Skip to playerSkip to main content
Don't let low-resolution images hold back your AI creations! In this tutorial, I take a deep dive into the "Extras" tab of Automatic1111's Stable Diffusion web UI to show you how to upscale and repair your images effectively.

I compare different upscaling algorithms—including the powerful but heavy LDSR model—and demonstrate how to use GFPGAN and CodeFormer to repair facial details that often get lost or distorted during the upscaling process. Whether you're dealing with blurry source photos or small AI generations, this guide will help you achieve clean, high-resolution results every time.

Original YouTube Tutorial: https://youtu.be/rTIxH4pJv4g

Video Details:
* Original Publish Date: March 29, 2024
* Focus: Image Upscaling / Facial Repair / Extras Tab Workflow
* Test using RTX 4060 TI 16 GB of VRAM

Follow lordcaocao2025 on Dailymotion for more technical AI research and generative workflow guides!

---
Connect with me:
📺 YouTube: https://www.youtube.com/@CaoCao2025
📱 TikTok: https://www.tiktok.com/@caocao20250
💎 Patreon: https://www.patreon.com/cw/Caocao2025

#StableDiffusion #Upscaling #Automatic1111 #AIArt #GFPGAN #CodeFormer #AITutorial #lordcaocao2025
Transcript
00:00Hello guys, welcome back with me, JoeChile2025. Today we're going to learn how to upscale existing
00:08image using automatic 11.11, that's double diffusion. What we need to do is click extras
00:15here, and then you can select batch process which from the inventory are single image.
00:21This time we're going to use single image, simply drag and grab a picture here. After
00:26you do that, select scale by or scale to. This you define the number yourself, but I prefer
00:35this, scale how many times, you know, this is 200, 250. So I'm going to resize this four
00:42times the original size. And then you set up the upscaler. This is the algorithm to upscale
00:49it. There's a bunch of it, and some you're going to download model, especially LDSR,
00:56is 1.9 gigs. So if you click this, it's going to download that. The other, you know, it's
01:03not so much, it's only a few max, some only 19 or 30. So it doesn't matter the other, but
01:11LDSR is quite heavy. Now we're going to do upscale. First we're going to use this length,
01:20because you need to select one of the upscaler. So if you click generate this, it's going
01:26to generate this fast. There are a bunch of upscaler. Sounds really not that great. But like
01:39Ashton 4X, this looks quite good, I guess. But it's give a lot of artifacts. But LDSR, this
01:50is the heaviest models. It usually turn out good. But now you just ignore this. I'm going
01:56to explain it later. Yes, it's finished loading. This look quite good. And there's a bunch of
02:06models too. I think the error is very fast, so I could do it. Ashton and Ashton 4x. And
02:18usually this bad, so let's win AR four times. Yeah. And I'm going to show you all the comparison
02:27from the original. Okay, this is the original image. It's 250. This is the upscale using... Oh, this is
02:37the still the original. This is the length cost. It's not different much, but it's already four
02:43times the size. This is nearest. This also doesn't change much. This is Ashton 4k. It has a different,
02:57but yeah, it's still bad. This is LDSR is the 1.9 gigs models. It's looking good, you know.
03:09It's way better. The image is clear. This is Ashton 4x+. Like I said, this is... This is a small
03:21model,
03:21but the result is pretty good actually. And this is the Ashton 4x+. Anyway.
03:30Not straight lines. And this is Kunet gone. It's not good at all. It's Kunet PNSR. Also doesn't look good.
03:42And Swinier for Axe. It looks pretty good too. But in this case,
03:52LDSR I think is the best. It's huge models, but it's worth it.
04:01Next we're going to image a character with face.
04:06Okay. This is the picture that I use. Same.
04:11Using lenses and the other. I'm going to show you the result of all the upscale image that we create.
04:19Okay. It's 223, 24G image. This is the length cost. Like usual, it's a really bad mirror. It's also
04:27as bad. I guess it's almost nothing different. So it's always fine to do that.
04:33Ashton 4k. It's... In this case, it's not working.
04:42LDSR. It's look good, but the face is still bad. Ashton 4x. I think it's good, but there's this thing.
04:55Ashton 4x anime. It's also good, this thing, but you know, all the face is bad.
05:04So what should we do if we meet a picture like this?
05:10You know, everything is bad at the face. That's what we're going to do.
05:17Use the... We're not going to use upscaler 2, because it's just add upscaler, and you check the
05:24visibility, but I don't think that's necessary. We're going to use cfgon and codemoner feasibility.
05:31Cfgons, it's like algorithm to repair the face, and so is codemoner.
05:39Codemoner feasibility, you need to increase this.
05:43Now, for example, to 0.7. And this is the weight. The more zero it is, the more good the
05:52image,
05:52but the more it's far from the original image. You know, it's like the more it is, the more good,
06:00but the image is going to be far from the original. But here, you know, it's not going to have
06:06a change at all,
06:08because it's been a matter of fact.
06:13So yeah, for example, we're going to do this with lancos.
06:21Let's see, it has the image better already. And, you know, if you make sure that you have the
06:30algorithm set up for the better plooch or the better upscaler, and then repair the face,
06:37because the face is good, but the other is ruined. So yeah, that's that. And to fast forward,
06:45I used the codeformer feasibility using the AFGON 4.0, codeformer 7, and the other one, the width also 7.
07:00This is the result using the ANIM 4X.
07:05The AFGON ANIM, the one that is here. AFGON 4X ANIM 6P.
07:14Looking good, I guess. We could repair this in a painting.
07:21Or the LDSR. I don't know why the cloth is ruined,
07:29but it is what it is. I think this is the best thing in this case.
07:38Okay, next. Okay, next we have a 72512768 image.
07:47And I manually reduce its size to a quarter of its original size, so it's very small.
07:57And now I'm going to upscale it and see the ability to have a very similar
08:02color with the original.
08:06We'll look the same with the original. Okay, this is Lanchos. I don't think we're gonna
08:12do Lanchos or Nears again, because that's really bad.
08:15The AFGON 4K is actually pretty good already. You know, it looks the same.
08:23But the face, the face is quite good, but you know, it's still blurry a little.
08:31LDSR is also good, but the face is also blurry. Okay, in this case, I'm not used CFGON or Codeformer.
08:39Okay. Asking for KX, the picture is also bad. And the offscale is also bad.
08:48Anyway, now it looks like a painting, so it's really bad for realistic image.
08:57And Sweeney R4K, it's also, the face also looks bad, but the body is okay.
09:03So in this case, we're also going to use the Codeformer and CFGON here. In this case,
09:12I use the LDSR and CFGON. It looks almost the same, but you know, the eye here is not working.
09:25I guess you could add 5 after, but for me, you know, from this
09:32original here to here, I guess it's really, really great, you know.
09:39And next time, we're going to add details and stuff using image to image. But for now, this is
09:47what I can share. Plus the guy, I generate now a picture that are
10:01five, one, two, seven, six, eight. I rescale it or upscale it to twice its original size.
10:10So you see it's 161%, it's 82%. And the result is quite good, I guess, if you see it from
10:19here.
10:19But if you zoom out, zoom in, I guess.
10:27It's a 100%, so it's still bad. But, you know, it's, it's, it's quite
10:35has this little thing using the Airgun 4K X.
10:43And if we use LDSR, it's quite clear.
10:51And if you use AirS gun for X plus, it's also pretty good.
11:03And for X anime, it makes the face look like anime.
11:08Make it like a painting, I guess. Swin 4K.
11:13It's also pretty good, I guess. Swin IR 4X, I mean.
11:21And this is if you have codeformer and stuff. I see Airgun 4X.
11:34The eyes look good too, since we repair it with codeformer.
11:40And the face also, we can do it. So yeah, even if you resize a picture that are already clear,
11:50using codeformer. And as I've gone, if you resize the character, it's still important.
11:55Because, see, the result is, it's pretty great, you know, instead of MTI and stuff.
12:07So anyway, guys, I think that's it for today's video.
12:12Thank you for watching. Don't forget to like, share, and comment.
12:15And see you again on the next episode. Have a nice day.
12:19Bye-bye.
12:19Bye-bye.
12:22Bye-bye.
Comments

Recommended