This is a quick becnhmark of different Stable Diffusion sampling methods using an Nvidia Geforce GTX 1660 Super video card with a simple prompt and no LoRA.

Hardware involved for this test

• Gainward GHOST Geforce GTX 1660 Super (6GB VRAM) (view settings)
• Nvidia driver version 531.68
• Kingston Fury 16GB DDR4-3200 RAM
• Core i3 9100F CPU

Benchmark results

1. DPM++ 2M SDE: 1.70 it/s
   
2. DPM++ 2M SDE Karras: 1.70 it/s
   
3. DPM++ 2M: 1.72 it/s
4. DPM++ 2M Karras: 1.72 it/s ◀ best speed/quality ratio in this case
   
5. Euler a: 1.73 it/s
6. Euler:  1.73 it/s
7. DDIM: 1.84 it/s ◀  but achieves good results in less steps
8. DPM++ SDE Karras: 1.17 s/it
9. DPM++ SDE: 1.18 s/it

As you can see, this card cannot generate 512×512 images faster than a quarter of iteration per second. Lowering the resolution (not recommended) does produce much faster results.

Stable Diffusion sampling methods comparison

• 2M Karras: Clear winner here, result are less prone to glitches and imperfections.
• 2M SDE: Fast, however both methods produce malformed/distorted images in this case.
• SDE Karras: Good quality, but twice slower than 2M Karras.
• DDIM: Further testing conclude that DDIM is faster in the sense that good results can be achieved in less steps than other methods.

Euler, Euler a and DDIM steps comparison

Both Euler a and DDIM achieved good results in 20 steps. Euler took 30 steps to create a glitch-free image.

As shown here, increasing steps can potentially begin to distort a previously fine image. DDIM, which is generally known for achieving good results in less steps, made an image with less glitches than Euler a in just 20 steps. Notice how DDIM got a proper lamp post to the left in 20 steps. More steps mean longer generation time.

Conclusion

There is no absolute winner when it comes to sampling methods. DDIM is known to be cheaper in the sense of achieving goods results in less steps, however, I prefer to use DPM++ 2M Karras alone which is slightly slower, but generates more consistent results.

Some methods are known to produce better results for specific checkpoint models, so when testing a new checkpoint, always test it with the sampling methods that you are most comfortable with. If results are identical in quality, then you should pick the fastest among them.