ouhenio / sd-img2img-speed-benchmark

Stable Diffusion img2img speed benchmarks

What is this?

The goal of this repo is to measure and compare different img2img acceleration methods.

Table of results

To replicate the results, simply run main.py.

Note that these measurements where done with an RTX 3090.

Method	speed (seconds)
Vanilla	8.72
float.16	3.28
torch.compile	8.58
torch.compile + float.16	3.15

Todos

• Use different torch.compile backend methods.
• Use Pytorch's 2 scaled_dot_product_attention.
• Measure the difference between a manualy compiled pipeline (via torch.jit.trace) versus torch.compile.
• Experiment with different batch sizes.