Hello, I wanted to try using Stable Diffusion, so I ran some benchmarks.I used a random cityscape as the prompt and tried out various models.It feels like there are slight differences in the fine details depending on the machine and tool used, but overall, they seemed quite similar. The combinations of machines I used involved GPUs: T4 and H100. I only used Automatic1111 with the H100. Automatic1111<a href="https://github.com/AUTOMATIC1111/stable-diffusion-webui" rel="noopener noreferrer" target="_blank">https://github.com/AUTOMATIC1111/stable-diffusion-webui</a> Forge<a href="https://github.com/lllyasviel/stable-diffusion-webui-forge" rel="noopener noreferrer" target="_blank">https://github.com/lllyasviel/stable-diffusion-webui-forge</a> T4 + Automatic1111 / 640*360<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/SDT4_640.png"/> T4 + Forge / 640*360<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/SDT4F_640.png"/> H100 + Automatic1111 / 640*360<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/SDH100_640.png"/> The result for 640x360 pixels are<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/%E3%82%BF%E3%82%99%E3%82%A6%E3%83%B3%E3%83%AD%E3%83%BC%E3%83%88%E3%82%99.png"/>And I tried larger images. T4 + Automatic1111 / 1920*1080<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/SDT4_1920.png"/> T4 + Forge / 1920*1080<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/SDT4F_1920.png"/> H100 + Automatic1111<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/SDH100_1920.png"/> The benchmark is,<img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2024/03/18/%E3%82%BF%E3%82%99%E3%82%A6%E3%83%B3%E3%83%AD%E3%83%BC%E3%83%88%E3%82%99%20(1).png"/> I feel like the images are quite different as well, but I did make sure to use a fixed seed for the runs.The H100 is overwhelmingly faster, as expected. I feel like using Forge might make it even faster.I wonder about the cost of providing such services for the companies involved...I plan to continue benchmarking little by little in the future.

Stable Diffusion speed comparison between Forge and Automatic1111, using T4 and H100 GPUs

Yuichiro Minato