Instruct Pix2Pix not working as expected on 1.6.0

[Gregwar]

Hello, I was trying Instruct Pix2pix on 1.5.0 with the [extension][https://github.com/Klace/stable-diffusion-webui-instruct-pix2pix), and it was working as expected:

I updated to 1.6.0, where Pix2pix should be supported by default, but it is always giving me giberish output, just like the denoising is actually doing its typical job:

The settings are the same in both cases (512x512 output, 1.25 image CFG, ~7 CFG), and set Denoising to 1 in 1.6.0.
The checkpoint is the same

Did I miss something ?

[rabbitflyer]

Same issue here.
I'd like to add a few points:

• At least for me, the image CFG scale setting has no effect on the output image.
• AFAIK pix2pix was natively supported prior to 1.6.0, and worked as intended without any extensions present.

[rabbitflyer]

@Gregwar
Update: Based on a bisect I just did, the first bad commit to master is 75336df.

[makkad0]

Same issue

[Gregwar]

Looks like the following rollback gets pix2pix working again on master:

diff --git a/modules/processing.py b/modules/processing.py
index e124e7f0..8b95abd9 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -296,7 +296,7 @@ class StableDiffusionProcessing:
         return conditioning
 
     def edit_image_conditioning(self, source_image):
-        conditioning_image = images_tensor_to_samples(source_image*0.5+0.5, approximation_indexes.get(opts.sd_vae_encode_method))
+        conditioning_image = shared.sd_model.encode_first_stage(source_image).mode()^M
 
         return conditioning_image

As a partial revert of 21000f1
ping @KohakuBlueleaf if you have any idea

[Gregwar]

I don't know what TAESD is actually, I didn't change any setting from 1.5.0 to 1.6.0, did any default change ?

[KohakuBlueleaf]

Thx for information
no default change

Will try to figure this out

[LTusche]

@Gregwar How do I roll it back as temporary fix? I tried to checkout 21000f1 and 6346d8e but got a runtime error when using the instructpix2pix model RuntimeError: Input type (float) and bias type (struct c10::Half) should be the same.

[Gregwar]

Try adding --no-half in that case, does it helps?

[LTusche]

Worked for this error but now there is another one RuntimeError: Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same or input should be a MKLDNN tensor and weight is a dense tensor.
This one seems to be a bit more complex so I guess I'll have to wait until Kohaku has a fix for the original problem.

[w-e-w]

I believe this issue was fixed in 1.7