Trying to match results after updates - Original installed version 43cca93

[creeduk]

I started with this build - 43cca93

Loved it had a few memory issues if I pushed it too far but running and on 2gb of VRAM so I was very happy. I updated recently to these versions (Friday, Saturday and yesterday) fe40f20, 0c45b5 and currently 3a1c5b9.

I noticed some burn out on some of my images like over configured, so I tried to replicate my earlier testing. One example was the dog sketch from the provided sample folder. I posted on reddit as well but figured it was time to ask here as maybe somebody knows a setting or if the code changed to a point where old results are no longer possible?

Sample used as sketch:

Results from 43cca93:

highly detailed, masterpiece, dog,
Negative prompt: longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality
Steps: 20, Sampler: DPM++ SDE, CFG scale: 7, Seed: 925894899, Size: 512x512, Model hash: 629ddef959, Model: inkpunk-diffusion-v1, ControlNet Enabled: True, ControlNet Module: scribble, ControlNet Model: controlnetPreTrained_scribbleV10 [c508311e], ControlNet Weight: 1

Results from 3a1c5b9:

highly detailed, masterpiece, dog,
Negative prompt: longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality
Steps: 20, Sampler: DPM++ SDE, CFG scale: 7, Seed: 925894899, Size: 512x512, Model hash: 629ddef959, Model: inkpunk-diffusion-v1, ControlNet Enabled: True, ControlNet Module: scribble, ControlNet Model: controlnetPreTrained_scribbleV10 [c508311e], ControlNet Weight: 1, ControlNet Guidance Start: 0, ControlNet Guidance End: 1

Tried numerous ways to match I would assume guidance start 0, end 1 and weight 1 but none of my test subjects match. I can confirm model has not changed and if I test old generations without controlnet they are exact (tested in PS with difference mode)

[ljleb]

Not sure, but maybe try toggling the "Only use mid-control when inference" option in the settings? May be related to #373

[creeduk]

I saw that post a few days ago when I had the image burn out and at first I thought oh great the colors are fixed, but then I realized it was not following my sketch or any outlines at all. Not sure how that works it is like having a really low weight from what I could tell or a low cfg.

Before all the new settings and sliders I was wondering what were all the default hardcoded values in 43cca93?

[Mikubill]

Tested with 96f2e1c

3a1c5b9	96f2e1c

This change was introduced by #350 10 hours ago - already reverted.

[creeduk]

I confirm that the weekend update and todays update create same output with only a tiny difference on the collar stiches. I still have different output from 43cca93. I plan to temp roll back to double check. Did some fixes that maybe make the function better change the output? Sometimes I know we have to move on and the old cannot be recreated.

Thank you for checking and providing feedback. I know consistency is a huge challenge with this technology.