Question: img2img alternative test - significantly less useful since the last updates (since sigma checkbox)

[cmp-nct]

This is just a question to see if my personal experience matches with you guys.

I've been using the img2img alternative test method quite successfully and frequently, it worked very well for me.
I used it for many hours, so I did have some experience. I took a break for a week and now it's useless to me.
There were a few GIT updates to it, included randomness addition and sigma checkbox, those changes came with rendering it useless to me.

I've tried it in many cases, from negative to positive CFG values, from few to many steps.
Before those changes it simply worked, sometimes I had to tweak steps (30-80) and play a bit with CFG (0.5-1.8)
Now it doesn't work anymore for me, in every single case I gave up and used alternative methods to get what I want.

Is that consistent with your experiences since the latest changes ?

[d8ahazard]

Have you tried adjusting the denoising strength all the way up to 1?

[cmp-nct]

Have you tried adjusting the denoising strength all the way up to 1?

Sure, I've spent hours on it (before the latest updates and after).
I tried a ton of configurations, never was happy with the results anymore. Often no result at all.

Are you using the feature with success ?

[EtherealIntellect]

Also from my own subjective experience - since i havent had the time to test thoroughly, the randomness update and before the randomness update produced different results compared to eachother, even when at randomness zero. Would need someone to confirm or deny tho, or i might after collab lifts its gpu restriction for me again

[EtherealIntellect]

Okay, so after some limited testing, the main difference i'm noticing is that the denoising slider actually works now - and its been mentioned in the features showcase new documentation that you need to set it to 1 to get the old/best effect. Can you check if maybe it was at 0.75 and if having it at 1 makes it behave as you'd expect again? If not i can look at it some more in the next few days

Also in collab you can use
!git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
%cd /content/stable-diffusion-webui/
!git checkout 9c48383
!git reset --hard
!pip install -r requirements.txt
!mkdir repositories
!git clone https://github.com/CompVis/stable-diffusion.git repositories/stable-diffusion
!git clone https://github.com/CompVis/taming-transformers.git repositories/taming-transformers
!git clone https://github.com/sczhou/CodeFormer.git repositories/CodeFormer
!git clone https://github.com/salesforce/BLIP.git repositories/BLIP
!pip install -r repositories/CodeFormer/requirements.txt
!wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth

to get the old version for testing purposes

[cmp-nct]

I'll check the collab method you mentioned, that's quite interesting. Never used it before, thanks.

I mostly tested with 1.0, only lowered it when desperate.
I just gave it another test with the current version and tried an easy image.

I used this prompt : "a woman with blue and purple hair and with green eyes is posing for a picture in front of a teal blue background"

I got a somewhat acceptable reconstruction with decode cfg of 1.3 but from what I recall I had visually perfect reconstructions before (especially the eyes didn't work out, I was not able to get them look right in the current version always bad pupils)

I tried changing the background to yellow, that was not possible.
Hair color change worked, getting the eyes right was not possible.

I'll need to dig into the collab like you said to get a fair comparison, maybe my previous experience just was lucky with the images I used.

[cmp-nct]

Okay, so after some limited testing..

That's not related though I guess you did look into the Cross Attention code ?
https://github.com/bloc97/CrossAttentionControl/blob/main/CrossAttention_Release_NoImages.ipynb

I stumbled upon it an hour ago and was surprised no one adapted it into this repository yet

[srizzo]

Unchecking all the Override checkboxes and setting the values manually + putting more weight to the updates, e.g. (blue:1.3) seems to help. Credits to #4939