Stable Diffusion - Better Images with FreeU - Fooocus

Video Statistics and Information

Video
Captions Word Cloud
Reddit Comments
Captions
Hey everyone, Rodney here at Kleebz Tech,  and today we're going to be talking about   FreeU in Fooocus. Free what, you ask? Well,  FreeU: Free Lunch and Diffusion U-Net. Yeah,   I know, that didn't really explain anything.  Now, this isn't specific to Fooocus only,   but I have had a few people ask about the  settings in Fooocus. Where are these settings,   you ask? Well, they happen to be hidden in  the debug settings, and I did some research   and experimenting to hopefully give you some  information about them. I'm not going to be giving   a super detailed explanation, since I don't have  a complete understanding myself, and I couldn't   find a lot of detailed information online. But  I will show you some examples of what it can do. Now, the claim is that FreeU can be used to  improve image quality without extra training.   As for it automatically improving image quality,  it's hard to say yes or no. I found for some   things it did, and others, it did not. But the  biggest thing I discovered is there's no correct   setting. I did some digging online, came up with  some recommended settings, and went ahead and   tried those out. I would love to go through  every possible setting to show comparisons,   but that would be pretty time-consuming, and I  don't think it would really show much. As with   many of these things, what looks best  for one person might not for another. So, where do we find the FreeU  settings in Fooocus? Well,   you're going to go into the Advanced  tab, and then the Developer/Debug mode,   and then you're going to find the FreeU  tab in there. And when you want to use it,   make sure you enable it. If you just  move the sliders, it won't do anything. Now, as for what those do: the B value is the  backbone, the S is for the skip connections. So,   the B value modifies mostly the core part of  the image, the backbone, and focuses on the main   structure of the image, and the S value emphasizes  more on the finer details and the textures. Now, as for which recommended settings to  use, I did some digging online. I found a few   different ones from the GitHub page and some  other sources and went with those. For SDXL,   and showing some of my examples, you're going to  find a lot of it is all subjective on whether you   think those are best or not. For some images,  I found they worked well; for other images,   not so well. And I think these two images will  give you an idea. So, in the top left was my   initial image, starting with everything set as  one, and then I slowly lowered the S values to   get down to the last one. It's pretty noticeable  how it reduces the textures and the details in   the image by lowering that S value. Similarly,  with this image, I went the opposite direction,   and I left the B values at one and increased  them over that. Now, I haven't found many that   suggested doing that, but I wanted to see  what would happen. And as it went through,   you can see it added textures and details,  makes the image look pretty bad once you get   to a certain point. So, that gives you a little  bit of an idea of what those values might do. Okay, so now, to start to look at the images that  I ran through just for comparison. I won't spend   a bunch of time on each one. You can make your  decision, pause it, look at it if you want to. So,   the center one is without the FreeU. So for  this one, I actually think that one's better,   in my opinion. But again, it really depends  on what you're trying to get out of the image.   I've been experimenting with it, and what I  usually do is generate an image and, if I like   the image for the most part but want to tweak  it, I'll go into the FreeU settings to try to   either increase or lower details. For our mushroom  scene with the giant glowing mushrooms, the same   idea applies. Definitely, the center image has  more detail in it and looks more realistic.   You can see with some of the other images, the  lower details and textures, and if that's what   you're trying to achieve, then adjusting the  FreeU settings might be what you want to do. Now, for our wizard, it's a similar idea. As  with any of these, it really comes down to the   style and look. It's almost like a different  style for some of these images on how they   come out. Sometimes it actually improves  the image; I personally like the bottom   right one myself. This demonstrates  how, sometimes with realistic images,   you can end up with oversaturation, especially  on the bottom left or even a little on the top   left. Often, it can be just a balance  of slightly adjusting those numbers. For the woman eating a burger, the  same idea. Some of these definitely   don't look that realistic. So, if you're  looking for a more photorealistic one,   the FreeU disabled on this one, I feel,  is probably better, or even the top right. Now, for our tranquil forest at sunrise, I'm a  big fan of the one without FreeU on this one,   but again, that's if I'm going for  more photorealistic. Whereas these   other ones definitely have  a different style to them. For our still life of assorted fruits,   I think that almost all of the FreeU ones  are better than the original one. So,   this example shows that in some images,  it definitely can make a big improvement. The astronaut one here, I'm not seeing much of  a difference between all of them, so overall, I   couldn't say whether that would make a difference.  I probably wouldn't mess around with it after   trying a few different settings; I'd realize that  maybe I'm not going to get anything different. Our Viking ship, the same idea. The  center one might look more realistic,   but I actually like some of the other ones.  So, I think here it demonstrates what you   can get for different looks from an  image by adjusting those settings. A lot of people, a lot of stuff I read  online, said that anime was better versus   the photorealistic ones, which may be  the case, especially depending on how   much detail you want to add or the textures and  things like that. I'm not huge on anime, so I   can't really say a lot of comparisons, whether  I think one of these is better than the other. And similar with this one, a lot of these, it's  hard for me to say, you know, you can see the   differences, but I don't necessarily know  which is better or not with a lot of these. Now, for this example, I went with a  more photorealistic one. I liked most   of the results of these. The bottom  left one here, I'm not a big fan of;   I think it looks too fake, but most of the other  ones, I think, actually came out pretty well. So, what's the best way to use FreeU? What I've  been doing is once I generate images, if I want   to play around with it and I get an image that's  pretty close to what I want, a lot of times I'll   go in, I'll enable FreeU. Now, I want to get the  same image, of course, so I'm going to go in here,   uncheck 'random', and I'll set my seed the same.  So, now I have my two images, and I'm going to   adjust it on these. At this point, I would  just go ahead, enter in the changes I want,   and generate based on that seed, so I can  adjust these to see how it impacts the image. Now, I'm also going to change this down  to just one. I'm just going to go with   this first image and have it generate that,  and then I can go ahead and compare them and   see if I like the results. If I don't like the  results, then I can continue tweaking further. Now, the one on the left here is the one that  I just adjusted versus the one on the right,   and then I can continue adjusting from  there, depending on what I want to do. So that's pretty much how I've been playing  around with it and testing it. The problem   is there's no real guaranteed settings, but  I think if you play around with it enough,   you get a feeling for what each one will do and  how it may impact the image that you're working   on. So, I would suggest keep playing around with  it and discovering what you can do with that. I do plan on using it a few times on images  that came out pretty close to what I wanted,   I tweaked it around a little  bit, but I will admit overall,   I don't use it a lot yet.  So, we'll see how that goes. I think that covers what I've learned so far with  FreeU and what I wanted to pass along. Hopefully,   it gives you an idea of what you can do with it  and maybe start getting better images. Sadly,   I don't have all the answers when it  comes to FreeU because I haven't had   enough time to really experiment enough  with it. I have not made up my mind on   if it makes better images or not. I  think it just makes them different. If you found the video helpful, please  consider hitting the like button,   and don't forget, keep on creating.
Info
Channel: Kleebz Tech AI
Views: 7,282
Rating: undefined out of 5
Keywords: stable diffusion tutorial, stable diffusion, stable diffusion secrets, fooocus tutorial, stable diffusion install, foocus, stable diffusion ai, stable diffusion xl, freeu stable diffusion, freeu free lunch, freeu free lunch in diffusion unet, better images stable diffusion, how to get better stable diffusion images, freeu, fooocus settings, stable diffusion video
Id: Vao6QnBI_jQ
Channel Id: undefined
Length: 8min 35sec (515 seconds)
Published: Tue Apr 02 2024
Related Videos
Note
Please note that this website is currently a work in progress! Lots of interesting data and statistics to come.