stable-diffusion-1.5 optimized for AMD GPU
Original Model
https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5
_io32/16
_io32: model input is fp32, model will convert the input to fp16, perform ops in fp16 and write the final result in fp32
_io16: model input is fp16, perform ops in fp16 and write the final result in fp16
Running
1. Using Amuse GUI Application
Use Amuse GUI application to run it: https://www.amuse-ai.com/ use _io32 model to run with Amuse application
2. Inference Demo
Use the code below to get started with the model.
With Python using Diffusers OnnxStableDiffusionPipeline
Required Modules
accelerate
numpy==1.26.4 # Due to newer version of numpy changing dtype when multiplying
diffusers
torch
transformers
onnxruntime-directml
Python Script
import onnxruntime as ort
from diffusers import OnnxStableDiffusionPipeline
model_dir = "D:\\Models\\stable-diffusion-v1-5_io32"
batch_size = 1
num_inference_steps = 30
image_size = 512
guidance_scale = 7.5
prompt = "a beautiful cabin in the mountains of Lake Tahoe"
ort.set_default_logger_severity(3)
sess_options = ort.SessionOptions()
sess_options.enable_mem_pattern = False
sess_options.add_free_dimension_override_by_name("unet_sample_batch", batch_size * 2)
sess_options.add_free_dimension_override_by_name("unet_sample_channels", 4)
sess_options.add_free_dimension_override_by_name("unet_sample_height", image_size // 8)
sess_options.add_free_dimension_override_by_name("unet_sample_width", image_size // 8)
sess_options.add_free_dimension_override_by_name("unet_time_batch", batch_size)
sess_options.add_free_dimension_override_by_name("unet_hidden_batch", batch_size * 2)
sess_options.add_free_dimension_override_by_name("unet_hidden_sequence", 77)
pipeline = OnnxStableDiffusionPipeline.from_pretrained(
model_dir, provider="DmlExecutionProvider", sess_options=sess_options
)
result = pipeline(
[prompt] * batch_size,
num_inference_steps=num_inference_steps,
callback=None,
height=image_size,
width=image_size,
guidance_scale=guidance_scale,
generator=None
)
output_path = "output.png"
result.images[0].save(output_path)
print(f"Generated {output_path}")
Inference Results
