Give Them Custom Shirts! No More Garbage Text - Stable Diffusion

00:18:25
https://www.youtube.com/watch?v=MnGYEB_oOqE

摘要

TLDRIn this video, Rodney from Kleebz Tech demonstrates how to generate accurate text on t-shirts in images, using various techniques and tools. He explains the importance of configuring settings in Fooocus, particularly focusing on the in-painting and text prompt areas. Rodney shares tips for selecting fonts, adjusting the denoising strength, and managing text size to ensure better clarity. He also discusses using Magnific AI for upscaling images and GIMP for additional editing adjustments. Throughout the tutorial, Rodney emphasizes the need for practice and experimentation, especially with more challenging images, and encourages viewers to share their own methods in the comments.

心得

  • 🛠️ Foocus is great for editing t-shirt images with text.
  • 🔤 Use bold fonts for better text clarity.
  • ⚙️ Adjust denoising strength for realistic results.
  • 🖼️ Magnific AI can enhance image quality post-edit.
  • 📸 Experiment with background colors for text.
  • 📏 Larger text works better for visibility.
  • 🎨 Use GIMP for advanced image layering.
  • 👕 Techniques are applicable beyond t-shirts.
  • 💻 Practice improves results over time.
  • 👍 Share your own methods to enhance community learning.

时间轴

  • 00:00:00 - 00:05:00

    Rodney introduces a tutorial on generating accurate text on t-shirts in images, acknowledging that while it's not perfect, decent results can often be achieved. He emphasizes the importance of image quality and preparation before diving into the process and mentions prior videos for context.

  • 00:05:00 - 00:10:00

    In the second part, Rodney details his setup with Fooocus, adjusting various settings for in-painting and image prompts. He stresses the importance of proper text prompts, image ratios, and the in-painting denoising strength. He indicates the need to find a balance in settings to achieve realistic effects on the t-shirt designs.

  • 00:10:00 - 00:18:25

    Rodney demonstrates the generation of images while providing insights on refining denoising strength and prompt adjustments. He also highlights actionable tips for working with logos and various styles. The segment ends with considerations for upscaling images and experimenting with different software tools, emphasizing the necessity for tweaking to achieve optimal results.

思维导图

视频问答

  • What software is recommended for generating text on t-shirts?

    Rodney recommends using Fooocus and mentions Magnific AI for upscaling images.

  • What should I consider when choosing text fonts?

    A thicker, bolder font is preferred for better results.

  • How can I improve the text generation quality?

    Adjust the denoising strength and ensure the text prompt is included in the generation settings.

  • Is it possible to use this technique for designs other than t-shirts?

    Yes, the techniques can be applied to various designs beyond t-shirts.

  • What are some common challenges in generating text on t-shirts?

    Text can appear too small or lack clarity; experimentation is needed for different images.

查看更多视频摘要

即时访问由人工智能支持的免费 YouTube 视频摘要!
字幕
en
自动滚动:
  • 00:00:00
    Hey everyone, Rodney here at Kleebz Tech, and  today I'm going to show you how to get accurate
  • 00:00:03
    text on t-shirts on people in your images, like  these. Now, this is not something that works
  • 00:00:07
    perfectly, but it does often do a pretty decent  job. It depends on the image that you're working
  • 00:00:13
    with, and I'm going to make things difficult on  myself by not using easy images to do this. So,
  • 00:00:19
    we're going to try to generate some stuff like  this. I do recommend watching my video I have
  • 00:00:24
    on text already that this is building  upon. This is a little more difficult
  • 00:00:30
    than that, which works very well.  This one's a challenge sometimes,
  • 00:00:33
    but I'm going to give you tips on how you  can try to get better results from it.
  • 00:00:37
    So, why don't we just start and get right  into it? There's no perfect way of doing this,
  • 00:00:44
    and it doesn't always work well with every image  generated. I'm going to show you the system of
  • 00:00:50
    how I do it. Hopefully, these come out well. I've  gotten some to come out well, and others can be
  • 00:00:56
    a challenge. The idea behind this is trying to  generate t-shirts with words on them, sayings,
  • 00:01:02
    or designs with words and getting accurate text.  This is purely about text, but it can also be used
  • 00:01:09
    for other stuff as well. One of the issues is,  a lot of times, if you try to get what you want,
  • 00:01:15
    it's not going to work. If you use in-painting, it  looks copied and pasted. There are other ways of
  • 00:01:20
    doing this sort of stuff. For example, in this one  I did in [__], which I can show at the end of this
  • 00:01:26
    video real quick how I did this. This is useful  for doing it that way if you have a specific
  • 00:01:32
    design you're trying to get on there. It doesn't  always work perfectly, but it is an option.
  • 00:01:37
    Let's say you want to generate one, and you want  to have the text and stuff like that to actually
  • 00:01:43
    work and spell the proper words. Well, that's  what I'm going to show you today. For this one,
  • 00:01:48
    let's go ahead and just start off  with one of these images, but first,
  • 00:01:51
    I'm going to set everything up. I'm using Fooocus;  you should be able to do this in pretty much any
  • 00:01:55
    of the UIs. There's a little bit different on how  they work. I do like the main engine for Fooocus,
  • 00:02:01
    and I personally think it's the best, even if  you just use it for in-painting. We're going to
  • 00:02:06
    check off the input image tab down at the bottom.  We're going to be mostly working in the in-paint,
  • 00:02:10
    out-paint, and the image prompt area down  here. I'm going to check off Advanced as well,
  • 00:02:15
    the advanced tab. We're going to check  off the developer debug mode, and in here,
  • 00:02:19
    we're going to go into the control tab.  We need mixing image prompt and in-paint;
  • 00:02:23
    we need to check that off. Then we're just going  to click on the in-paint tab and have that open.
  • 00:02:27
    A lot of times, if you were just to go ahead  and try to use the image prompts with the
  • 00:02:32
    text that I'm going to show, it's not going to  work properly; things won't line up. But I'll
  • 00:02:37
    show you how to get that to work. If you have  seen my other video on how to get good text,
  • 00:02:41
    this is basically the same idea with some  slight modifications. The basic concept is
  • 00:02:46
    still the same. If you haven't seen that,  I do suggest looking at that one. It works
  • 00:02:52
    well on flat surfaces for signs or anything  like that. The shirts are a little harder,
  • 00:02:57
    as I said. It doesn't always work for every shirt,  doesn't always work for every design. It takes a
  • 00:03:01
    little bit of practice, but once you get used to  it, you can get halfway decent results. They may
  • 00:03:06
    not always look perfect, though; I will warn  you of that. And if you see any flickering or
  • 00:03:11
    anything while I'm doing this, my video card,  I think, is about to die, so hopefully not.
  • 00:03:17
    I'd like to be able to continue recording  videos and not have to put that on pause.
  • 00:03:20
    So, let's start off with this image down here.  Yeah, we'll do this one. Drag it down, then in
  • 00:03:30
    the in-painting area, I'm going to select a larger  mask. If you're not familiar with in-painting,
  • 00:03:36
    I do suggest watching my in-painting tutorial.  Now, what we need to do at this point is we're
  • 00:03:44
    not going to generate anything. It doesn't really  matter; we can get rid of some of these things in
  • 00:03:48
    the prompt. You want to put your text in the  quotes. Now, we're going to be using something
  • 00:03:53
    different for this, but we are going to just  prepare the prompt at the moment. The in-paint
  • 00:03:58
    respective field, that's what we want to pay  attention to first right here. Now, if we go
  • 00:04:02
    all the way to one, when it goes to in-paint, it's  going to use the whole image. It won't upscale,
  • 00:04:07
    won't do anything, but what happens is, at that  point, we're going to be putting some text into
  • 00:04:11
    the image prompt, and when we do that, it's  going to have a hard time keeping that text
  • 00:04:17
    together because it's going to be so small at  that point. So, we need to have the image zoom in,
  • 00:04:23
    and the way to do that is to slide this  down. Normally, it's around six or so,
  • 00:04:28
    but we actually are going to slide this not all  the way down because then it doesn't take any of
  • 00:04:32
    the image around it into effect. And we want  that because it's useful for reference. So,
  • 00:04:39
    I'm going to bring this down a little bit,  probably to about 0.2. We'll give that a shot.
  • 00:04:43
    Now, the first time we generate, we're actually  not going to generate anything. I just need to
  • 00:04:47
    get some numbers, and if you've watched my other  video on text, you'll know why I'm doing this.
  • 00:04:52
    But at this point, once we go ahead and hit  the generate button, we're going to want the
  • 00:04:57
    command window open. Now, my system, when I  go to do this in upscale while I'm recording,
  • 00:05:04
    can have issues. So, I'm going to try to leave  this open to show you exactly what happens. So,
  • 00:05:08
    let's hit generate. Now I'll minimize my window  here and we'll see if it actually does this.
  • 00:05:17
    There's one number you want to pay attention to  here, that's the 425x425, that's the section of
  • 00:05:22
    the image that it's going to be working on.  Now, the other thing I do here as well is,
  • 00:05:28
    for reference, I'm going to take a screenshot  of this. That's just so I have a rough idea of
  • 00:05:39
    what it's working on. So, I'm going to save that.  Now that we have that, we can work on getting our
  • 00:05:47
    text. We don't need this to finish, but I'm  going to let this finish because I do want
  • 00:05:51
    to show something. One of the things to keep in  mind when doing this is, the shirt doesn't look
  • 00:06:01
    right. We're wanting to get more of a faded  look, a worn look. So, this doesn't always
  • 00:06:08
    work well when you try to do the in-painting on  the shirt. I'll show you how to get around that.
  • 00:06:15
    Now we want to get our text. I've already got one  prepared here, but I'm going to change the size
  • 00:06:20
    because I don't have the proper size because I  didn't know. So, we're going to put 425. I'm going
  • 00:06:25
    to resize this. The other thing that I do is I'll  bring in that image, and that's close enough for
  • 00:06:35
    what we're doing here. I've got my text where I  want it. You can use any tool; I do this in Adobe
  • 00:06:41
    Express just because I like the text effects.  Working with text in here works very well. The
  • 00:06:49
    other thing to keep in mind is when you're working  with the text, the font, it's usually best to have
  • 00:06:54
    a thicker, bolder font, and it won't always keep  the same font. We'll go with this one; we're going
  • 00:07:01
    to leave it there. And the background color,  I find you don't want to go with transparent;
  • 00:07:07
    that doesn't work well. I'm still trying to figure  out which colors work the best. I tend to go with
  • 00:07:15
    the same background color. Then we're going to  go ahead and download and save that image. We're
  • 00:07:20
    going to go to our image prompt tab, and then I'm  going to bring in that text that I just created.
  • 00:07:26
    I'm going to put this on CPDS. Pyrocanic can  work as well. Play around to determine the stop
  • 00:07:33
    value, which is when it stops influencing it.  You want that usually pretty high, otherwise,
  • 00:07:38
    the text can start to fall apart. But the higher  you go, it does impact the rest of the generation.
  • 00:07:45
    Now, the other thing is, if you want to do  anything else, like I'm going to drop this image
  • 00:07:50
    in here just as a reference. For the robot, there  we go. Now I have everything set. I have my prompt
  • 00:07:59
    set, I do have the text in the prompt and here, so  that helps influence each other. If you don't put
  • 00:08:06
    the text in the prompt, I find it doesn't work as  well. Now we're not going to touch this in-paint
  • 00:08:11
    respective field at this point because we've  already set this, and if we were to change that,
  • 00:08:15
    it can mess up the image prompt that we're  using. The other thing we need to do is the
  • 00:08:22
    main denoising strength. One of the reasons this  looks kind of fake is because it regenerates
  • 00:08:29
    that whole image right from scratch where  we've masked. So, if we go in and look here,
  • 00:08:35
    that gets completely regenerated. We want  to make the shirt look more like it's worn,
  • 00:08:42
    like the graphics are on the shirt and  to maintain some of those wrinkles and
  • 00:08:49
    things like that and contours that make it  look more believable. We're going to need
  • 00:08:54
    to lower this down. That way, it doesn't  completely destroy what's underneath it;
  • 00:09:00
    it still uses it. Now, a high number  means that it mostly starts from scratch.
  • 00:09:05
    If you bring it way down, it barely will change  anything at all. Actually, if you go down there,
  • 00:09:12
    it's not going to do anything. For the most part,  you're going to need to be at least over 0.5 to
  • 00:09:16
    get any results. Let's try around 0.75. This  really depends on the look you're going for,
  • 00:09:23
    and each image is going to be slightly different.  You're going to have to play with that as you
  • 00:09:27
    start generating it. Let's go ahead and generate  this and see what we start getting for results.
  • 00:09:34
    Okay, let's look at our images we have here.  These actually came out pretty decent. You can
  • 00:09:42
    see the letters do kind of follow some of the  contours. Now, some of these aren't going to
  • 00:09:47
    come out perfect, but we get a pretty good job.  I'm going to lower this down a little bit more.
  • 00:09:58
    I'm going to put that down closer to 0.7. Okay,  let's see if we can change this up a little bit
  • 00:10:06
    and get it a little more faded. A couple of  these actually, this one actually I thought
  • 00:10:12
    came out pretty good. At this point, let's  say we wanted that but didn't like the design;
  • 00:10:17
    we could in-paint just that area of it. Once  you get the text right, you can always work
  • 00:10:22
    around it. Let's go ahead and try another shot.  This came out halfway decent. Now you can adjust
  • 00:10:31
    the CPDS and other things to get better text.  You can also have it zoom in more if you want
  • 00:10:37
    to maintain that structure, especially if you  have smaller text. If you want to really go in,
  • 00:10:43
    you can zoom in even more similar to how I show  in my other video on how to do text. The biggest
  • 00:10:49
    thing here is going to be the denoising strength.  It's the big difference because to keep the shape
  • 00:10:56
    better so it looks more realistic, you need to  turn that down. If you turn it down too much,
  • 00:11:02
    then it also gets more difficult. The other thing  to keep in mind, which I didn't show here, but
  • 00:11:06
    this will work for more than just text. You could  include images to help guide it, logos, whatever
  • 00:11:15
    you want. That's entirely up to you how you do it.  This just gives you the basic idea of how to get
  • 00:11:21
    that. Obviously, once you get the basic design,  you can then go from there and in-paint it.
  • 00:11:28
    Now, I will mention one thing you might want to  do. If you're going to be upscaling your images,
  • 00:11:32
    you can upscale them first and then do this  because you'll be working with a smaller area
  • 00:11:38
    when you do the in-painting and that respective  field. You can do that with an upscaled image
  • 00:11:44
    as well. Okay, so I've decided let's go ahead  and do an upscaled image. I've got this image
  • 00:11:51
    here I generated. It's going to be a little bit  tougher because it's further back. That would
  • 00:11:59
    normally be very difficult if you try doing  it under the normal image prompts without
  • 00:12:04
    adjusting things because your text would  be so small that it wouldn't maintain the
  • 00:12:09
    structure. But we shouldn't have that issue  because we want to do the same technique.
  • 00:12:15
    I've gone ahead and already masked the area,  got that out of the way. In the image prompt,
  • 00:12:19
    I've changed things a little bit. I've put the  robot and everything right on that image. That
  • 00:12:25
    way I can get more influence. I'm going to  increase the weight a little bit. The stop
  • 00:12:29
    value in the image prompt just for the  style, the look, and everything to help
  • 00:12:33
    influence that result. The prompt is all set.  I've got the in-paint denoising strength set.
  • 00:12:39
    Let's go a little bit lower on that. I've measured  everything out, so this should work. Like I said,
  • 00:12:50
    this has been upscaled. This is already an  upscaled image. Don't worry if the eyes are weird.
  • 00:12:55
    I didn't work on any of that yet. I usually do  that after the fact. Let's go ahead and do this.
  • 00:13:02
    Looking at these four, I'd say overall they're not  too bad. This one didn't get the text right, but
  • 00:13:09
    this is a perfect example. Let's say you did like  this but just the text part at the bottom. You
  • 00:13:14
    could just start and mask that area, remeasure,  and figure it out from there. Just continue. This
  • 00:13:24
    one didn't come out that good. This one's not too  bad. I actually like this one the most. What I
  • 00:13:35
    would normally do at this point is, I would mask  this area to remove the text I don't want. This
  • 00:13:59
    one, I am perfectly fine with that text. We'll  clear this out. I'm going to get rid of these
  • 00:14:04
    things now because let's say I just want to  finish touching those things up. I'm going to
  • 00:14:18
    change the denoising strength now because I don't  want that to influence generation. Put that in
  • 00:14:33
    there. I think we're good to go. Let's give this  a shot. So, we've got a couple of images. This
  • 00:14:42
    one's fine. Now we have our images. I'll give  you a couple of bonuses. This should give you a
  • 00:14:49
    good idea of how to get to this point. I didn't do  many image generations; I just wanted to show the
  • 00:14:56
    technique. This gets a lot easier if you're not  trying to do it on images like the ones I'm using
  • 00:15:03
    that are dirty and worn out. If you're putting  it on somebody that's got a smoother shirt,
  • 00:15:12
    it gets easier to get a better result. But I  wanted to show the more difficult situations.
  • 00:15:20
    Now, if you really want to take that to  the next level and make it look phenomenal,
  • 00:15:24
    you can use Magnific AI. This does cost a bit of  money unless you do a lot of upscaling and want
  • 00:15:32
    a lot of detail. I don't necessarily recommend it,  but if you're using it enough, it is worth it. The
  • 00:15:38
    first one I'm going to do is the one I just did.  I'm going to put my prompt in here. Let's go with
  • 00:15:43
    a scaling factor of two. I'm going to keep the  resemblance up. We're going to leave everything
  • 00:15:50
    else fine. Let's go with that. We're going to  go ahead and upscale that. If you're wondering,
  • 00:15:58
    as I mentioned before, about getting a design on  a shirt in GIMP, I can show you how to do that.
  • 00:16:07
    Let's see if I can remember how to do this.  Open as a layer. Let's scale this down. I'm
  • 00:16:26
    not going to go for perfection here. At this  point, you could rotate it or anything like
  • 00:16:33
    that. I'm not going to worry about that right  now. Then I'm just going to change this. If
  • 00:16:40
    we go down to the mode on that layer, and  you can try different ones, but usually,
  • 00:16:46
    I find overlay works pretty good. There we go.  This shirt doesn't look perfect, and you can
  • 00:17:00
    lower the opacity as well. That helps it blend  in. There's a lot of little tweaks you can do,
  • 00:17:06
    but that's pretty much how you would do it in  GIMP. You could change to other modes. Soft
  • 00:17:14
    light will do it as well. Play around  with different ones, and sometimes it
  • 00:17:19
    depends on the image. That's how you can do  something like that in GIMP if you want to.
  • 00:17:26
    Let's check our Magnific upscale. Here we go.  Now, it went a little extreme, but that one
  • 00:17:36
    actually came out really good. There we go. I  will say I really love Magnific; just wish it
  • 00:17:48
    wasn't so expensive. If you really want to up your  game, that's the way to do it. Well, that covers
  • 00:17:53
    everything for this video. As I said, it does take  a little tweaking and practice. It may not always
  • 00:17:59
    work with every image, but it is one of the better  ways I've found. If someone has a better way,
  • 00:18:04
    I'd love to hear it in the comments. I'm always  looking for improvements and better ways of doing
  • 00:18:09
    things. This is not limited to t-shirts and can  be used for other things as well. If you found
  • 00:18:15
    this video helpful, please hit the like button  so others may discover it as well. Thank you
  • 00:18:19
    to everyone who has donated to help the channel,  and thanks for watching. Now, have fun creating.
标签
  • text generation
  • t-shirts
  • Image editing
  • Fooocus
  • Magnific AI
  • GIMP
  • text prompts
  • denoising
  • image manipulation
  • video tutorial