.
pip install -r requirements.txt
- Run the setup script.
./setup.sh
First, Rename the .env.example`` file to .envand configure the Stable Diffusion Engine and API key as appropriate. Example.env` configuration:
SD_ENGINE_ID="stable-diffusion-v1-5"
SD_API_SECRET_KEY="sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
Run the following command on your Linux terminal to generate a 1024x1024 pixel image.
python main.py --seed=<YOUR_SEED> --prompt=<YOUR_PROMPT>or
python main.py -s=<YOUR_SEED> -p=<YOUR_PROMPT>Replace <YOUR_SEED> and <YOUR_PROMPT> with the desired seed and prompt for image generation.
Example:
python main.py --seed=12345 --prompt="A beautiful sunset over the mountains."This will produce a 1024x1024 pixel image based on the prompt and seed provided.
The code for generating 512×512 pixel images using stable-diffusion-v1-5 can operate independently.
Pattern1: CLI Execution
python generate.py <ENGINE_ID> <YOUR_PROMPT>Pattern2: Class Invocation(No Return Value)
from generate import SDImageGenerator
engine_id = os.getenv('SD_ENGINE_ID')
generator_key = os.getenv('SD_API_SECRET_KEY')
engine_id = "stable-diffusion-v1-5"
generator = SDImageGenerator(engine_id, generator_key)
# No return value
generator.generate_image(prompt, <OUTPUT_SAVE_DIR>)Pattern3: Class Invocation (Return Value: PIL Image)
from generate import SDImageGenerator
engine_id = os.getenv('SD_ENGINE_ID')
generator_key = os.getenv('SD_API_SECRET_KEY')
engine_id = "stable-diffusion-v1-5"
generator = SDImageGenerator(engine_id, generator_key)
# Return value: PIL Image
output_image = generator.generate_image(prompt, <OUTPUT_SAVE_DIR>, return_image_data=True)