OneTrainer - My tips for preparing a data set

00:38:45
https://www.youtube.com/watch?v=1qI3K6RBNL0

الملخص

TLDRThe video offers guidance on creating effective datasets for training AI models, especially focusing on character representation. Key aspects include gathering diverse images, ensuring high quality, and removing extraneous characters from training images. The importance of adjusting image metadata, using unique tokens, and linking names to specific characteristics is stressed. The video emphasizes the necessity of using appropriate software versions and troubleshooting potential issues in model training times. This information is applicable to both SDXL and other models, providing viewers with a solid framework for improving AI-generated art accuracy and quality.

الوجبات الجاهزة

  • 📷 Use high-quality images for training.
  • ✨ Edit out unwanted characters from images.
  • 🔑 Implement unique tokens for better recognition.
  • 🔄 Ensure diverse angles and poses in your dataset.
  • 🛠️ Keep software updated for optimal performance.
  • 📊 Understand the importance of image resolution in training.
  • ❌ Avoid using solely illustrations; real images are preferred.
  • 🚫 Remove duplicate characters to focus AI learning.
  • 📏 Resize images adequately before training.
  • ⚙️ Adjust training parameters if necessary.

الجدول الزمني

  • 00:00:00 - 00:05:00

    视频介绍了如何设置数据集以训练AI模型，包括重要提示和技巧。讲者提到了一些观察，尤其是在创建Laora时，需要考虑图像的质量和类型。

  • 00:05:00 - 00:10:00

    讲述了关于Gonzo的设置，强调了获取角色的不同角度和视图的重要性。同时指出并非所有图像都需要是原始角色，还可以使用已创建的玩偶图像来填充数据集。

  • 00:10:00 - 00:15:00

    讨论了图像的分辨率和尺寸对AI学习的影响，指出较大的图像如何帮助模型理解角色特征，同时建议在训练数据集中包含真实图像而非绘画。

  • 00:15:00 - 00:20:00

    讲解了在设置数据集时删除包含其他角色的图像的重要性，以避免模型混淆。强调应专注于所训练的角色，确保AI能够清晰地学习。

  • 00:20:00 - 00:25:00

    展示了一些编辑后的高分辨率图像，并解释了如何在收集数据时保持角色的一致性和多样性。建议使用多个角度和姿势提升模型的学习效果。

  • 00:25:00 - 00:30:00

    讨论了在创建文本文件时，如何更好地标记和描述图像以减少混淆。强调了独特标记的重要性，以帮助模型在训练时给予更多关注。

  • 00:30:00 - 00:38:45

    总结了训练过程中的注意事项，比如图像数量对训练步数的影响，以及不同工具和版本对AI性能的重要性。鼓励观众尝试讲述的方法以提高最终图像的质量。

اعرض المزيد

الخريطة الذهنية

فيديو أسئلة وأجوبة

  • What is the ideal image size for training datasets?

    Images should have a minimum smaller size of 124 pixels, but larger images generally produce better quality results.

  • How can I improve the quality of my AI model training?

    Use a varied dataset with high-quality images, remove unwanted characters, and include descriptive tags.

  • Why do I need tokens when training my model?

    Tokens help distinguish your model's unique characteristics, preventing confusion with similar terms in the dataset.

  • What should I do if my model is not auto-adjusting during training?

    Check for potential installation issues and ensure that necessary components and software are up to date.

  • Is it necessary to remove other characters in my dataset images?

    Yes, removing other characters helps the AI focus on the main subject, reducing confusion during training.

  • How many images do I need for effective training?

    While fewer images can work, generally more images provide better training outcomes, with around 37 images yielding quick training results.

  • What is the importance of image quality in training datasets?

    High-quality images provide better details for the AI to learn from, leading to improved results in generated outputs.

  • Can I use illustrations or drawings in my dataset?

    While you can use them, real images provide more texture and detail for better learning outcomes.

  • What is the recommended process for editing training image metadata?

    Edit the metadata to include relevant tokens and descriptive keywords that clarify the image's content.

  • How do I handle images of different orientations?

    Use images of the same subject in various orientations to give the AI a better understanding of the subject's appearance.

عرض المزيد من ملخصات الفيديو

احصل على وصول فوري إلى ملخصات فيديو YouTube المجانية المدعومة بالذكاء الاصطناعي!
الترجمات
en
التمرير التلقائي:
  • 00:00:00
    hello and welcome to stream tabulous So
  • 00:00:02
    today we're going to do a quick video on
  • 00:00:05
    um your data set for one trainer so how
  • 00:00:09
    to set it up uh what's
  • 00:00:12
    important and the text files for that
  • 00:00:15
    data set uh because I'm getting a lot of
  • 00:00:18
    well I'm creating the uh Laura and it's
  • 00:00:21
    not looking like my um my data set so
  • 00:00:24
    the images that you're feeding into it
  • 00:00:27
    so I want to talk about uh little tips
  • 00:00:29
    and tricks
  • 00:00:30
    um what to consider when creating the
  • 00:00:33
    Laura uh for particularly
  • 00:00:37
    sdxl uh this does apply to the um the
  • 00:00:40
    512s with um just the SD 1.5 as well so
  • 00:00:45
    it will help both out so we'll go over
  • 00:00:48
    to the intro and then we'll take a look
  • 00:00:50
    at that
  • 00:00:55
    [Music]
  • 00:01:06
    okay so we're going to take a look at
  • 00:01:07
    the um the data set for Gonzo and we're
  • 00:01:11
    going to talk about that quickly uh so
  • 00:01:13
    I've gone on and I've collected as many
  • 00:01:17
    images as possible now you know I've
  • 00:01:18
    already created this Laura and if you
  • 00:01:20
    have head over to Civ AI you'll actually
  • 00:01:24
    see images created with this Laur to
  • 00:01:26
    give an idea how well it actually worked
  • 00:01:29
    out the same principle applies for a
  • 00:01:32
    person as well and uh we'll talk about
  • 00:01:35
    both while we're doing this so I've gone
  • 00:01:38
    in and I've done the data set now
  • 00:01:42
    there's not a lot of images available on
  • 00:01:44
    the character uh unless you're going to
  • 00:01:46
    go through the show
  • 00:01:48
    and get as many different angles and
  • 00:01:50
    many different views as possible of the
  • 00:01:52
    character and screenshot
  • 00:01:55
    them now not all of these are actually
  • 00:02:00
    the um these ones might be okay it would
  • 00:02:03
    have been
  • 00:02:04
    the the data set for um animal a lot of
  • 00:02:10
    the images that were used in that were
  • 00:02:12
    actually from the purchasable puppet
  • 00:02:16
    that you can get so the um because
  • 00:02:18
    there's legs in that so it's where
  • 00:02:20
    people have purchased it and just set up
  • 00:02:22
    photos and put that in so keep that in
  • 00:02:25
    mind you don't necessarily need every
  • 00:02:27
    picture to be of the the um the
  • 00:02:31
    character in The Way of the uh the TV
  • 00:02:34
    show you can use puppets that people
  • 00:02:37
    have actually created and uh characters
  • 00:02:40
    and cartoon characters of it when the AI
  • 00:02:43
    looks at it it looks at it as an overall
  • 00:02:46
    and it
  • 00:02:47
    will it doesn't copy the
  • 00:02:51
    image so you don't have to worry about
  • 00:02:53
    that you looking for the
  • 00:02:55
    likeness uh physical photos of the real
  • 00:02:58
    puppet have a a lot of information in it
  • 00:03:03
    where we can see the
  • 00:03:04
    textures uh when it comes to a painting
  • 00:03:07
    the painting is not going to have or
  • 00:03:09
    drawing is not going to have the uh the
  • 00:03:11
    textures and the details so we want to
  • 00:03:14
    populate it with more real pictures than
  • 00:03:17
    drawings and so forth but uh with um
  • 00:03:21
    animal there weren't a lot of photos
  • 00:03:23
    where the uh the legs were in it which
  • 00:03:26
    is the other thing people uh Wonder well
  • 00:03:28
    you know how do you get the
  • 00:03:30
    but you got to remember this is creating
  • 00:03:33
    and populating as many images as
  • 00:03:35
    possible and most of them are from the
  • 00:03:37
    waist up but it is a Laura and you're
  • 00:03:41
    using it with a base model so the AI can
  • 00:03:44
    extrapolate and imagine the lower part
  • 00:03:47
    so you may get more human body figure
  • 00:03:51
    for the lower part of the model but it
  • 00:03:54
    will work it out and it will fill it in
  • 00:03:56
    and we can see here that we do have some
  • 00:03:58
    legs it will work out the missing
  • 00:04:01
    sections of it so we don't have to be
  • 00:04:03
    perfect so now all these images were
  • 00:04:06
    tiny as we can see here this is only 199
  • 00:04:09
    pixels by um 253 pixels and we talk in
  • 00:04:13
    pixels we don't talk in inches um we
  • 00:04:15
    don't talk in centimeters we're talking
  • 00:04:17
    pixels because it's um it's the computer
  • 00:04:21
    related image that we're actually
  • 00:04:23
    talking about uh we can
  • 00:04:26
    also mention the PPI pixels per square
  • 00:04:29
    Square in on top of that and the higher
  • 00:04:32
    that is obviously the better quality
  • 00:04:34
    you're going to have so the first thing
  • 00:04:36
    you want to do is go through and resize
  • 00:04:39
    all these to
  • 00:04:42
    124 by whatever so the smaller size of
  • 00:04:46
    the image needs to be
  • 00:04:47
    124 so the width of this one for example
  • 00:04:51
    will be 124 but the height will be over
  • 00:04:54
    124 it's the same if you're doing a
  • 00:04:56
    person it'll be 124 and if it's the full
  • 00:04:59
    body in the picture it may be 3,500 for
  • 00:05:02
    the height so that is perfectly fine it
  • 00:05:06
    doesn't have to be 124x 124 when we're
  • 00:05:09
    training a data set you need to remove
  • 00:05:13
    that picture from your head that okay
  • 00:05:15
    it's going to create this picture it's
  • 00:05:16
    not going to create that picture it's
  • 00:05:18
    not going to look at that picture that
  • 00:05:19
    picture doesn't exist what exists are
  • 00:05:23
    keys and what I would mean by that is it
  • 00:05:27
    knows what an i roughly looks like so
  • 00:05:30
    the AI will go through and it'll go okay
  • 00:05:31
    this is an eye or this is what a human
  • 00:05:34
    eye looks like this is close to an eye
  • 00:05:36
    I've been trained with eyes of artwork
  • 00:05:39
    and this is very much what I know an eye
  • 00:05:41
    to look like so it will use your image
  • 00:05:45
    to create that eye you might not know
  • 00:05:48
    what this is but it knows it's on the
  • 00:05:50
    face and it knows it's where the beak is
  • 00:05:52
    so it's going to go through and create
  • 00:05:54
    this as well and that's what it's going
  • 00:05:56
    to do with the whole image so it's going
  • 00:05:58
    to go okay well that's a mouth I've got
  • 00:06:00
    artwork and um you know I've been
  • 00:06:02
    trained with uh other puppets not this
  • 00:06:05
    puppet and so forth so it has an idea
  • 00:06:08
    what it's looking at and it's learning
  • 00:06:10
    that information it is not learning this
  • 00:06:13
    image so it may learn well this is a
  • 00:06:16
    shirt and this is um you know a diamond
  • 00:06:19
    pattern on the shirt with red and white
  • 00:06:21
    lines and it's gray and black so it will
  • 00:06:25
    learn that so when you prompt it you may
  • 00:06:27
    get something which looks very very
  • 00:06:29
    close to this but you will not get this
  • 00:06:34
    image created you're getting something
  • 00:06:36
    that is new okay
  • 00:06:41
    so keep in mind that if we take a look
  • 00:06:44
    we'll come back to this one we'll take a
  • 00:06:45
    look at this and we'll zoom in it's very
  • 00:06:47
    pixelated it's not high quality at all
  • 00:06:50
    so when it's learning that it's going to
  • 00:06:52
    say well this isn't a very good quality
  • 00:06:54
    eye and it's going to learn a bad
  • 00:06:56
    representation of this we want the best
  • 00:07:00
    possible
  • 00:07:01
    quality so larger images are better
  • 00:07:05
    what's the downfall of a large image if
  • 00:07:07
    you can train this at 4K if you want to
  • 00:07:11
    but then when you try to create a canvas
  • 00:07:13
    of say uh 1024x 1024 it's going to be an
  • 00:07:18
    absolute mess because it's going to be
  • 00:07:20
    trying to compress that 4K image down
  • 00:07:23
    into that and that is something to keep
  • 00:07:26
    in mind when you're if you're training
  • 00:07:27
    an sdxl of a person and you want that
  • 00:07:30
    high quality
  • 00:07:31
    person your canvas size needs to be um
  • 00:07:35
    larger so if you're going for a portrait
  • 00:07:38
    view you need 1024 by here and you'll
  • 00:07:41
    need higher here it could be 1,600 and
  • 00:07:45
    above so keep that in mind and that will
  • 00:07:48
    give you the better quality so the
  • 00:07:50
    larger the canvas the better the quality
  • 00:07:52
    you're going to get when you're using
  • 00:07:53
    your model trained with large image data
  • 00:07:57
    sets if you're not training at at
  • 00:08:01
    1024x above then you might as well just
  • 00:08:05
    be training a SD 1.5 and using that and
  • 00:08:09
    then you've got the problem that the um
  • 00:08:12
    the core of the AI isn't as good for
  • 00:08:15
    knowing how to do poses and backgrounds
  • 00:08:18
    and changing it and you're limited to
  • 00:08:20
    what you can actually do with it that's
  • 00:08:22
    the difference with the sdxl sdxl is
  • 00:08:26
    better across the board for
  • 00:08:28
    understanding what you want to create
  • 00:08:30
    when you're putting in your prompt
  • 00:08:33
    authoring so we resized this another
  • 00:08:36
    thing we need to look at is things like
  • 00:08:38
    this we got other characters in there do
  • 00:08:41
    we really want those other characters in
  • 00:08:43
    there where on the off chance the AA
  • 00:08:47
    might come through and go this is a
  • 00:08:48
    green hand cuz we can see it looks like
  • 00:08:50
    knuckles I mean obviously there's a hand
  • 00:08:52
    in the puppet so it may get mistaken for
  • 00:08:54
    a hand and then occasionally you might
  • 00:08:57
    get a green weird looking kmit hand on
  • 00:08:59
    Gonzo so it's best to come through and
  • 00:09:03
    mask them out you can mask in one
  • 00:09:05
    trainer I don't bother doing that cuz uh
  • 00:09:08
    obviously I'm going through and I'm
  • 00:09:09
    editing all my photos to get the best
  • 00:09:11
    quality first so I might as well just
  • 00:09:13
    eliminate the characters when I'm doing
  • 00:09:16
    that so something like an animal the AI
  • 00:09:21
    is going to know this is an animal and
  • 00:09:23
    it's pretty much got to know that it's
  • 00:09:24
    chicken to be quite Frank so we don't
  • 00:09:27
    really have to mask that out we can
  • 00:09:28
    leave that in in there and we can have
  • 00:09:31
    um I think it's
  • 00:09:33
    CA as the chicken and we can leave that
  • 00:09:37
    as so if we put the word Chicken in
  • 00:09:39
    there we'll end up getting similar
  • 00:09:41
    results to what these chickens look like
  • 00:09:44
    uh so keep that in mind anything which
  • 00:09:47
    has multiple characters we really want
  • 00:09:50
    to remove those multiple characters
  • 00:09:54
    okay so let's take a look at what that
  • 00:09:56
    looks like
  • 00:10:00
    these are the edited these are all large
  • 00:10:03
    and we can see here this one is on its
  • 00:10:05
    height is
  • 00:10:08
    1,214 this one on its height is
  • 00:10:14
    1,279 and um yeah I mean if it's a long
  • 00:10:18
    photo then it's going to be a lot taller
  • 00:10:20
    than that we probably come through and
  • 00:10:22
    find one that's bigger than that uh I
  • 00:10:24
    have certainly trained a data set with
  • 00:10:26
    um 3,000 for the height numerous times
  • 00:10:30
    and even for the width so now when we
  • 00:10:33
    come through if we can find um that same
  • 00:10:35
    photo that we're looking at
  • 00:10:38
    before want
  • 00:10:40
    the this one here now this has been
  • 00:10:44
    upscaled so when we come in now we can
  • 00:10:47
    see this is a lot cleaner image so the
  • 00:10:50
    AI will have a better understanding of
  • 00:10:53
    representing that and then the more data
  • 00:10:56
    there is on what that looks like we get
  • 00:11:01
    more information for the AI to actually
  • 00:11:04
    learn it and of course there's the
  • 00:11:06
    different poses the different angles so
  • 00:11:09
    AI can extrapolate that and have a
  • 00:11:12
    better understanding of the way this
  • 00:11:15
    character actually is going to
  • 00:11:17
    look now with one trainer we've gone
  • 00:11:20
    over uh using the um the extra tools to
  • 00:11:24
    create the text files now if we open
  • 00:11:27
    that
  • 00:11:27
    up what we may get is something which is
  • 00:11:32
    obviously my token token is something
  • 00:11:35
    unique keep that in mind because uh if
  • 00:11:38
    it's just the first name with the last
  • 00:11:42
    initial you're going to find there's
  • 00:11:44
    probably a lot in the data set of the
  • 00:11:46
    core model which has first names with
  • 00:11:49
    last initials and it's going to create
  • 00:11:52
    conflict so what a lot of people do is
  • 00:11:55
    they just put an X at the beginning and
  • 00:11:56
    an X at the end to create their token
  • 00:12:02
    now when we come through this we'll
  • 00:12:03
    notice here there's some things which
  • 00:12:05
    aren't quite right so bird I don't want
  • 00:12:08
    this confused with a bird I want this to
  • 00:12:11
    actually read as Gonzo so if we come
  • 00:12:14
    back here and we'll leave that one open
  • 00:12:17
    we'll go through and have a look at the
  • 00:12:19
    edited one and we'll compare that top
  • 00:12:22
    and
  • 00:12:25
    bottom okay
  • 00:12:29
    so you can see now that I've come
  • 00:12:31
    through and I've removed the word bird
  • 00:12:34
    I've replaced it with the word Gonzo by
  • 00:12:37
    itself and you can replace it with the
  • 00:12:40
    Single Character name so um know maybe
  • 00:12:45
    Claudia C so for example now keep in
  • 00:12:48
    mind Claudia C may get confused with the
  • 00:12:51
    famous superm model when it's doing it
  • 00:12:53
    but we have our token and the token's
  • 00:12:55
    the key one that we want to be using
  • 00:12:56
    when we're creating our prompt anyway
  • 00:12:58
    but we can have a basic name here and
  • 00:13:01
    then the double understanding the AI
  • 00:13:05
    goes I've seen that word with the token
  • 00:13:08
    in the model so I will pay more
  • 00:13:11
    attention to the Laura model than the
  • 00:13:14
    base model that you're using when
  • 00:13:16
    creating it because of what can happen
  • 00:13:18
    and this happens when you're creating a
  • 00:13:20
    picture of someone it can get confused
  • 00:13:24
    with the representation of the um the
  • 00:13:28
    the train training language with the
  • 00:13:31
    training of the model language for a
  • 00:13:33
    similar person so black hair blue eyes
  • 00:13:35
    white skin we're pretty much going to
  • 00:13:39
    cover a lot of different people in the
  • 00:13:41
    world and a lot of training set so the
  • 00:13:44
    name alone might not just be enough so
  • 00:13:48
    the more information you put in so you
  • 00:13:50
    put your token in and then you put the
  • 00:13:52
    name in and then you put black hair
  • 00:13:54
    white skin blue eyes it will help point
  • 00:13:57
    it more if you're use that name more in
  • 00:14:01
    your training data to use your Laura and
  • 00:14:05
    then you've got your weight as well so
  • 00:14:07
    increase your weight so it pays more
  • 00:14:08
    attention to your Laura as well but if
  • 00:14:11
    you go too high you start getting
  • 00:14:12
    artifacts so that's something to keep in
  • 00:14:14
    mind so we come through and yeah this
  • 00:14:18
    one I've got no human he's a muppet and
  • 00:14:21
    if we come through and we have a look at
  • 00:14:22
    some more of our training data and this
  • 00:14:26
    is our final one so I've come through
  • 00:14:28
    and edited that but we'll go back
  • 00:14:30
    through and we'll take a look
  • 00:14:32
    at this one actually does have a bird in
  • 00:14:35
    it and it has a chicken in it so I've
  • 00:14:37
    left that it still says no humans but it
  • 00:14:40
    generally works pretty fine with the
  • 00:14:41
    training data now what I come through
  • 00:14:44
    and worked out later on is I should have
  • 00:14:47
    added tokens for this pose because uh
  • 00:14:51
    when you put the word flying
  • 00:14:53
    in there's a lot of things which are
  • 00:14:56
    like balloons that are just flying in
  • 00:14:58
    the air so what you get is you get a lot
  • 00:15:00
    of vertical rather than
  • 00:15:03
    horizontal images of people flying so
  • 00:15:06
    the way around that would have been to
  • 00:15:09
    come through and actually
  • 00:15:13
    have would have been to come through and
  • 00:15:16
    where the word flying is I should have
  • 00:15:19
    put the word flying up here in the same
  • 00:15:21
    manner so I should have put the um the
  • 00:15:25
    comma and then I should have put flying
  • 00:15:27
    but I should have tokened the word fly
  • 00:15:29
    flying so if I had to put um X flying X
  • 00:15:33
    then whenever I wanted the Superman pose
  • 00:15:36
    of it flying it would have forced with
  • 00:15:39
    the token to use my Laura model over top
  • 00:15:43
    of the core AI which would have given me
  • 00:15:47
    these poses instead so without that in
  • 00:15:52
    the Lura it is very much hit and miss to
  • 00:15:56
    get these flying and is not enough
  • 00:16:00
    images of him flying to actually get
  • 00:16:02
    that in and I don't know if the balloon
  • 00:16:05
    one might have um flying as
  • 00:16:07
    well just says in the air so that's
  • 00:16:10
    something to keep in mind so you can do
  • 00:16:13
    multiple tokens as well to force it to
  • 00:16:16
    it uh
  • 00:16:17
    so that is useful if there's something
  • 00:16:20
    specific in the IM image in the pose
  • 00:16:23
    that you want to force it to so if you
  • 00:16:25
    have the pose of Gonzo flying of course
  • 00:16:29
    that would have been better off as a
  • 00:16:31
    token so there the things to keep in
  • 00:16:34
    mind and we noticed down here uh the the
  • 00:16:37
    name is wrong so when we come through
  • 00:16:42
    it's picked up this weird person a lot
  • 00:16:47
    for
  • 00:16:50
    um for the character description I don't
  • 00:16:53
    know why um as far as I can tell it's a
  • 00:16:58
    middle eastern
  • 00:16:59
    nationality person when I Google it and
  • 00:17:03
    um looks nothing like Gonzo so unsure
  • 00:17:06
    why the AI was really confused with that
  • 00:17:09
    so editing the um that data is so
  • 00:17:13
    important to come through and make sure
  • 00:17:16
    that you're coming through and editing
  • 00:17:18
    every one of these so doing the Muppet
  • 00:17:21
    is a lot of editing it's very important
  • 00:17:24
    to go through every one of your images
  • 00:17:27
    and add extra information to it or
  • 00:17:29
    remove information there's a lot of
  • 00:17:31
    people that say if you're training a
  • 00:17:32
    person don't bother having this in there
  • 00:17:36
    just have it as this so if you're
  • 00:17:37
    training a face or a person you could
  • 00:17:39
    try that without actually having the U
  • 00:17:42
    the rest of it and every file just
  • 00:17:45
    having just the token um I prefer to
  • 00:17:49
    have it with the description of clothing
  • 00:17:52
    and that because if I'm creating it I
  • 00:17:54
    like to be able to trigger uh that
  • 00:17:58
    information for the clothing style or I
  • 00:18:01
    like to be able to put in the negative
  • 00:18:02
    prompt and remove that clothing style if
  • 00:18:05
    it's forcing it too much from my um from
  • 00:18:09
    my Laura model if you have just the word
  • 00:18:11
    Gonzo in for example and every single
  • 00:18:14
    image
  • 00:18:15
    is the same um say this top and you
  • 00:18:20
    don't have information in there for
  • 00:18:22
    saying like a Tweed top um or whatever
  • 00:18:25
    then every time you render it it's going
  • 00:18:28
    to to push that same Red Top every
  • 00:18:31
    single time uh even if you're putting it
  • 00:18:34
    in the negative because uh it doesn't
  • 00:18:37
    understand that that is a red top
  • 00:18:40
    because it's not in its data set for its
  • 00:18:44
    language so keep that in mind a lot of
  • 00:18:47
    people train themselves and they just
  • 00:18:49
    use their name uh but it's might be
  • 00:18:53
    quite fine if you've got a mix of
  • 00:18:54
    clothing myself uh I love black it's
  • 00:18:57
    slimming I'm a short person so it also
  • 00:18:59
    helps make me look a little bit taller
  • 00:19:01
    uh because that slimming aspect of it so
  • 00:19:04
    I love black and I love dark colors for
  • 00:19:06
    that reason now when it comes to um if I
  • 00:19:11
    was to train a data set with just my
  • 00:19:13
    name it's going to force black black
  • 00:19:15
    black black black so I'd never be able
  • 00:19:17
    to really control it the way I want to
  • 00:19:21
    so that's something to keep in mind and
  • 00:19:23
    you could maybe train the model twice of
  • 00:19:26
    just having it as the um
  • 00:19:29
    the token name and then you could also
  • 00:19:31
    train another model with all the other
  • 00:19:33
    information and you can use those two
  • 00:19:35
    models together and that uh can give you
  • 00:19:39
    more control so you can lower the weight
  • 00:19:42
    on just the uh one with the name the
  • 00:19:45
    token and then you can have a higher
  • 00:19:47
    weight on the one with the descriptive
  • 00:19:50
    clothing and that can actually help
  • 00:19:53
    refine so it goes okay this is that face
  • 00:19:55
    this is that face but I now have
  • 00:19:57
    information and I can see that this is
  • 00:19:59
    the same face but I'm removing clothing
  • 00:20:01
    I have found that actually works for me
  • 00:20:04
    um if I've gotten a Laura of say RoboCop
  • 00:20:08
    which I did
  • 00:20:09
    recently both luras were subport par one
  • 00:20:14
    Laura was exceptionally bad but using
  • 00:20:17
    the two together was better for creating
  • 00:20:21
    the character so if you've seen my
  • 00:20:24
    RoboCop image um using the zisha
  • 00:20:29
    version 4 model which I've got to do a
  • 00:20:31
    review on that is absolutely fantastic
  • 00:20:33
    it's a mindblowing model um I actually
  • 00:20:37
    used two luras of
  • 00:20:39
    Robocop to get the weights right and mix
  • 00:20:43
    that to get the image so perfect crisp
  • 00:20:46
    and get the details that I was looking
  • 00:20:48
    for and then the prompt offer for that
  • 00:20:52
    is quite literally an A4 page long um
  • 00:20:55
    there's a lot of descriptions on Tech
  • 00:20:58
    textures and macro textures of the um
  • 00:21:01
    the metal to so when you zoom in you get
  • 00:21:04
    those fine little um grains because they
  • 00:21:08
    when they painted the RoboCop armor for
  • 00:21:11
    the movie uh they painted it with
  • 00:21:13
    automotive paint I believe it had a um
  • 00:21:17
    uh it might it was a pearlescent and I
  • 00:21:19
    believe it had a touch of chameleon in
  • 00:21:21
    the blue so it had a bit of a rainbow
  • 00:21:24
    effect to it so the blues to the sort of
  • 00:21:27
    semi- purple pink tone
  • 00:21:29
    and up close you could actually see the
  • 00:21:32
    sort of stippled effect of the way that
  • 00:21:35
    um paint was done and with the
  • 00:21:38
    highdefinition movie you can really see
  • 00:21:40
    those details so I wanted that to come
  • 00:21:42
    through so the prompt author is
  • 00:21:45
    exceptionally long a lot of people think
  • 00:21:48
    that there's no art to doing AI believe
  • 00:21:52
    me if you want to create an image and
  • 00:21:54
    you've got that image stuck in your head
  • 00:21:56
    your prompt author which is fully
  • 00:21:58
    copyrightable in all countries is very
  • 00:22:02
    very unique and that output in the image
  • 00:22:06
    is unique so you'll know if someone's
  • 00:22:08
    pinched your prompt one to one because
  • 00:22:10
    you
  • 00:22:11
    can see it because it it's repeatable so
  • 00:22:14
    every time you render the image it comes
  • 00:22:16
    up the exact same details and uh that's
  • 00:22:20
    something to keep in mind that's when we
  • 00:22:22
    do these we have the information in
  • 00:22:24
    there and then when we're doing the rest
  • 00:22:26
    of our prompt or we're doing our prompt
  • 00:22:28
    authoring and you can put um you know
  • 00:22:32
    macro details on vest and then when it
  • 00:22:35
    comes through and it takes a look at
  • 00:22:37
    that image we'll get more details on
  • 00:22:39
    that vest the textures acrossed with the
  • 00:22:42
    core model as
  • 00:22:44
    well
  • 00:22:47
    so what else do we want to know we come
  • 00:22:49
    through here and um you notice the space
  • 00:22:52
    image I've removed the other characters
  • 00:22:54
    cuz I wanted the space helmet I thought
  • 00:22:56
    it was unique um I've got it one once or
  • 00:22:58
    twice creating a space character but not
  • 00:23:01
    quite to this it's more like a bit of a
  • 00:23:03
    bubble but it worked well so I wanted
  • 00:23:07
    that one in there I should have
  • 00:23:09
    tokened um I should have done uh Bowl
  • 00:23:12
    helmets and I should have really joined
  • 00:23:15
    that by
  • 00:23:16
    putting that and I should have put X and
  • 00:23:19
    an X at the end and tokened it and that
  • 00:23:22
    would have um helped point the that to
  • 00:23:26
    my model which would have been fantastic
  • 00:23:29
    to get closer but also it would mean
  • 00:23:32
    that I would only get a front forward
  • 00:23:34
    facing version of Gonzo uh so that's
  • 00:23:37
    something to keep in mind and again if
  • 00:23:39
    we come through here we'll notice that
  • 00:23:41
    there's um no images of the other
  • 00:23:45
    characters because I've gone through and
  • 00:23:47
    edited those out so if we can go back
  • 00:23:51
    through to
  • 00:23:52
    one we'll come through and say this one
  • 00:23:56
    here 21 I don't know if it's still 21 in
  • 00:23:58
    the
  • 00:24:01
    final it is not ah but we can see it
  • 00:24:05
    here I have cropped it
  • 00:24:07
    out uh and some of these where the other
  • 00:24:11
    characters were in the background I've
  • 00:24:13
    simply gone through and I've painted
  • 00:24:15
    them out so that's something you can do
  • 00:24:18
    as well is just go through and paint
  • 00:24:20
    them out and I'd like
  • 00:24:23
    to show one if I can
  • 00:24:30
    okay we got that in the background there
  • 00:24:32
    and I believe I removed that so if we
  • 00:24:35
    can find
  • 00:24:37
    him ah we can see I've cropped that
  • 00:24:40
    down so I'm removing characters but I've
  • 00:24:43
    also um just painted them out uh because
  • 00:24:45
    that's all that masking is it just
  • 00:24:47
    paints them out so if I'm can find one
  • 00:24:51
    give an
  • 00:24:54
    idea ah there's a classic one so we
  • 00:24:57
    didn't want the rat for into the picture
  • 00:24:59
    so
  • 00:25:06
    38 here we go so I have removed it from
  • 00:25:10
    the image to ensure that when I did the
  • 00:25:12
    flying that I would get this pushed
  • 00:25:15
    without having a bit of confusion
  • 00:25:16
    between the characters so I've done that
  • 00:25:19
    everywhere where there was another
  • 00:25:20
    puppet I've gone through and remove them
  • 00:25:23
    to create that data set to concentrate
  • 00:25:27
    on just my character I was okay leaving
  • 00:25:30
    Camille the
  • 00:25:31
    chicken probably pronouncing that wrong
  • 00:25:34
    um so this one was edited to remove the
  • 00:25:36
    background cuz I really just wanted to
  • 00:25:38
    concentrate on the character and again I
  • 00:25:41
    put this one in twice I left the
  • 00:25:43
    background on this one and I removed it
  • 00:25:45
    on this one sometimes I've done that
  • 00:25:47
    with um multiple images sometimes I've
  • 00:25:51
    flipped the image so it might be facing
  • 00:25:53
    in One Direction with this hand raised
  • 00:25:55
    and then I've put the exact same image
  • 00:25:57
    in and I flipped it so I've got left
  • 00:25:59
    hand raised right hand raised so you can
  • 00:26:03
    do that as well so you can use the same
  • 00:26:05
    image and flip it so left side of the
  • 00:26:09
    character flip it get the right side of
  • 00:26:12
    the character and vice versa so you can
  • 00:26:14
    use that same image multiple times to
  • 00:26:16
    create more of a data set and you can
  • 00:26:19
    see that there so that gives me both
  • 00:26:21
    sides of the character so then when I'm
  • 00:26:25
    prompting I'm not always going to be
  • 00:26:27
    Force this EX exact same facing
  • 00:26:29
    direction if I put side view it's going
  • 00:26:32
    to give me both directions so I get more
  • 00:26:36
    availability of um options with the
  • 00:26:40
    character so that's something very
  • 00:26:42
    important to keep in mind uh feed the
  • 00:26:44
    data set multiple
  • 00:26:47
    times so now now when it comes to
  • 00:26:49
    training um I trained a model last night
  • 00:26:53
    on 37 images I believe it was and it
  • 00:26:57
    trained in less than an hour at 17 steps
  • 00:27:02
    now I've got some people saying that
  • 00:27:04
    they're not getting that uh if they've
  • 00:27:07
    got 37 images they're getting 37 steps
  • 00:27:10
    if they got 200 images they're getting
  • 00:27:13
    200 steps it's not Auto adjusting um
  • 00:27:17
    with the Adam West and my settings that
  • 00:27:20
    is meant to auto adjust and calculate
  • 00:27:23
    the amount of steps that it's actually
  • 00:27:25
    doing if it is not auto adjust in to do
  • 00:27:29
    the steps there is a strong possibility
  • 00:27:33
    that the AI is not learning the images
  • 00:27:38
    correctly and that may create a uh Aura
  • 00:27:42
    which isn't working so I'm wondering if
  • 00:27:45
    that's the case for some people uh you
  • 00:27:48
    can leave a comment below if you have
  • 00:27:50
    that problem and if you're having that
  • 00:27:52
    problem is your Laura actually still
  • 00:27:55
    working uh what would it affect if your
  • 00:27:58
    Laura is working well obviously if you
  • 00:28:00
    got 37 images and it's not doing it 17
  • 00:28:02
    steps it's going to double the amount of
  • 00:28:05
    time that it actually takes to create
  • 00:28:06
    the
  • 00:28:07
    Laura uh so how could we lower the steps
  • 00:28:11
    we can go through to where it says
  • 00:28:13
    repeats in one trainer so I might just
  • 00:28:15
    open that up quickly which is going to
  • 00:28:16
    take a moment so we'll come
  • 00:28:21
    back okay so this is a screen recording
  • 00:28:24
    from the other system because one
  • 00:28:26
    trainer on this one doesn't want to open
  • 00:28:28
    up at the moment um it's coming up with
  • 00:28:31
    an error so I'm currently updating it so
  • 00:28:36
    this is just from the other system so I
  • 00:28:38
    don't have to wait cuz it's taking
  • 00:28:40
    forever to update and um
  • 00:28:43
    yeah so let's hit play here so what you
  • 00:28:45
    want to do is come through obviously you
  • 00:28:47
    got General and if you come through to
  • 00:28:50
    um Concepts and then you click on your
  • 00:28:52
    images you'll see here there is
  • 00:28:56
    repeats now
  • 00:28:58
    I leave that on default and um it auto
  • 00:29:03
    adjusts for me so 37 images it does it
  • 00:29:06
    17 to 18 steps now as I said if you're
  • 00:29:11
    um if you've got 37 images and you're
  • 00:29:13
    getting 37 steps that means it's not
  • 00:29:16
    Auto
  • 00:29:17
    adjusting uh it's likely to be something
  • 00:29:20
    that you've got running um or installed
  • 00:29:23
    incorrectly which is affecting one
  • 00:29:25
    trainer what that is there is numerous
  • 00:29:28
    things it's really hard to
  • 00:29:31
    tell now if you're finding that your
  • 00:29:33
    training is working uh still so you'll
  • 00:29:37
    have to leave a comment below and let me
  • 00:29:38
    know if it still works without the auto
  • 00:29:41
    adjusting you may have to adjust this so
  • 00:29:45
    for 37 steps um it used to be you could
  • 00:29:48
    write the steps in it actually used to
  • 00:29:50
    say steps not repeats and you could put
  • 00:29:54
    the number of steps for the training
  • 00:29:56
    which was a more logic iCal better way
  • 00:29:59
    to work so I could have put 17 in there
  • 00:30:01
    and I'd know I'd get 17 steps and for
  • 00:30:04
    that there was the mathematical equation
  • 00:30:06
    which we talked about in qu where you do
  • 00:30:09
    I think it was 3500 divided the number
  • 00:30:11
    of images and you'd get your um how many
  • 00:30:14
    steps you should be
  • 00:30:17
    doing cabbage truck's finally coming
  • 00:30:19
    around so just ignore that in the
  • 00:30:21
    background if you can hear it if the
  • 00:30:22
    noise cancellation isn't picking it
  • 00:30:24
    out now so 37 steps you could put 0
  • 00:30:28
    point in a
  • 00:30:30
    number what that number would be to get
  • 00:30:33
    those steps I'm not sure what the
  • 00:30:35
    mathematics behind this is to determine
  • 00:30:38
    the calculation same if you wanted more
  • 00:30:42
    steps uh you could put two in so let's
  • 00:30:45
    say 30 images you put two in it's not 60
  • 00:30:49
    so I don't know what the magic math is
  • 00:30:52
    for this it was far simpler when it just
  • 00:30:54
    said
  • 00:30:56
    steps uh so that something to keep in
  • 00:30:59
    mind there is the drop down here to
  • 00:31:01
    change that but I'm not sure how that
  • 00:31:04
    actually affects one trainer and what
  • 00:31:06
    that actually does as I said I haven't
  • 00:31:08
    played with it cuz I don't need to uh my
  • 00:31:12
    one trainer is working correctly and
  • 00:31:15
    it's automatically adjusting for
  • 00:31:21
    me so that's something to keep in mind
  • 00:31:25
    um I'd like to know below what you'll
  • 00:31:27
    get getting for your steps and that uh
  • 00:31:30
    okay so we'll take a look at what I
  • 00:31:32
    actually have installed because uh that
  • 00:31:35
    might be affecting it for you and it's
  • 00:31:36
    something to know about
  • 00:31:39
    now because I hadn't used one trainer
  • 00:31:43
    before when I installed it I didn't see
  • 00:31:46
    any like U Magic Pudding um or anything
  • 00:31:49
    specific that it was saying that you
  • 00:31:50
    needed to install to get it going when I
  • 00:31:52
    read the installation so I'm not sure if
  • 00:31:56
    there's something which
  • 00:31:59
    is just been installed when I did queer
  • 00:32:01
    which is important to running one
  • 00:32:05
    trainer I'm not entirely sure but I did
  • 00:32:08
    find out that uh these particular tools
  • 00:32:11
    that I have installed are actually
  • 00:32:13
    important for all AI based programs so
  • 00:32:18
    whether it be um editing the image
  • 00:32:20
    creating luras and so forth uh so we'll
  • 00:32:23
    come over and take a look at
  • 00:32:26
    those Okay so here we are um in the
  • 00:32:29
    folder which I have for what I call my
  • 00:32:31
    AI toolkit pack um these are the
  • 00:32:34
    essentials for running
  • 00:32:36
    AI now I did have 12 uh the Cuda tool
  • 00:32:41
    kit 12.4 installed but I was Finding I
  • 00:32:45
    was having issues with that uh come to
  • 00:32:48
    think of it with Adam West not the 8bit
  • 00:32:52
    which is also meant to auto adjust for
  • 00:32:56
    its um
  • 00:32:58
    steps that wasn't working originally for
  • 00:33:01
    me and I ended up rolling back to
  • 00:33:06
    11.8 and uh reinstalling one trainer
  • 00:33:10
    After Rolling everything back
  • 00:33:13
    and Adam West 8bit which is what I used
  • 00:33:16
    thereafter has been working flawlessly
  • 00:33:20
    so Nidia toolkit and we've I've talked
  • 00:33:24
    about this we've talked about this um
  • 00:33:26
    when it came to installing queer and
  • 00:33:29
    what you actually need to install so
  • 00:33:33
    11.8 um or specifically
  • 00:33:36
    11.80
  • 00:33:38
    52206 now um I did put the toolkit up
  • 00:33:42
    for the the zip file uh on
  • 00:33:47
    um on my uh cloud and that does have
  • 00:33:55
    yeah versions in it so what I will
  • 00:33:58
    do in the comments wherever YouTube
  • 00:34:01
    decides to slap them so whether they're
  • 00:34:03
    left right up down wherever I will put a
  • 00:34:06
    link to um my toolkit uh because this
  • 00:34:09
    might actually resolve your
  • 00:34:12
    issues so I have that installed um fmeg
  • 00:34:17
    I have installed as well but of course I
  • 00:34:19
    have it installed via the K light codak
  • 00:34:22
    pack uh git I have installed I have
  • 00:34:26
    python installed and I believe I am on
  • 00:34:30
    the 3106 version I'd have to check my
  • 00:34:33
    uninstaller for that um but either way
  • 00:34:37
    uh 310 was your base and that was the
  • 00:34:40
    update from there and I think in the
  • 00:34:42
    toolkit let's have a look what we've
  • 00:34:47
    got hang get that toolkit to
  • 00:34:51
    open that
  • 00:34:54
    one yeah 3106
  • 00:34:58
    so 3106 is the one that I'm likely to
  • 00:35:02
    have
  • 00:35:03
    installed um I can't remember it's
  • 00:35:06
    updated if it is I will leave it in the
  • 00:35:09
    uh description the version that I'm
  • 00:35:11
    actually running and then of course I
  • 00:35:13
    have the um the visual Basics installed
  • 00:35:18
    so I have python installed underneath
  • 00:35:21
    that uh I have uh C++ underneath that
  • 00:35:25
    one I would have to go back or you guys
  • 00:35:27
    can go take a look at the um the queer
  • 00:35:31
    for what I actually installed cuz I
  • 00:35:33
    mentioned it in that video uh actually I
  • 00:35:36
    think it was um automatic 11 and it was
  • 00:35:40
    for a facial recognition um feature that
  • 00:35:44
    you can add to automatic 11
  • 00:35:48
    so yeah I will
  • 00:35:50
    um I'll leave a link to download the
  • 00:35:53
    screenshot of what is installed for that
  • 00:35:57
    specific spefic
  • 00:35:58
    toolkit because they are also very
  • 00:36:06
    important so there are some things that
  • 00:36:09
    uh you want to keep in mind and I think
  • 00:36:11
    the those toolkits um those specific um
  • 00:36:15
    programs and um the development tools
  • 00:36:18
    are important for the specifics of what
  • 00:36:22
    is installed uh for certain AI programs
  • 00:36:26
    to work efficiently and properly
  • 00:36:28
    uh they're all very very fussy if
  • 00:36:30
    they've got wrong versions of things or
  • 00:36:32
    if they require certain versions they
  • 00:36:34
    tend to kick out errors and not work
  • 00:36:37
    properly now one train uh uh when you do
  • 00:36:41
    the update which I'd recommend it
  • 00:36:43
    usually goes through and it says your
  • 00:36:45
    Python's out of date uninstalling and it
  • 00:36:48
    installs newer version for what it
  • 00:36:50
    actually requires and I have not seen
  • 00:36:53
    the update of that break um the Cryer
  • 00:36:58
    which is a good thing so I would say
  • 00:37:01
    update that and let it do its thing for
  • 00:37:03
    going through uh it also comes through
  • 00:37:05
    and it mentions arguments uh for command
  • 00:37:09
    line codes to install certain things
  • 00:37:11
    that it requires if it's out of date and
  • 00:37:13
    doesn't have permissions to do so so
  • 00:37:16
    copy those and paste them down because
  • 00:37:18
    you may need to run the command line to
  • 00:37:20
    install those and we can talk more about
  • 00:37:22
    that uh if you need to know more about
  • 00:37:25
    that comment below hopefully you guys
  • 00:37:28
    understand that well
  • 00:37:29
    enough but that's Basics on how to um
  • 00:37:33
    the tips that I've gotten how I do my
  • 00:37:35
    data set to get the quality that I do
  • 00:37:39
    and I get fantastic quality extreme
  • 00:37:41
    quality for um the number of images that
  • 00:37:44
    I use the number of steps that I use and
  • 00:37:46
    the number of epoch that I use so give
  • 00:37:49
    all that a try again I will leave links
  • 00:37:53
    to my settings for the Adam West 8bit so
  • 00:37:57
    you can just download that drop that
  • 00:37:59
    into your one trainer uh I'll leave a
  • 00:38:01
    link to the uh the pack I'll leave a
  • 00:38:04
    link to the image and um I'll put as
  • 00:38:07
    much in the description as I can for
  • 00:38:09
    what I'm running there and of course I
  • 00:38:11
    will see you in the next stream table
  • 00:38:13
    this
  • 00:38:16
    video thank you for watching my video
  • 00:38:18
    and sticking around to the end if you
  • 00:38:20
    like my videos it really help me out if
  • 00:38:22
    you could like And subscribe it helps
  • 00:38:24
    the YouTube algorithm to push my videos
  • 00:38:27
    out there to more viewers which in turn
  • 00:38:30
    helps me and helps everyone so thank you
  • 00:38:33
    for watching my video and hanging around
  • 00:38:34
    to the end and I will see you in the
  • 00:38:37
    next video
الوسوم
  • AI training
  • dataset setup
  • image quality
  • tokens
  • character representation
  • one trainer
  • AI models
  • image editing
  • training steps
  • image metadata