Skipping last few iterations

[FartyPants]

Some checkpoints, especially over trained would butcher the last few steps. I created script using store_latent that will save the iteration at certain step.
So for example I set steps 20 and the image is pretty much overcooked at step 20, but saved with the script at steps 15 it is perfectly fine.
Now saving the image early is not the same as setting steps lower, so if I set steps to 15, the image will be still ruined.

Is there any way to skip the last few steps (plugin/script/settings) just as I described?
So if I set steps 20, I want to actually terminate it at step 15.
Or is there some other way this is handled?

The script to save at certain step works, but it works only for txt2img since I just call sd_samplers.sample_to_image and images.save_image at certain step as per the wiki suggestion.
This way it doesn't work for inpainting obviously because it would save current iteration that is a cropped image, not the final assembled image.

[FartyPants]

I figured it out and wrote a script that will bail earlier (you set sampling steps ex. 20 and the system is going for that, but then it will bail at the number set by script f. ex 16 - which seems to produce good results for overtrained ckpts) but I'd still need some help to polish it. I'd like to talk to some developers before committing this anywhere)

If you are interested, let me know and I'll post it on github after some clean up.

If not - don't let me know.

[Inosss]

same. I tryed inkpunk diffusion model and very often a see so perfect generation in preview below last sampler steps, but dont know how to get it :(

[JD1234JD1234]

i think its something to do with automatic1111 code/update

[f-rank]

I am very interested in this.

[CurtisDS]

i noticed this issue after updating a few weeks ago (when they added the extra networks tab). I think it might have to do with live previews though. The issue seemed to go away when I turned live previews off. Although... now that I cant see the image as its being created, maybe I just dont realize that its still messing up the image in the final step.

[evilblade666]

I'm extremely interested! Any chance of a link to this script?..

[Nacurutu]

Clip Skip
This is a slider in settings, and it controls how early the processing of prompt by CLIP network should be stopped.

A more detailed explanation:

CLIP is a very advanced neural network that transforms your prompt text into a numerical representation. Neural networks work very well with this numerical representation and that's why devs of SD chose CLIP as one of 3 models involved in stable diffusion's method of producing images. As CLIP is a neural network, it means that it has a lot of layers. Your prompt is digitized in a simple way, and then fed through layers. You get numerical representation of the prompt after the 1st layer, you feed that into the second layer, you feed the result of that into third, etc, until you get to the last layer, and that's the output of CLIP that is used in stable diffusion. This is the slider value of 1. But you can stop early, and use the output of the next to last layer - that's slider value of 2. The earlier you stop, the less layers of neural network have worked on the prompt.

Some models were trained with this kind of tweak, so setting this value helps produce better results on those models.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#clip-skip

[JD1234JD1234]

this didnt fix it

[FartyPants]

CLIP skip has nothing to do with skipping iteration steps.

[Nacurutu]

"Saving steps of the sampling process"

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#saving-steps-of-the-sampling-process

[JD1234JD1234]

is there a way to edit the script so it saves the 'second to last' frame? instead of saving all of them, also when i do 20 steps it saves 7 images?

[Nacurutu]

is there a way to edit the script so it saves the 'second to last' frame? instead of saving all of them, also when i do 20 steps it saves 7 images?

I don't know how to code, maybe @vladmandic or @Gerschel can help you with that.

[ClashSAN]

@JD1234JD1234 @FartyPants use
https://github.com/AlUlkesh/sd_save_intermediate_images extension

Its mean to be an upgrade/update of that custom script

[vladmandic]

just a quick note, save at steps - x is not the same as final_step - x as many schedulers bail out when image generation is considered complete regardless of current steps (e.g. if you set steps to a high number, most likely it will never get reached). its not the case with all samplers as some are always-incremental. so to actual save at final_step - x, you'd need to buffer final x steps in run tasks and save in postprocessing task once completed.

[FartyPants]

That's too complicated for me. The version I made works well for my purpose. I took the scripting example in wiki and modify it and came up with this hacked solution. I have no idea how the whole A1111 works and pretty much very limited python knowledge. I'm not even skipping the steps. I couldn't figure out how to bail out the process so I'm just storing the image at a certain step then sending it to interface/saving it. Took me a good 30 minutes to figure out how to overlay the output image when inpainting as that's the problem with prior save at step solution

[Gerschel]

@FartyPants I think you can do it, think of the buffer as a disposable list. You'll learn more if you try and fail. Keep a list of references "n" wide, push and pop to keep the length even, and on completion pull those from memory, display all, allow you to select the better one for save and clear the list.

[pablonaj]

Hey thanks for this, it's a nice script and works for what you wanted. Do you think it is very hard to actually skip the steps instead of only saving the X frame you are after? I think this would be super useful as a preview for seeds, you could make 100 images and stop at step 10 then chose the best ones to actually fully render.

[FartyPants]

Hey thanks for this, it's a nice script and works for what you wanted. Do you think it is very hard to actually skip the steps instead of only saving the X frame you are after? I think this would be super useful as a preview for seeds, you could make 100 images and stop at step 10 then chose the best ones to actually fully render.

Yeah, that's how I wanted it in first place. I did find how to interrupt the process, that was easy - but then I'd have to copy more of the process_images code from main app to the script so the image can be finished nicely, stuff like color correction and this and that that I'm not always sure what it does. We will see.
This is really just hacking a code that was written by many people without documentation and it is hacking by someone who doesn't have slightest idea how it works. So not very ideal. :) Not to mention that my knowledge of python is abysmal.

[FartyPants]

@Gerschel Thank you for believing in me, but I first need to actually know what I'm doing - which is only true on an extremely superficial level. But as you said: "You'll learn more if you try and fail"
And yeah, those suggestions would be actually awesome.

[FartyPants]

I updated it from some of the ideas above - so now it shows multiple last steps in the interface and can also fully bail out - and I learned a bit more of python

I'm going to rewrite it using @vladmandic suggestion, finally I actually understand what he meant and of course he is 100% right

[Eleksar387]

Your repo does not exist anymore. Has it moved? I really liked it!

[PhreakHeaven]

Your repo does not exist anymore. Has it moved? I really liked it!

Yea...what the heck? Came to find out how to fix the "store_latent" error and see there WAS a new version, but now it's all just gone... :(

[PhreakHeaven]

I updated it from some of the ideas above - so now it shows multiple last steps in the interface and can also fully bail out - and I learned a bit more of python

I'm going to rewrite it using @vladmandic suggestion, finally I actually understand what he meant and of course he is 100% right

So...what's going on with this script now? It worked great, but is now broken (first version) and it disappeared from your repos...? Tried to get the updated version when it stopped functioning (store_latent error), but cannot find it anywhere. Please help!

[ghostsquad]

I too am looking for this. Is there a new extension or something that does this?

[HerrGitHuber]

For everyone that is still looking for a fix >>>
There is an Extension called "Anti Burn" https://github.com/klimaleksus/stable-diffusion-webui-anti-burn.git

It does the trick!

(can also be installed directly from the "Available" List in the Extensions Tab)