[Issue]: Hi-Res Fix is not working correctly for the first image on the batch

[TiagoSantos81]

Issue Description

The first pass is using the same dimensional information -- width and weight -- as the supposed final data, although the correct information is shown in on "Hires resize: from XXX:XXX to XXX:XXX". Obvious effect on first pass processing times and the subject duplication effects.

Might be related to the change in resize_mode parameters in c5b089c.

Version Platform Description

This occurs on windows platform on the latest updates. Did not happen two days ago.

Version:
app: automatic.git
updated: 2023-07-09
hash: e8c29597

Platform:
rch: AMD64
cpu: AMD64 Family 23 Model 24 Stepping 1, AuthenticAMD
system: Windows
python: 3.10.9

Relevant log output

No response

Acknowledgements

• I have read the above and searched for existing issues
• I confirm that this is classified correctly and its not an extension or diffusers-specific issue

[vladmandic]

that was fixed, you're using quite an old version.

[TiagoSantos81]

that was fixed, you're using quite an old version.
No, it was updated before the commit you .
The info I am showing is indeed updated at 09-07-2023 but I had made git pull -r before writing this ticket. I just had copy pasted without even looking properly at the info, assuming it to be correct.

Did another git pull moments ago, with success, and I still get the same data in the 'System Information' Tab.

This is my 'From Text' Tab:

As you can see, it contains the most recent layout changes.

So, it seen there is also some problem in the system info tab, that is not updating correctly the Version section.

[TiagoSantos81]

This issue seems to be related only with the first image in the batch. Repeating the seed in another position of the batch seems to fix it, as shown by both images with seed 705512689, shown below:

Second position in the batch:

First position in the batch:

[TiagoSantos81]

And the grids showing the issue:

Starting on seed 705512688:

Starting on seed 705512689:

[vladmandic]

ok, i'll reopen. but this may take a while since i have upcoming trip.

[TiagoSantos81]

No worries. Thank you for looking into it.

[StableInfo]

This issue seems to be related only with the first image in the batch. Repeating the seed in another position of the batch seems to fix it, as shown by both images with seed 705512689, shown below:

Second position in the batch:

First position in the batch:

Hello @TiagoSantos81 Do you mind sharing your workflow for this? and prompt?and modelname? This image is so much high quality I am impressed.
Do you do: text2image + hires -> img2img + ultimate sd upscale (+controlnet tiles)? What am I missing?

[TiagoSantos81]

Hello @TiagoSantos81 Do you mind sharing your workflow for this? and prompt?and modelname? This image is so much high quality I am impressed. Do you do: text2image + hires -> img2img + ultimate sd upscale (+controlnet tiles)? What am I missing?

Hi! Thank you. I did not know this was that good, since I am mostly testing these technologies, mostly to gain a better developer perspective, potentially for future projects.

This is one of a few first tries with RPGv5, but all other main models I've tested seem to have similar quality, AFAIK.
I just use txt2img, CFG=12 but CFG Scale Fix=7, and the HiRes Fix using Latent (Nearest) Upscaling. To not abuse 1.5 limits I go from Hires resize: from 384x704 to 720x1280. If GitHub does not strip the metadata, you might try downloading the picture and checking it in your own webui image browser.

[vladmandic]

can you check if the issue still persists?

[TiagoSantos81]

can you check if the issue still persists?

Unfortunately, no. Generated just now, after a git pull -r :

The gen info is:

(masterpiece), (extremely intricate), fantasy, (((an on (old witch)))), (((wearing a black witch hat and black clothes))), (perfect composition:1.4), aspect ratio 1:1, beach, deviantart hd, artstation hd, concept art, detailed face and body, award-winning photography, margins, detailed face, professional oil painting by Ed Blinkey, Atey Ghailan, Jeremy Mann, Greg Manchess, Alex Gray, trending on ArtStation, trending on CGSociety, intricate, high detail, sharp focus, dramatic, award winning matte drawing cinematic lighting octane render unreal engine volumetrics dtx
Negative prompt: UnrealisticDream, (BadDream), nsfw
Steps: 20, Seed: 705512687, Sampler: DPM++ 2M SDE Karras, CFG scale: 12, Size: 384x704, Parser: A1111 parser, Model: rpg_v5, Model hash: bb7e29def5, Latent sampler: DPM++ 2M SDE Karras, Image CFG scale: 6, Denoising strength: 0.5, Refiner start: 0.5, Secondary steps: 20, Version: 8db3797, Pipeline: Original, Operations: "txt2img, hires", Dynamic thresholding enabled: True, Mimic scale: 7, Threshold percentile: 100, Hires resize: 720x1280, Hires steps: 20, Hires upscaler: Latent (nearest)

[vladmandic]

i just spent an hour trying to reproduce this. using different batch sizes and batch counts.
plus downloaded same model you've used and used your exact settings.
all-in-all, i cant reproduce at all.

example grid:

[TiagoSantos81]

Mind UnrealisticDream, (BadDream) are embeddings, which might change the overall look, but they should not influence the duplicated head issue.
I am also using xformers with medvram, which might not help in reproducibility. I believe I have disabled all not-preinstalled extensions on the first runs I have reported, but I will try to test that again, also without extensions.
You catch me in a busy work period, but I will try to get back to you, ASAP.

[TiagoSantos81]

I have made a clean git pull -r, and the issue seems to be solved now, although results seem to be somewhat different for other seeds. The other repo, continues to produce the same issue, despite having an effective git pull -r, and all extensions deactivated.
As far as I can tell, the issue is solved.
Thank you for your time looking into this and solving it.

[TiagoSantos81]

Closing this to help with maintenance. Please, reopen if it is still relevant.