[zoom_enhance] txt2img broken on latest Automatic1111

[JDadders]

On the new version of Automatic1111, [zoom_enhance] throws the following error:

"ERROR - Could not check length of img2img_images"

And fails to fill in the mask (the output shows just the image with the black mask over the area needing enhancing).

However, img2img seems to work as expected.

Fully expected that the new version of Auto would break things, I'm just hopeful posting the error message might help you track down the issue. Thanks for this great tool!

Also unfortunately I'm still seeing the lower performance since v8 that I mentioned here but that's a seperate issue :)

[Poliveris]

Did you ever get this figured out? I can't even get 7.8 to work with SD A1111 1.0

[JDadders]

I didn't. I had a backup of my previous working setup (SD A1111 1.0 + Unprompted from March 25th which is basically 7.91 plus another fix) and I've been using that ever since the SD A1111 update.

I noticed your comment thread you posted earlier and I have to agree - I love this extension and loads of love to the creator but it's a real shame that the latest versions of unprompted have downgraded the quality of zoom_enhanve so much. As you say, it used to be absolutely essential anda game changer but now I'm stuck on old versions of A1111 and Unprompted because the 'upgrade' is so much worse than the old version. Hopefully there'll be a fix that brings back the old quality in the future.

[Poliveris]

It is unfortunate :/ I do hope this is eventually fixed.

Might I ask what your other fix was for 7.9.1? I don't have any errors anymore but it doesn't seem to generate that new image and/or stitch. What additional fix did you need to get it running? Currently got 1.0 installed with 7.9.1

It says it's downloading the clipseg but does nothing after and only shows original in generation

[JDadders]

Ah I just meant that the 25th of March checkin to github which I'm using is v7.9.1 + another fix that the author commited on the 25th, I haven't done anything to it personally.

In theory that version with v1 SD A1111 should work. I hope you manage to get it up and running again soon, I know how frustrating it can be to be on a broken version.

[Poliveris]

What version of python are you on?

Does your beginning process look like this: Python 3.10.9 (tags/v3.10.9:1dd9be6, Dec 6 2022, 20:01:21) [MSC v.1934 64 bit (AMD64)] Commit hash: 48a15821de768fea76e66f26df83df3fddf18f4b Installing requirements for Web UI Installing requirements for Unprompted - img2pez Installing requirements for Unprompted - pix2pix_zero

[JDadders]

I'm on Python 3.10.8

[Poliveris]

Tried 3.10.8 still nada; now sometimes it doesn't do the clipseg download. Not sure what else to do at a low at this point I guess. Although I really do need this to work.

[Poliveris]

Whats your command prompt look like after an image is generated? Mine does the clipseg thing to 100% but no image is generated in temp and it doesn't showcase the 512x512 upscale in the preview like it used too

And also curious what version of torch are you using? 1.13 or 2.0?

[JDadders]

Hey, sorry for the late reply. My Torch version is 1.13.

After a generation my command prompt just shows the iteration progress bar after a generation. I don't see the clipseg download as I already have it locally.

Did you manage to make any progress on fixing your copy?

[Poliveris]

I couldn't get any <8.0 version to work; and eventually even 9.2.0 was messing up by not stitching the image together or when it did it was just the unmodified version ending up in %TEMP%.

I end up having to use a modified Zoom_enhance.py file provided by the creator a few months back for this specific fix.https://github.com/ThereforeGames/unprompted/issues/84

So for now just gonna stick with 9.2.0 the results aren't what I was used too; but it's definitely still pretty helpful overall. Just hoping the creator adds the old Img2img option.