StableDiffusion Deep Dive - HiRes Fix - How to avoid twinning and losing composition
Vložit
- čas přidán 24. 01. 2023
- This video explains what the HiRes Fix in the Stable Diffusion Automatic1111 Web GUI is useful for, describes how the feature works, then identifies which settings are important and what the appropriate range for these settings are to ensure you can consistently generate good images.
Overview of Hires Fix - 01:05
Methodology - 01:40
Steps and Samplers - 02:13
Denoising - 03:40
Upscaling - 04:29
Upscalers - 04:50
Recommendations - 06:40
Settings Chart - 7:10
#aiart #stablediffusion #upscaling #automatic1111
Link to Google spreadsheet with the figures shown in this video as well as some useful prompt constructors. It is read only, so make a copy for yourself if you want to use the constructors.
docs.google.com/spreadsheets/...
Thank you so much! This solves a lot of questions I had, mainly "what upscaler should I use and when?". This has also let me understand a bit more of how the whole upscaling process works. Great video!
I was looking for this guide for so long. forgot it was a video and not an article so google searches got me stuck. This is the best deep dive around.
great info, love your charts and summaries of recommendations. Really enjoyed your upscaling video a month ago so now going back through most of your videos, good stuff.
Thanks for the quick and clean information, this video every second counts.
This tool really is a puzzle to use. It's time consuming but in a good way.
all your videos are amazing !!
Really well made and helpful. Thank you so much!
Thanks! I'm glad you enjoyed
Thank you for this video, very good information!
Very helpful video, thanks!!
Very clear, thank you very much. I just stepped up my SDXL game.
Great info, thank you very much. Subbed :)
thank you very much for the great explanation, it helped me a lot!
Really useful info!
Thank you very much, really helpful.
Thanks A LOT for the video!
I was really curious about that subject and my graphics card is way too weak for me to make a complete analysis like that myself.
Thank you
A follow-up on your hands testing might be worthwhile to see impact that hi-res fix can have on improving hands and faces.
I came to similar conclusions with my own tips research and noob testing, as i have a 2070s im not going to make a big xyz grid so this video is good to see it all laid out. Wish i saw this earlier, would have saved me loads of time, but hey the experimenting was good for learning a1111, id still say im a noob though.
From my experience the latent upscalers are better in adding more details in the initial phase of image generation
Tried it out and it works, but when I go below 0.5 denoising it starts to look bad many details are lost. I tested 0.5-0.66 and results are decent, on 0.70+ it starts to generate weird horror stuff. Good tip thanks.
Did u manually generate the images in the spreadsheet or is there an extention to selec all samples?
2:43 you would recommend keeping hires steps "glow at 30"?
I don't have the highres fix checkbox? Has this been removed in newer versions?
a year later and i still have that little upscale vs denoising chart as a jpg on my desktop for quick review
Any insight or help with images made in defourm? Colors are never consistent as more images are made. I have color coherence set to LAB. I don’t know what to do. Cfg 7
hi, how many hires steps do i need to use to upscale by 2x? i keep getting errors
Great video! It took me 1 hour to process this 8 minutes video. Weed doesn't help, but man that was a lot of amazing information. Thanks!
When i only apply hi rez fix its emmidiately changes original picture to something elce even with 0 denoizing.
mine creates a double image, meaning it creates in original than at 50% it breaks off and creates another image, sometimes with two people and never looking as good as the original, the second image gets the upscale and not the first.
I know this video is 1 year old, but I wonder of through a similar technique can you achieve similar results as Korea or Magnific. Have you tried something like that?
i keep getting this when trying to do hires steps: RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 128 but got size 64 for tensor number 1 in the list.
How do you manage to generate one face, and not two, or three in one?
There are so many tutorials and theories, but I just don’t understand this point.
Good video but the sound volume is very low for me.
and what about hires fixes like ESRGAN? are those the non-latent ones?
Yes, those are the non-latent ones.
Hello, perhaps you can help me, English is not my first language and although I understand most of it, concepts like "converging" & "fractualization" are hard for me to understand. What exactly do you mean, in laymens therms when you say that some of the samplers are converging ? I think I understand the fractualization part, it's just making some weird repeating patterns , right ?
Convergence means that the image approaches a "final form" that won't change with more steps, so high step counts just waste time.
The non-converging samplers add randomness into their input, so the image keeps changing forever.
A fractal is a geometric pattern that continues as you zoom in/out on the object. For a famous example, see the Mandelbrot set.
Here, it basically just means that a single subject starts getting split into multiple smaller subjects
Can someone tell me why I would use Hires fix instead of upscaling in the "Extras" menu?
...
The audio level is too low ! Its very hard to listen this on a cellphone. Thanks any way, nice video.
Hi, do you know how i can upscale above x2? Everytime i try to create some images and want them to upscale to more than x1.9 times the resolution, i get out of vram errors.
Base resolution 768x768px, Euler a, 30 Steps, Hires Steps 10 on RTX 3090
i upscale 1280x720 to 2560x1440 with hi-res-fix and I only have a 3060ti 8GB lol
@@Flashdog97 i think i found the problem, didn't set --no-half-vae in webui-user.bat after that i get an output for 95% of my images, before that, around 40% where just black (testet with RTX 2060s, 3070 Mobile and 3090)
@@cosciblog also add --xformers to the bat. sped up my game by 3-times
Hires fix can also have a different Sampler, different pos./neg. prompt so much left unsaid. : )
can use a different prompt for the Hires. fix ?
I found the option in settings/User interface : Hires fix: show hires prompt and negative prompt (requires restart)
dang thats good to know!
Does this allow you to generate at higher res than normal? my 6GB vram only allows up to 768 x 768.
I'm not sure whether hires fix reduces the amount of RAM needed for larger images, but I would guess it doesn't.
The SD upscaler script in img2img might help since it breaks down larger images into tiles, but I haven't verified that either.
@@siliconthaumaturgy7593 i tested it (3 times already) on my 3050ti laptop gpu with 4gb vram and my resolution was 552x504 (sampler: dpm++ SDE Karras | Steps: 150) and upscaled it by 2 (R-ESRGAN 4x+ Anime6B | highres steps: 30 | denoise 0.33) and it worked just fine .. not sure if thats important, but it said smth about "Tile x/9" in my console. Also: i start automatic1111 with "--medvram" and "--xformers" (not sure about xformers, might delete it again, but medvram will stay :D)
how do you get a batch with labels like that sampler/steps, sampler/noise etc
There is a drop down near the bottom of the image generation called scripts. Select "X/Y" plot from the dropdown. My video on Prompt S/R kind of covers it (since Prompt S/R is used within in the X/Y plot)
@@siliconthaumaturgy7593 And there is a green book button on the right of the values box that lets you add all the existing, non-numerical values such as the sampler names, so you don't have to type them yourself, and you can delete the ones you don't want to.
why when I set Denoising below 0.75, everything get blur
I assume you are using a Latent upscaler (any upscaler with Latent in the name). I would use a non-Latent upscaler to see if that improves things
Cloning happens on almost all my upscales. That is an image of a man AND a woman, or a character and a croud in the background. In these cases almost all my upscales make all the faces the same.
7:16
Your sound is too low.
Could you please turn the volume down more, that way I can officially blow my ears out on the next video instead of just going def.... thanks...
Learn to encode audio.
Hi, you're doing a good service here. I don't mean to be rude but might I suggest maybe dialing back some of the technical stuff. I think you would get a lot more viewership if you just kept it really simple with examples of your experiments and a simple synopsis of recommended settings and reasons for those settings. Not trying to be a jerk just giving some unsolicited practical advice. The name of the channel is a bit complex as well, don't know if u plan on doing mostly AI content but it's hard to remember and hard to search. Again no disrespect just some things to consider. Thank you for this information.
I say keep it techie and people just need to catch up to you, siliconthingamajiggy. There is no 'one set of params fits all'.
I do understand that a lot of people just want the final settings instead of the technical details, so I have a comment in the video description of where to go if you just want to see that. For this one, it starts around 7 minutes.
I disagree. In the age of AI when everyone is slamming faceless accounts with the same redundant content everyone is pulling from chatbots, I'm glad he's pumping out technical content with actual value.
i came here to watch acutaly someone press the button and say what it does, instead i saw alot of pictures, dissapointed
you video is too long man. You need to divide it into two parts. in a minute here i what matters: 1. 2 . 3.. then you can go into depth of each part.