Stable Diffusion 1.5 ONNX FP32 Model Released

What changed

A Stable Diffusion 1.5 text-to-image model has been independently exported to ONNX format with FP32 precision by user JasonYANG170. This export is based on the stable-diffusion-v1-5/stable-diffusion-v1-5 base model and does not utilize third-party ONNX weights. The model is configured for 512px image generation using the PNDM epsilon scheduler. To address potential allocation failures, specific runtime UNet graph optimizations have been disabled. The conversion process and associated files, including SHA-256 checksums and validation evidence, are detailed in conversion.json.

Why it matters for builders

Developers can leverage this ONNX version to deploy Stable Diffusion 1.5 in environments that natively support ONNX inference. This offers an alternative to PyTorch or TensorFlow deployments, potentially enabling broader compatibility across different hardware and software stacks. The FP32 precision ensures a balance between accuracy and performance, while the specific optimizations address known runtime issues.

Practical impact

Builders can explore integrating this ONNX model into applications requiring text-to-image generation capabilities. The provided conversion/browser-diffusion.mjs harness suggests potential use cases within web environments, with VAE execution noted to work in WASM. Further investigation into the conversion.json file will provide details on file sizes and validation, aiding in integration planning.

Caveats and source limits

The source indicates this is not a performance certification for 16GB integrated GPUs. The model was tested on an RTX 5080 Laptop for WebGPU, but broader performance benchmarks are not provided. The exact performance characteristics and compatibility across various ONNX runtimes and hardware configurations require further independent testing by developers. The model is licensed under creativeml-openrail-m.