This repo contains an implementation of Stable Diffusion inference running on top of ONNX Runtime, written in Java. It's a modified port of the C# implementation, with a GUI for repeated generations and support for negative text inputs. It is intended to be a demonstration of how to use ONNX Runtime from Java, and best practices for ONNX Runtime to get good performance. We will keep it current with the latest releases of ONNX Runtime, with appropriate updates as new performance related ONNX Runtime features become available through the ONNX Runtime Java API. All the code is subject to change as this is a code sample, any APIs in it should not be considered stable.

This repo targets ONNX Runtime 1.14. The version number is in two parts <sd4j-version>-<onnxruntime-version>, and the initial release of sd4j is v1.0-1.14.0. We'll bump the sd4j version number if it gains new features and the ONNX Runtime version number as we depend on newer versions of ONNX Runtime.

The project supports txt2img generation, it doesn't currently implement img2img, upscaling or inpainting.

By default it uses a fp32 model, and running on a 6 core 2019 16" Intel Macbook Pro each diffusion step takes around 5s. Running on better hardware, or with a CUDA GPU will greatly reduce the time taken to generate an image, as will using an SD-Turbo model. There is experimental support for the CoreML (for macOS) and DirectML (for Windows) backends, but proper utilisation of these may require model changes like quantization which is not yet implemented.

These are a few example images generated by this code along with their generation parameters:

Text: "Wildlife photograph of an astronaut riding a horse in the desert", Negative Text: "", Seed: 42, Guidance Scale: 10, Inference Steps: 40, Scheduler: Euler Ancestral, Image Size: 512x512.

Text: "Press photo of an America's Cup catamaran sailing through the sands of Mars, high resolution, high quality", Negative Text: "water, sea, ocean, lake", Seed: 42, Guidance Scale: 10, Inference Steps: 40, Scheduler: Euler Ancestral, Image Size: 512x512.

Text: "Professional photograph of the Apollo 11 lunar lander in a field, high quality, 4k", Negative Text: "", Seed: 42, Guidance Scale: 10, Inference Steps: 50, Scheduler: Euler Ancestral, Image Size: 512x512.

Text: "Professional photograph of George Washington in his garden grilling steaks, detailed face, high quality, 4k", Negative Text: "painting, drawing, art", Seed: 42, Guidance Scale: 10, Inference Steps: 60, Scheduler: Euler Ancestral, Image Size: 512x512.

The SD4J project supports SD v1.5, SD v2 and SDXL style models. For models which do not support classifier-free guidance or negative prompts, such as SD-Turbo or SDXL-Turbo, the guidance scale should be set to a value less than 1.0 which disables that guidance. Models like SD-Turbo can generate acceptable images in as few as two diffusion steps. The difference between SDv1 and SDv2 models is autodetected, but SDXL must be supplied as the model type for SDXL models otherwise it will throw an exception on generation. In some cases the autodetection of v1 and v2 may fail in which case supplying the --model-type {SD1.5, SD2, SDXL} argument with the appropriate parameter will fix the model type.

This project requires Apache Maven, Java 17 or newer, a compiled ONNX Runtime extensions binary, and a Stable Diffusion model checkpoint. The other dependencies (ONNX Runtime and Apache Commons Math) are downloaded by Maven automatically.

There are many compatible models on Hugging Face's website. We have tested the Stable Diffusion v1.5 checkpoint, which has pre-built ONNX models. This can be downloaded via the following git commands (skip the first one if you have already configured git-lfs):

git lfs install
git clone https://huggingface.co/runwayml/stable-diffusion-v1-5 -b onnx

The Stable Diffusion v1.5 checkpoint is available under the OpenRAIL-M license. For other SD models there is a one or two stage process to generate the ONNX format models. If the model is already in Hugging Face Diffusers format then you can run the convert_stable_diffusion_checkpoint_to_onnx.py file from the diffusers project as follows:

python scripts/convert_stable_diffusion_checkpoint_to_onnx.py --model_path <path-on-disk-or-model-hub-name> --output_path <path-to-onnx-model-folder>

If the model is an original stable diffusion checkpoint then you first need to run:

python scripts/convert_original_stable_diffusion_to_diffusers.py --checkpoint_path <path-on-disk-to-checkpoint> --scheduler_type lms --dump_path <path-on-disk-to-diffusers-output>

If the model is an SDXL model then it needs to be exported from the Hugging Face Hub using optimum:

optimum-cli export onnx --model <model-hub-name> <path-to-onnx-model-folder>

The scripts require a suitable Python 3 virtual environment with diffusers, onnxruntime, optimum and onnx installed.

You will also need to check out and compile onnxruntime-extensions for your platform. The repo is https://github.com/microsoft/onnxruntime-extensions, and it can be compiled with ./build_lib.sh --config Release --update --build --parallel which generates the required library (libortextensions.[dylib,so] or ortextensions.dll) in the build/<OS-name>/Release/lib/ folder. That library should be copied into the root of this directory.

The GUI can be executed with mvn package exec:exec -DmodelPath=<path-to-stable-diffusion-model>. It constructs a window where you can specify the parameters of the image you'd like to generate, and each image creates its own window where it can save the image as a png file. Saved png files contain a metadata field with the generation parameters.

The com.oracle.labs.mlrg.sd4j.SD4J class provides a full image generation pipeline which can be used without the GUI directly from other code.

To use the GPU you need to modify the pom file to depend on onnxruntime_gpu and swap <argument>CPU</argument> for <argument>CUDA</argument> in the exec-maven-plugin block.

This code provides a thin Tensor wrapper object which is a tuple of a direct ByteBuffer instance and a long shape array, which is used to provide easy access in and out of ORT's OnnxTensor objects. There's a Scheduler interface which the two available schedulers (LMS and Euler Ancestral) implement. The SD4J pipeline object is a suitable entry point for use without the GUI, and there is an example of such usage in the CLIApp class.

This project welcomes contributions from the community. Before submitting a pull request, please review our contribution guide.

Please consult the security guide for our responsible security vulnerability disclosure process

The code is available under the Universal Permissive License (UPL). It requires a Stable Diffusion model architecture checkpoint to work, and any Stable Diffusion models should be used under their licenses. There are 1000+ compatible models available on Hugging Face each of which are licensed separately, though many use a variant of the OpenRAIL-M license.

The tokenizer onnx model is taken from the C# implementation, and is available under the MIT license. More details on the tokenizer are available in its README file.