Stable Diffusion XL on an AMD GPU

Hello. I thought there might be people who want to run Stable Diffusion on AMD, so I decided to give it a try.

The CPU I used is a Ryzen 9 5950X, and the GPU is a Radeon 7900XTX (with 24GB VRAM).
Since the VRAM capacity is sufficient, I wanted to check the performance and usability.

I utilized Stable Diffusion's SDXL for this test.

You need to install PyTorch that is compatible with ROCm, AMD's standard. After that, I installed the necessary libraries using pip as usual:

pip install diffusers invisible_watermark transformers accelerate safetensors

The usage is the same as with NVIDIA GPUs. Since I'm using the standard SDXL setup without any special libraries, I didn't encounter any issues.

from diffusers import DiffusionPipeline
import torch

# Specify the GPU
device = "cuda"

# Set a fixed seed
generator = torch.Generator(device).manual_seed(100)

# Write the prompt
prompt = "An astronaut riding a green horse"

# Build the pipeline and set to fp16.
pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
pipe.to(device)

# Generate and retrieve the image
images = pipe(prompt=prompt, generator=generator).images[0]
images

There was no noticeable difference from the CUDA version.

For a 1024x1024 image with 50 steps (the standard number of steps), the generation speed was 16 seconds.

While the generation speed is slightly slower, it was still quite usable.