A1111 - Best Advice for amazing Stable Diffusion Images
Summary
TLDRThis video tutorial offers valuable tips for achieving stunning AI-generated images. It guides viewers on selecting the best models, utilizing negative embeddings, and leveraging various models and lora adjustments for optimal results. The host also explains how to use the interface effectively, including adjusting sliders for clip skip and choosing the right upscale models for high-resolution fixes. Additionally, the video covers the use of extensions like 'a detailer' for face tracking and 'control net' for body tracking, providing a comprehensive guide to enhance image generation.
Takeaways
- 😀 Choose the right AI model for better results based on community ratings, hearts, and downloads.
- 🔍 Read model descriptions and prompts to understand how to effectively use them for desired outcomes.
- 📚 Utilize negative prompts and embeddings to exclude unwanted elements from AI-generated images.
- 🎨 Explore different models like Realistic Vision, ref-animated, and deliberate for various artistic styles.
- 🔧 Understand the importance of settings such as sampler, CFG scale, and high-res fix for image quality.
- 📸 Use the 'after detailer' extension for automatic face tracking and higher-resolution in-painting.
- 🌟 Discover and apply settings from successful images by checking prompts and parameters used.
- 🔗 Recognize that models, lora styles, and embeddings must be compatible with the base model version you are using.
- 🛠️ Adjust the weight of lora styles to prevent overemphasis and achieve a balanced result.
- 🖼️ Use the image-to-image tab for making subtle changes to existing images, like altering ethnicity while keeping the same composition.
- 🔧 Employ extensions like 'a detailer' for face tracking, 'control net' for pose tracking, and 'Roop' for face replacement to enhance image detail and customization.
Q & A
What is the main purpose of the video?
-The main purpose of the video is to share tips and tricks for achieving amazing results with AI, covering both basics for beginners and advanced techniques for experienced users.
Why is choosing the right AI model important for the outcome of the AI-generated images?
-Choosing the right AI model is important because a better model generally leads to better results in terms of image quality and accuracy. The model's ratings, hearts, and downloads can indicate its popularity and effectiveness within the community.
What are 'negative embeddings' and how are they used in AI image generation?
-Negative embeddings are small models trained on content that the user does not want to appear in the image. They are used in the negative prompt to ensure that the AI-generated image avoids certain elements.
How can one find suggestions for using a specific AI model?
-Suggestions for using a specific AI model can be found by scrolling down on the model's page, where there is information on the best positive and negative prompts, as well as other settings like the sampler and CFG scale.
What is the role of the 'after detailer extension' mentioned in the script?
-The 'after detailer extension' is used for automatic face tracking and in-painting the face in higher resolution, enhancing the detail in the facial features of the AI-generated image.
Can you explain the significance of the 'CFG scale' in AI image generation?
-The CFG scale determines how closely the AI sticks to the provided prompt. A lower value allows the AI to take more creative freedom, while a higher value makes the AI adhere more strictly to the model's training data.
What is the 'high-res fix' and how does it improve the quality of AI-generated images?
-The 'high-res fix' is a feature that allows users to upscale the resolution of an AI-generated image for better quality. It involves using specific upscale models and adjusting settings like denoise strength to enhance image details.
What are 'loras' and how do they influence the style of an AI-generated image?
-Loras are smaller versions of models that can influence the style of an AI-generated image. They add specific details or characteristics to the image, such as additional sharpness or artistic flair.
How can one adjust the settings for the 'image to image' tab in the AI software?
-To adjust the settings for the 'image to image' tab, one can change the prompt, adjust the denoise strength, and use various extensions like 'a detailer' or 'control net' to modify specific parts of the image.
What are some tips for using the 'in painting' tab to modify parts of an AI-generated image?
-In the 'in painting' tab, one can use a brush tool to paint over areas they wish to modify. It's important to adjust the render size to match the area being modified and to set an appropriate denoise strength to allow for changes while maintaining image quality.
How can one ensure that the AI-generated image matches the desired ethnicity or facial features?
-To match the desired ethnicity or facial features, one can use the 'image to image' tab to adjust the prompt accordingly and set an appropriate denoise strength that allows the AI to make the necessary changes while staying close to the original image.
What is the purpose of the 'Roop' extension in AI image generation?
-The 'Roop' extension allows users to replace the face in an AI-generated image with a photo of a real person. This can be useful for creating images that have a specific person's likeness but may require the original image to be close to the body and head shape of the person for a natural result.
Why is it recommended to use a lower batch size or batch count for users with slower computers or older GPUs?
-Using a lower batch size or batch count helps prevent overloading the computer or GPU, as it allows images to be rendered one at a time or in smaller groups, improving performance on less powerful systems.
Outlines
🤖 Choosing the Right AI Model for Amazing Results
The video script begins by emphasizing the importance of selecting the right AI model to achieve the best results. It suggests looking at community ratings, hearts, and downloads to identify popular and effective models. The script introduces the concept of models and luras (smaller model versions influencing style), and provides tips on using specific models like 'Realistic Vision' and 'ref animated'. It also discusses the use of negative embeddings to exclude unwanted elements from AI-generated images and mentions the significance of choosing the right sampler and CFG scale for high-quality outputs.
🔍 Navigating the Interface and Enhancing Image Quality
This paragraph delves into the user interface of the AI software, focusing on adjusting settings like clip skip and selecting appropriate vae models. It explains how to add quick settings for clip stop and sdvae, which are crucial for customization. The script also covers the high-res fix feature, which improves image resolution and quality using upscale models like '4X Ultra Sharp'. The importance of setting the right denoise strength for image rendering is highlighted, and the process of placing models, luras, embeddings, and vae files into the correct folders within the software is outlined.
🎨 Working with Models, Luras, and Embeddings
The script discusses the compatibility of models, luras, and embeddings with different versions of AI software, such as SD 1.5 and sdxl. It explains the process of adding negative embeddings to the prompt field and adjusting the weight of luras to achieve better results. The paragraph also covers the use of the CFG scale to control how closely the AI adheres to the given prompt, and introduces the image to image tab for making subtle changes to existing images, such as altering ethnicity while maintaining the same composition.
🖌️ Advanced Image Editing with Inpainting and Extensions
This section introduces advanced image editing techniques available in the AI software, such as the inpainting tab for making detailed changes to specific parts of an image. It explains how to use the 'only mask' feature and adjust render size for局部 image re-rendering. The script also highlights various extensions that enhance the AI's capabilities, including 'a detailer' for face tracking and replacement, 'control net' for body tracking, and 'Roop' for face replacement using photos. Additionally, it provides tips for users with slower GPUs or older computers on how to use tiling for upscaling images without overwhelming their hardware.
🛠️ Optimizing AI Image Generation for Different Hardware
The final paragraph offers advice for optimizing the AI image generation process based on the user's hardware capabilities. It suggests using the 'batch count' and 'batch size' settings to manage the rendering of multiple images, which can be beneficial for those with slower or older GPUs. The script also mentions using the 'SD upscale selected' script for tiling-based upscaling to减轻GPU负担. The video concludes with a call to action for viewers to leave a like and comment if they have further suggestions or enjoyed the content.
Mindmap
Keywords
💡AI Models
💡Prompts
💡Negative Embedding
💡Sampler
💡CFG Scale
💡High-Res Fix
💡Loras
💡Extensions
💡In-Painting
💡Batch Processing
Highlights
The importance of choosing the right AI model for better results based on community ratings, hearts, and downloads.
Understanding the use of models with detailed instructions and suggestions for positive and negative prompts.
The role of negative embeddings in refining AI-generated images to exclude undesired elements.
The significance of samplers and CFG scale in achieving high-quality AI image generation.
How to utilize the high-risk fix with specific models for enhanced image details.
Insights on using extensions like After Detailer for automatic face tracking and inpainting.
Recommendations for models like 'ref animated' and their compatibility with different VAEs and negative embeddings.
The application of different models such as 'Deliberate' for photorealism and 'Epic Realism' for natural lighting.
The concept of 'loras' as smaller model versions that influence the style of AI creations.
Guidance on matching the correct version of models, loras, and embeddings for optimal AI performance.
How to navigate the Automatic1111 interface for effective use of additional parameters like clip skip and VAE model selection.
The process of implementing high-res fix for image upscaling using recommended models like '4X Ultra sharp'.
Instructions on where to place different AI components such as models, loras, embeddings, and upscalers in the Automatic1111 folder structure.
The use of 'textual inversion' for negative prompts and how to integrate it with AI models.
Adjusting the weight of 'loras' to prevent overemphasis and achieve a balanced AI-generated image.
The function of the 'image to image' tab for making subtle changes to existing AI images.
How to use the 'in painting' tab for modifying specific parts of an AI-generated image.
The integration of extensions like 'a detailer' for face tracking and 'control net' for body tracking in AI image generation.
The use of the 'Roop' extension for replacing faces in AI images with real photos.
Strategies for handling AI image rendering on older computers or with slower GPUs using tiling and batch processing.
Transcripts
today I want to share with you some of
the best tricks to get amazing results
with AI this video is great for
beginners but there's also a lot of
stuff in here for the pros so let's get
started to get amazing results you want
to make the chop for you as easy as
possible and this starts with the choice
of the model because the better the
model is the better your results will
look of course there's a lot of
different models here so when you go to
a Civic AI you can also see here how
many ratings does it have how many
hearts how many downloads does it have
so that you can see which are the most
beloved models by the community those
are often working the best of course
there's different categories I want to
show you some of my favorite models for
example realistic vision is one that I
really like and here already I want to
show you something that is really
important to understand how to work with
such a model so for that you want to
scroll scroll down and here you have a
lot of information of how to use the
model for example here is a suggestion
for the best positive prompt all of
these different words that you want to
use next to the description of what you
want to have as the image then there is
also here a suggestion for the negative
prompt in here on what you want to use
for all these words so that you get the
best out of that model there's two
different suggestions in here so that is
pretty awesome and you can also see here
in Orange there is a negative embedding
this is a very small Model A textual
inversion as it is called this is
trained basically on the stuff you don't
want to have in the image and that is
then put into the negative prompt so
here you can click on that link and this
will then go to the page where you can
download the negative embedding we will
talk about it in a minute where to put
that then when you scroll deeper here
you can also see you have here for
example the sampler you want to use so
either Euler a or DPM plus plus sde
carers below that it says CFG scale of
3.527 is good and then also for the high
risk fix use the 4X Ultra sharp model
for that again I'm gonna show you in a
second where to put that there's a lot
of information here that you can use but
there is even more information that you
can use so when you scroll up you can
see these images and you can check out
the images that you like for example
this is really nice with the neon light
in the background so click on that and
when you click on that image you can see
here on the right side The Prompt that
has been used the negative prompt that
has been used the sampler the model the
CFG scale the steps that have been used
and the seed you can see here it says in
the first comment the after detailer
extension was used to create this image
that extension is something that does
for example automatic face tracking and
then automatically in painting the face
in a higher resolution so that's also
very nice I have a video about that
right here to check out that extension
other models I like are for example ref
animated now again you want to scroll
down and here on that page you'll find
something really useful so when we
scroll down again you find here a lot of
interesting information but on top of
that you see that this model works well
with the vaes orange mix and the klf8
anime 2 model and then also the Blast 2
vae so that you can use that below that
you find here the negative embeddings
that I am so fond of because they are
really helpful here's a very nice list
of them you can click on these links to
download them again I'm going to show
you where to put all of that in a second
so you really want to check out all of
that information and for each of these
models you want to scroll up click on
images you like here to look what kind
of settings prompt negative prompt has
been used to create that and you want to
start with this information and then
change the prompt to work your way into
that model to get amazing results right
from the start now there's other models
out there like for example deliberate
and this can be trained on different
styles for example this can do
photorealism and digital painting at the
same time there is also epic realism
which is really good has a little bit
more natural looking light so that's
also very nice and then another thing
that is really important is different
lauras you want to use now what loras
are is that they are smaller versions of
models and they can basically influence
the style that a model can create one
that I like really much I have a video
about that here is the XL more Artful
Laura XL also means that this Laura is
supposed to be used with the sdxl model
you can see here on the right side in
the information the type is Laura the
base model is sdxl 1.0 so that means
this does not work without SD 1.5 model
and it does not work with a St 2.x model
it has to be a sdxl model on the other
hand here we have another Laura it's
called a detailed tweaker now what this
does is that it adds some more details
to the image so that can be very useful
but this one as you can see on the right
side is for the base model SD 1.5 so not
for 2.x and not for sdxl and you will
find that most models and most lures at
the moment on cvdi are trained for SD
1.5 they have a lower resolution but
often still have a very good image
quality so you can absolutely start with
SD 1.5 and get an amazing results out of
that often even better than sdxl at the
moment because like I said there are so
many different loras for that so many
different models that are already
trained for that now let's have a look
at the interface of automatic 1111 what
are important things you might want to
know here so first of all we have up
here a slider for clip Skip and we have
a Chooser for the vae model you can see
I have for example orange mix and I have
the vae 840 000 and also 560 000 so I
can choose them here so they fit to the
model and I can play around with that
where do you get these sliders well you
want to go to settings and then on the
left side you want to scroll down until
you find user interface you click on
that you scroll down a little bit more
until you find quick settings and in
here you can add these additional
parameters so when you click in an empty
space you get a long list from where you
can can select these different editions
and what you want to add here is clip
stop at last layer and also
sdvae so these two choices when you add
them they will add these two additional
choices up here and then you want to
click of course on apply settings so
that the settings are saved and then
when you reload the UI you have these
two choices up here another thing this
is really important for good quality is
the high rest fix because usually with
the SD 1.5 model you are for example
rendering with 512 by 768 so with the
high res fix if you click on that this
is opening you have some additional
settings here you want to use some good
upscale models so I suggest to you the
4X Ultra sharp model the 4X nmkd model
and also the 8X nmkd model both of them
are super scary male versions I will
link all of that below the video and
then of course you want to upscale two
times and you want to set the denoise
strength between 0.2 or 0.35 not too
high because then it changes the image
too much although you can also
experiment with that now I want to show
you the difference this is an image I've
rendered with just the 512 by 768
resolution and as you can see the face
doesn't look too good the other details
are kind of okay but overall the image
could be better now here I have the high
res fix and this as I think we can all
agree looks amazing the face is fixed
everything looks beautiful this is not
using restore face this is just using
the high-res fix in that case I'm using
the 4X nmkd model so at this point we
have talked a lot about different models
and loras and embeddings where do you
put them in the automatic take 11 11
folder now the easiest one for that is
the negative embeddings these go into
the embeddings folder in here so you
just want to double click and then you
save them in here you can see I have
already a bunch in here then for the
models you want to go into the models
folder in here
and in there there's two important
folders there is this stable diffusion
folder this is where all of your models
go for normal rendering and then up here
you have the Laura folder this is where
all of your loras go so that's very
important other folders that are
important in here for you is the es
argon folder this is where all of the
different app scalar models go and then
down at the lower end you find that the
vae folder this is of course where all
of your vae models go there's another
important folder for you this is for the
control net extension because that has
its own models for control net so for
that you want to go into the extensions
folder and then in there you want to
look for the SD web UI control net
folder and in there you want to look for
the models folder and this is where all
of your control net models go I'm going
to provide a link under the video where
you find a lot of different appscaler
models here for example you can see the
the 4X Ultra sharp upscaler when you
scroll down a little bit more you see
here the 4X nmkd super scale model is
different versions of that and then also
down here you'll find the 8X nmkd super
scale model so that you can experiment
with these upscales to see which of them
give you the best results now when we
talk about all these different models
and lures and also control that models
the important thing to remember here is
that they have to fit the version that
you want to use so if you want to have a
SD 1.5 model you also need to have an SD
1.5 Lora an SD 1.5 negative embedding
and also for control net you want to
have a SD 1.5 control net model and of
course when you're working with sdxl the
Lora has to be sdxl and the control net
model also has to be sdxl now this is
not true for the vae and this is all
also not true for the upscaler that you
use in high-res upscaling these models
can be used on any model now let's look
at some other interesting facts here you
can see down here I have a tab for
Generation but next to it there is
textual inversion and there's also Laura
textual inversion is the embeddings and
this is mostly used for the negative
embedding so when you click on that you
can see here I have a list of the
negative embeddings that I can use you
want to click here into your negative
prompts field and then click on one of
these negative embeddings and as you can
see here this is added now this does not
use any kind of bracket or any kind of
other information it is just simply the
name of the negative embedding this is
however different for your Laura so now
we're on the Laura tab you want to click
into your positive prompt and now when I
click here for example on ADD detail you
can see that I have the pointy brackets
but also at the end I I have a weight of
1 here now often what you want to do
here is to maybe try this out to lower
it for example to 0.8 or 0.7 as a weight
the reason for that is that often these
lures are over trained so they are
actually too strong to be used and you
want to reduce the weight to get a good
looking result with that also you will
sometimes find that the lower you try to
use does not want to play well with the
model you're using so to combat that to
still get good results there are two
ways you can try to fix that one is to
have higher amount of sampling steps
here than you would usually use so for
example use 40 or 50 steps and another
thing you can try is down here you have
the CFG scale so for that you want to
lower that a little bit for example to
five four or even three to see if this
will improve the image if nothing helps
the Laura just doesn't want to work with
your model also here's a little
explanation for the CFG scale and what
it actually does the CFG scale means how
close the AI is sticking to the prompt
that you are writing so when you have a
lower value of that the AI is taking
more freedom with the image but it can
also often create better results because
the AI can stick closer to what the
model was trained on when you set a too
high value for example 14 15 16 you
might get a lot of problems with the
results in the image and it might not
look good now here's another interesting
thing I want to show you on the right
side you can see we have here these
colorful buttons and one of them when
you Mouse over assess image to image tab
so when you click on that the image and
all of the image information is sent to
this image to image Tab and this image
to image method is actually surprisingly
powerful so here I have the image we
have seen before now I'm changing the
prompt up here instead of writing
beautiful French woman I'm writing
beautiful black African woman and I'm
sitting down here my denoise strength to
0.5 now again for the denoise strength
the lower the value is the closer it
sticks to the original image the higher
the value is the more it looks different
from the original image so with 0.5 it
stays relatively close but also allows
the AI to make changes and as you can
see here we have a beautiful French
woman on the left side and then when I
click here on the image I have as a
result we have a beautiful black African
woman as a result she's wearing the same
clothing it is the same scene with the
same background so that you can do
subtle changes in the image for example
changing the ethnicity but the
composition and the rest of that stays
the same and that can be a very powerful
tool to to adjust your image and play
around with an image you already have
and you like but you want to improve on
that so image to image is really
powerful to get better results another
thing you can do here is to use this
button that looks like a little
painter's palette to send it to the in
painting tab now the in painting tab is
also in the image to image tab but down
here you can see you have different tabs
and one of them is called in paint you
can see again I have the image here so
as you can see on the right side here we
have an icon for a paintbrush I have a
slider here for the size of the brush
and with that for example I can paint
out the face to replace it or I can put
sunglasses on it or I can have an older
person anything I want to do also for
any other part of the image now to make
this work and this is really important
you want to scroll down here and you
want to click here on only mask and then
also you want to adjust here the render
size so often for example in this case I
have something that is almost round so
for the render size I would set for
example 512 by 512 or 768 by 768 because
this is only rendering that part and
then in painting it into the existing
image so don't use too high resolutions
here because it is going to lead to bad
results also another thing here is
because we want to change the image you
need to have the denoise strength high
enough to enable that you want to play
around with that but for example a value
between 0.4 and 0.75 is good with that
to see how soon can I actually change
that so of course the closer the changes
to the original image for example I want
to have the same face but I want to have
a angry expression for the face I can do
with a lower value of the denoise but if
I want to have a more intense change of
the face for example making it the face
of a 70 year old woman you want to have
a higher denoise value to be able for
the AI to actually create that kind of
image now at that point let's go back to
our text to image Tab and we want to
scroll down because I want to show you
some of the most useful extensions you
want to have now one of them we already
talked about is called a detailer and
here for example you can track the face
the hence the person in total and also
more choices for the face now often this
works good here for the face what I can
do here for example is I put a Laura in
here for example here I have a Laura for
Nicki Minaj and then I can just replace
the face with Allura of Nicki Minaj onto
that body now the reasoning for that the
why you want to use this extension is
you want to use the Nicki Minaj Laura in
your main prompt with a low value for
example 0.5 so so that the face is
attempted by the AI but also the body is
fitting the face of the character but
then on top of that you still give
enough freedom for the model that you
are using so it's not overwhelmed by the
Laura so you get the full quality of the
model but then also you get the body
shape and the face shape the head shape
from the Laura now afterwards in the
second step with the a detailer which of
course you need to enable here this is
then rendering only on the face the face
of the character you want to have so
that you have the full quality of the
model on the rest of the image and then
mainly the lower quality on the face of
the person so it looks as much as the
person as possible another important
extension of course is control net with
that you can do all kinds of tracking I
have multiple videos about that that you
should check out you can use as second
extension with that so here you can see
I'm using open pose and this is tracking
the body of the person on the left side
in my preview here but I have another
extension called the open pose editor
installed this gives me this little edit
button here so when I click on that I
have here now a preview of everything
and I can move these points around to
adjust them if I feel like this tracking
hasn't done a good job or I want to
change the pose a little bit or there
are some body parts missing so you can
see down here I have person one I can
scroll down and there are some
informations here like the left hand the
right hand that have not been tracked so
I can add them to the image I can post
them in here and then when I'm happy
with everything I click up here on the
send post to control net button so let's
do some changes here and put this arm up
here like so
and then when I send this you can see
that now this has updated down here and
this can now be used to generate the
image so this is very very useful as an
extra extension another extension you
want to use here is the Roop extension
again I have a video about that here now
what this does is that you can use any
kind of photo and then the AI will use
the information of that face of the
person to render it onto the AI image
now the good thing here is that you only
need a single image to put it onto the
body onto the face of any kind of AI
image that you have rendered the
downside of that of course is that the
body type the head type and the rest of
the person does not match the face or
the body that the person actually has
because it's only replacing the face so
you might want to try to render an AI
image it is as close as possible to the
body of the person and also the head
shape and then use Roop to put the face
on that because otherwise it's gonna
look a little bit strange as a result so
here you can see an example of rupe in
action this is the original image with
the original face of the character and
here we have the same image but now we
have rendered the image of the other
character on top of that the rest of the
image as you can see is completely the
same now here's another useful
information especially for people with
slower gpus older computers when you
send your image to image to image to
upscale it or upscale it even further
what you can do instead of using the
size down here is that you leave this as
the original size you scroll a little
bit lower you go here to script where it
says SD upscale selected from that
pop-down list here and what this is
doing is it is creating tiles and this
is also used using the app scalers that
we have downloaded before words for
example you can see here the 8X nmkd
super scale upscaler this is using a
scale factor of two and what this is
doing is it is splitting it up into
smaller tiles so that you render the
images one after another and afterwards
they are combined into one big image and
last but not least I have another advice
here for you for the text to image tab
but also for the image to image tab when
you scroll down here you have batch
count and batch size this allows you to
render multiple images or image
variations at the same time now what
batch size is doing is that these images
are rendered parallel at the same time
if you have a slow computer and older
GPU that might be a problem for you on
the other hand when you use the batch
count these images going to be rendered
one after another and that should help
you with a slower compute or an older
GPU let me know if I left out any
amazing advice that I should put in
another video like this leave a like if
you've enjoyed this video thanks for
watching and see you soon bye
oh you're still here so uh This is the
End screen there's other stuff you can
watch like this or that's really cool
and yeah I hope I see you soon uh leave
a like if you haven't yet and well um
yeah
Voir Plus de Vidéos Connexes
FaceFusion: The Definitive Deep Dive & Walkthrough - Everything You Always Wanted to Know About...
How to Make Ai Influencer | Ai Influencer kaise banaye 2024
SDXL Local LORA Training Guide: Unlimited AI Images of Yourself
Mastering Leonardo AI: A Comprehensive Step-by-Step Tutorial for Beginners
The Ultimate Guide to A1111 Stable Diffusion Techniques
2024最新AI绘图教程,如何使用Leonardo ai创建人物一致性角色,AI 3D卡通人物生成,超详细教学,新手轻松掌握,ai制作卡通动画图片,怎么把图片转换视频
5.0 / 5 (0 votes)