Summary

<aside> 📢 Note: All tests done are done with Highres fix, when I have time I’ll try to do one too without the highres fix All provide decent results and adherence to the prompts, however

DPM fast, LMS require extra fixing to improve the image (Note that the highres fix is already applied to the images below)

• From experimenting, easiest fix is to try increasing sampling steps Ex. Increasing to 50 steps easily solves the clarity of DPM fast’s sampling at 20
• It is optimal to apply highres fix to all samplers however it is not necessary and uses up a lot of colab runtimes

(In hindsight, highres fix does help a lot with better quality images, however as it requires more time, I would suggest using no highres if you are experimenting)

• Personal favorites: LMS Karras, DPM ++2M Karras, Euler, DPM++SDE

All Samples Compared with the same master prompt and seed

</aside>

Euler A

Euler A

Euler

Euler

LMS

LMS

Heun

Heun

DPM2

DPM2

DPM2 a

DPM2 a

DPM++2S a

DPM++2S a

DPM++ 2M

DPM++ 2M

DPM++SDE

DPM++SDE

DPM fast

DPM fast

DPM adaptive

DPM adaptive

LMS Karras

LMS Karras

DPM2 Karras

DPM2 Karras

DPM2 a Karras

DPM2 a Karras

DPM++2S a Karras

DPM++2S a Karras

DPM++ 2M Karras

DPM++ 2M Karras

DPM++ SDE Karras

DPM++ SDE Karras

DDIM

DDIM

PLMS

PLMS

UniPC

UniPC

With Fix

Deviation and change from original settings to improve quality:

LMS Fix

LMS

LMS (Sampling Steps 50)

LMS + Inpainting

Since this is a small issue, inpainting can be used to remove the specs on the faceBut from further tests, simply increasing the sampling steps should be sufficient

DPM fast Fix

DPM fast

DPM fast (Sampling Steps 50)

Requires at least 50+ Sampling steps to get a cohesive image, but even so the image is not as good as using the other samplers