imgailab/sdxl-onnx-cpu

sdxl-onnx-fp32

ONNX optimized version of stabilityai/stable-diffusion-xl-base-1.0 with FP32 precision for maximum compatibility.

Available Components

• unet: FP32 optimized
• vae_decoder: FP32 optimized
• vae_encoder: FP32 optimized
• text_encoder: FP32 optimized
• text_encoder_2: FP32 optimized

Usage

Basic CPU Usage

from optimum.onnxruntime import ORTStableDiffusionPipeline

# Models use FP32 for maximum compatibility
pipe = ORTStableDiffusionPipeline.from_pretrained(
    "Mitchins/sdxl-onnx-fp32",
    provider="CPUExecutionProvider"
)

result = pipe("a red apple on a table")
result.images[0].save("output.png")

GPU Usage (CUDA)

pipe = ORTStableDiffusionPipeline.from_pretrained(
    "Mitchins/sdxl-onnx-fp32",
    provider="CUDAExecutionProvider"
)

Performance Benefits

• Compatibility: Works reliably on CPU and GPU
• Speed: ONNX runtime optimizations
• Stability: No type mismatch issues
• Quality: Full FP32 precision

File Structure

All models are FP32 for compatibility:

unet/

• model.onnx (3.9MB + 9794.1MB data) - FP16 precision

vae_decoder/

• model.onnx (188.9MB) - FP16 precision

vae_encoder/

• model.onnx (130.4MB) - FP16 precision

text_encoder/

• model.onnx (469.7MB) - FP16 precision

text_encoder_2/

• model.onnx (0.8MB + 2649.9MB data) - FP16 precision

Generated: 2025-08-08 11:05 UTC with onnxruntime 1.22.1