Inpainting

[Ramam001]

ive been using bens and camendurus notebooks but recently i switched it to this on andorid

ive been getting diffrent results on this collab

some what not prefred results on same seed on other collab with same models

[Linaqruf]

I'm not touching any backend code; the repo was straight-up cloned from camenduru's, and after that, I set the remote URL to automatic1111's. So, I don't know what's gone wrong without any comparable images provided.

All I can suggest is:

• Make sure the ENSD for both notebooks is the same. I'm not setting up the eta noise seed delta, but many Colab implementations out there use 31337. That was the old way of shifting the seed to ensure a 1:1 result to NAIFU.
• Remember if you've set the discard penultimate sigmas value in another Colab. I set it to False by default in my notebook because it was the default value, but you can change that in quick settings.
• Remember if you've set the old Karras implementation value in the other Colab. This is the culprit of seed breaking, so you can't get the same result created by the new implementation. This is not my fault, but they changed how it works on 01/01/23 - Fix Karras scheduler doesn't start from the actual max to min AUTOMATIC1111/stable-diffusion-webui#4373. You can revert the changes by setting the option to True in quick settings.
• Make sure you're using the right sampler. I set dpm 2m++ Karras as the default sampler instead of Euler a.
• Don't enable new ui option because they're not only changing the UI but also the backend.

You can read about other changes in this wiki: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Seed-breaking-changes#2023-01-01---Karras-sigma-minmax"