This repo holds Stability AI's AWS Marketplace containers.
Container images are all tested on AWS EC2 instance types g5.2xl, which are equipped with Nvidia A10 GPUs.
If you have problems, please open an issue in this repository or submit a ticket at https://platform.stability.ai/support.
For commercial use of these containers, please see Stability Memberships.
SDXL Turbo is an adversarial time-distilled Stable Diffusion XL (SDXL) model capable of running inference in as little as 1 step.
This container runs the SDXL Turbo model as a web service using FastAPI. It is available on the AWS Marketplace from Stability AI.
Once you have pulled the Docker image for the Marketplace product, run the container to start the server:
Run on a cuda device with:
$ docker run --gpus all -p 8080:8080 -t <image-name>
To test the model:
cd sdxl_turbo
pip install -r client_requirements.txt
python client_request.py
Take a look at sdxl_turbo/client_request.py for an example of a text-to-image request and an image-to-image request.
Stable Video Diffusion (SVD) Image-to-Video is a diffusion model that takes in a still image as a conditioning frame, and generates a 25 frame video from it.
image: The image to be used as a conditioning frame. Must be of size (1024, 576) pixels.
motion_bucket_id: An int This can be used to control the motion of the generated video. Increasing the motion bucket id will increase the motion of the generated video.
Defaults to 127.
noise_aug_strength: A float between [0, 1]. A multiplier, used to determine the amount of noise added to the conditioning image. The higher the values the less the video will resemble the conditioning image. Increasing this value will also increase the motion of the generated video.
Default 0.2.
Run on a cuda device with:
$ docker run --gpus all -p 8080:8080 -t <image-name>
To test the model:
cd svd
pip install -r client_requirements.txt
python client_request.py
The client_request.py demo code uses the ffmpeg Python bindings to change the codec of the output video to h264, so that the resulting video can be previewed in VSCode. This is not essential to the functionality of the container and can be removed.