Ultimate Guide to Installing Forge UI and Flowing with Flux Models

finally we have an update on Forge UI it was definitely needed in this video I will show you how to install the latest

version of Forge UI then I'll install three different versions of flux the new model everyone is talking about I'll

also add the art styles in this new version and show you how to quickly revert to an older version if you miss

features like SVD and other previous functionalities uh first I will show you how to install Forge go to this page I

will add links in the description and on Discord then if you scroll down you can find the installation section this works

only with Nvidia video cards and in this video I'll demonstrate the installation on

Windows also you should know that there there have been a lot of changes in the last few days so it has been quite

unstable however I will show you later how to revert to a previous version if something doesn't work click on the link

where it says download oneclick package they had a faster version 2.4 but it's broken so uh this one with version 2.3

works for me then create a folder on your drive and save the zip archive there um as you can see I have quite a

few versions uh installed in different folders to experiment with I I might name this folder Forge flux since it

supports flux but you can choose any name you like inside this folder save the zip archive in the instructions it

says to to uncompress the file then use the update. bat to update and finally run the Run bat file so let's let's do

that you can use software like windzip or win raar to uncompress the file which might take a while once you verified

that all the files are extracted without errors um you can delete the zip file to save space uh double click on the update

file once it's finished press any key like enter next you uh you have the file which you use to start forge UI you can

create a shortcut to the desktop for easy access next time as you can see I have this commit hash here a long string

of characters that shows the exact version you're running I'll show you later how to switch to older versions uh

let it run until it finishes the first run will take longer because it installs all the requirements

and downloads a stable diffusion model so you can can start generating right away this is how the interface looks

in the top left where it says UI you'll find some buttons that let you change the interface for the specific model

you're using such as Excel flux or SD in this case we choose SD since the downloaded model is SD View onb 5 we

only have that model in the list for now here you can add a quick prompt and then click the generate button to get an

image based on that text and a random seed if you click the button again you'll get a different image in the

negative prompt section you add words that describe what you don't want in the image from here you can adjust the

sampling steps and the width and height of the image when the seed is set to minus one it's random but you can

specify a fixed seed number if you click generate you'll get an image based on the prompt and that exact seed clicking

generate again with the same seed will give you the same image unless you change the prompt which will create a

variation click on the dice icon to set the seed to random um under the image you can find

more information about the generation including the prompt all the settings used and the time taken which is about a

second in this case in this case the SD 1.5 model was trained on 512 pixel images so when generating images at 10

24 pixels it may sometimes repeat or distort features because it's not used to to handling that larger

resolution so use the recommended sizes for each model in the command window you can find more information about uh

generations and errors if you want to uh close Forge so simply close this command window uh if I try to generate now you

get a connection error so close that browser Tab and then go to the webui folder here you'll find a file named

webui user. you can right click on it and choose to edit or open it with a text editor inside this file that you'll

see arguments and if someone tells you to use um it's a specific argument ments they mean to add them to this file for

example if you add-- theme dark you can change the theme to dark mode in this file if you remove the comments that Ram

in front you can add pass to different model locations if they are stored elsewhere save the file and then let's

test it um open the Run file which will start forge you'll see that it's launching with the arguments and at the

end you'll get the nice Dark theme for Forge uh for comparison this is the default theme and this is how it looks

now there are many uh models you can use right now we only have one and if you go to the civit AI website and click on

models you'll find all kinds of models a popular sdxl model for example is Juggernaut which I've used in many of my

videos if you go to the filters you can select the base model including checkpoints Laura and various types even

the new flux models let me show you where I got the flux models for this video from this page you can download

the flux Dev version which is the floating Point uh version it's 17 GB in size and you can download it using this

button this version is for those with newer video cards that have more vram but it provides very good image quality

this file should be placed in the webui folder then in the models folder and finally in the stable diffusion folder

if you download Laura or text encoders you should place them in their respective folders there are categories

for different types of models but checkpoints like SD sdxl and flux go into the stable diffusion folder um you

can see we're in the right location now because we can see the existing SD model here so save the flux model in this

folder as well there's another Dev version on this page called the nf4 version which is a bit smaller at 12 uh

gabt this version is likely suitable for video cards with 12 GB of vram or more to have a decent speed for your

generation they keep releasing new versions and right now version two is the best download and save this model in

the same location um on this page there's the flux Schnell version which is a faster model while the image

quality might not be as high as the dev versions it's quick and produces acceptable results they keep releasing

new versions but I'll download this one called BNB nf4 uh you can also try versions like b or Salto but I'll be

using this version for now so I'm downloading it and placing it in the same folder to recap we have the dev

floating Point 8 version the larger model with high quality results the dev nf4 version similar quality to fp8 but

smaller in size uh the Chanel version medium quality but faster all flux models require a good video card

compared to sdxl or SD models after the download is complete if your interface face was open you might not see the new

model in the list click on the refresh button and now you should see all three downloaded models let's start with the

best one the floating Point 8 version since we're using a flux model you'll want to switch to the flux interface if

you're wondering how it differs first uh it doesn't support a negative prompt so the CFG scale needs to be set to one it

also suggests some recommended settings with values around 10 24 pixels similar to the sdxl models flux can handle even

higher resolutions but it will take more time to generate the images let's add a prompt and and give it a test to see

what we get that's a cute cat isn't it um the first generation always takes more time because it needs to load the

model on my RTX 490 it took 28 seconds the first time but on an RTX 2060 it took even 9

Minutes the Next Generation should be faster so if I try again you can see that it now takes only 143 seconds you

know it's not as fast as sdxl but it's still a good speed for the quality it offers if you have a good video card

some people get better speeds with different options on the top like various combinations swap methods or

swap locations try them all to see what works best for you and for me the auto setting worked well uh just remember

that the a first generation always takes longer so I make sure to note the speed of the second generation you know after

you make any changes not the first uh you should visit this page to find more information on how different

options work as mentioned they're they're still working on it so some things might not work or they might work

now but change with it new versions at the end of the page you can see discussions where other people share

their problems and solutions or you can ask your own questions and maybe get some help in this video I covered all

the steps I took and sped up only the loading part so it worked for me but it might not work the same way on your

system check this page to possibly find a solution if you you have a model that requires a serate vae file placing that

file in the models and vae folder will make it appear here uh let's create a 10 24 pixel Square image as you saw with

the SD model it generated two heads uh for the cat but you know with flux we don't have that problem if you're using

the flux Dev version make sure the steps are set between 20 and 30 keep in mind that more steps will take more time and

after a certain point increasing the number of steps may not necessarily make the image look better it might even make

it worse they say it supports different resolutions and sizes up to 2 megapixels

a megapixel equals 1 million pixels so 2 megapixels equals 2 million pixels the total number of pixels in an image is

calculated by multiplying its width and height in pixels for example 1920 multiplied by 1080 is slightly over 2

megapixels here are some tests I did this one looks quite sharp a slightly larger size also looks good and can add

more details I tried even larger sizes still under 2 megapixels and they looked okay at the limit of 2 megapixels the

image wasn't the sharpest for this seed but I got lucky with another seed and still got a nice image however larger

sizes resulted in less Sharp Images the good news is that I didn't get duplicate heads or poor composition and some

images still looked quite good depending on the prompt keep in mind that at such large sizes the generation time

increases and the quality might not be as high for this seed at full HD size the image was good but with the second

seed it wasn't as great the same issue occurred with portrait sizes that went over 2 megapixels so keeping the

resolution under 2 megapixels usually yields nice results if if not try adjusting the prompt seed or Samplers

I'm selecting the second Dev model the one with nf4 and the name keep in mind that since it's a Dev model it's also

needs 20 steps to work well the first generation takes longer but on my system the second generation was faster than

the floating point8 model by half a second now let's test the Schnell nf4 version this one only needs four steps

to run so I will interrupt the Generation Um it takes a moment to stop and you can see where it stopped at the

beginning where there was still noise uh I'll set it to four steps but you can go up to eight steps if you prefer here's

the result um as usual the next Generations are faster this one is about four or five times faster than the dev

version taking only 3.5 seconds in the settings tab you can adjust various options such as the path to save the

images the format and file name for saving and other preferences if you go down to the other section you'll find

defaults you can click view changes to see how the current settings differ from the default ones click apply so that the

next time you open the interface it will use the same settings you have now and I have an older video where I discuss

these settings in more detail uh this version of Forge also comes with some Styles you can use it says Styles

integrated so you can choose one like the 3D style when you run it you get a 3D render style that's quite cute isn't

it these Styles should work with all models although sdxl is familiar with more art styles than flux still applying

these Styles will have an effect even if some art styles are not fully recognized if you've seen my other

videos you probably saw that I created over 300 art styles that work with both comfy UI and Forge UI these are positive

and negative prompts combined with your prompt to achieve a specific style you can click download and save them to your

web UI folder you'll know you're in the right directory if you can see this the integrated Styles already there um both

should be visible now if you have an older version of Forge you might not see these integrated Styles but this is the

folder where you should place the CSV Styles file if Forge was open when you downloaded the file you might not see

the new Styles in the list close your command window for Forge and run it again to see the updated list now my

Styles will appear first followed by the integrated ones you can easily recognize my Styles because they have

a category first followed by a vertical bar and then the style name Let's test it with a prompt using a flux Dev model

remember the steps should be set uh between 20 and 30 for the dev model I'll generate an image first without applying

a style so we can compare here's the result with the cat I'll reuse this seed so when I click

to generate again it's no longer random each time you change the prompt or use the same prompt with a different style

the generation will be different uh even if you use the same seat um for cinematic portrait style I got the same

cat but with a more cinematic look if you click on this folder it will open the output directory uh sometimes it

remains on the taskbar so if I open that folder I can see a folder with today's date inside containing all the images

generated with text to image this way you can compare the results before and after applying the Cinematic style you

can search for a word like cute uh and it will show you all the styles with that word in the name for example if I

test uh uh cartoon cute I get this cute cat you can also search by category or scroll down to see them organized by

type such as photography drawing 3D design and art if I choose dark Academia I get a

style reflecting that theme you can also combine it with um another style like cute resulting in a combination of the

two Styles this way you can create unique Blends based on different style influences when I tried to create a

vector design I ended up with a blurred image this issue occurred only with the dev model and not with the chel model

one solution I found was to try a different sampler and scheduler for example using DPM Plus+ a 2m uh with sgm

uniform helped me get a clearer image uh but with other prompts and seeds I encountered the same problem so

I I tried another sampler like one uh while this one fix the image it is much slower than other Samplers I'm only

changing to one or similar Samplers if a specific image doesn't work and I don't have any other options so different

Samplers can produce different results and vary in speed I also have some nice sculpture

Styles in the craft category I'll close Forge and show you something uh on this page you can find

commits commit refers to a specific update or change made to the forge projects codebase on GitHub each commit

represents a point in the Project's history where modifications were made such as adding new features fixing bugs

or updating documentation so uh the ones that are marked as verified are more stable and if you click on that button

you can copy the full SSH when you paste it into notepad you'll see a long string of numbers and

characters so if you scroll down and you'll see that more pages with commits uh that string of characters is

important because it allows us to get the exact same version if we encounter a bug we can revert to a previous stable

version and return to the main version once the issue is fixed if you watched an older video you'll see that it

features an earlier version with SVD and other elements that the new version doesn't have however it also lacks flux

in the new features so how do we revert to a different version I found a method on a Chinese site that is quite simple

in notepad paste this text uh you can get it from Discord this script sets the path to the git binary navigates to the

web UI directory and checks out a specific commit that long string using git this effectively reverts the code in

that directory to the state it was in at that commit for example when I recorded this video this was the latest verified

version with a commit starting with aad however by the next day many more edits had been made and many of these are not

verified if I keep going I can eventually find the version from yesterday that worked for me so if you

watch this video and something doesn't work as it does for me one possible reason could be that you have the latest

version because you clicked on the update bat file the latest version might have some bugs which is why it can be

useful to revert to a previous version so the commit I added here is quite old but worked well for me it includes uh

the SVD tab that people kept asking about you can try different newer versions to see what works best for you

go to file and click save as then navigate to the folder where you have the run and update files name it

rollback dobat and set the save as type to all files click save and it should appear in the folder now if you run that

rollback file it will indicate that it's switching to that commit version you can press enter to continue this process is

similar to changing a train's route to make it go in a different direction so uh now we can click on run to open Forge

you'll see it's now running the commit I specified and it will download all the necessary components to make it work the

first time might be slower but after that it should like start up pretty quickly um so well now we have an older

version which includes the SVD Tab and other tabs that the new version didn't have at the time of this um recording uh

this version works well with SD and in sdxl models but doesn't support flux models if I select a flux model and try

to run it you'll see an error in the command line this indicates that only the latest version of Forge is

compatible with fux models so how do we switch back to the main version to get the latest updates open notepad again

and paste this text inside now go to file save as and place it in the same folder as rollback and

update for save as type choose all files and name it something like roll forward dobat then click save now if we run the

roll forward bat file you'll see a message indicating that the previous commit with 29 has been switched back to

the main branch press any key to close that message now if I run Forge again you'll see that it's uh back to the main

version the version I used when um I installed an updated Forge you can see that this new version includes flux and

other updates but no longer has the SVD tab uh and some of the other tabs from the older

version if you go to the PNG info tab you can drag or upload an image you generated with Forge I'll pick one from

the output folder where the images are usually saved on the right you'll see the information like prompts and

settings used for that image from here you can send it to text to image which will add the prompt uh including Styles

and the same seed so if you generate it again you'll get the exact same image the result is this cool portrait you can

also make the image bigger by clicking on highr fix make sure to adjust the settings for it here I'll send it to

image to image where you can almost replicate everything except for Styles if you want a different style you can

add that here but I'll keep the same style to get a larger image for the size I'll keep the same ratio but increase it

to full HD I can do this because I have 24 gigabytes of vram but you may need to use a smaller

size if it crashes for Deno I'll use 0.35 and for seed I'll set it to random to generate multiple variations running

it now will give you a larger image in full HD with nice details so flux is quite capable with the right prompts

let's do some comparison between models to wrap up this video I added the model name on top and the number of steps on

the bottom the Schnell model is clearly the fastest of all followed by the dev nf4 which is slower and then the dev

floating point8 version which is slightly slower depending on your system for realism in photography the chel

model may not be the best but it works quite well for illustrations it tends to add more contrast and vibrant colors

compared to the other models and is fairly good with text especially when compared to sdxl the two Dev versions

produce very similar results but the floating point8 version offers more detail and accur accuracy following the

prompt a bit better if your video card can handle it you should test both models to see the difference in speed

for me the difference is only half a second so using the nf4 version may not be worth it when the floating point8

version provides slightly more detail if your video card is weaker try the Schnell version if it still can't handle

it try the sdxl models and if that's still too much go with the sdv one five models with an Nvidia card you can

generate images for free on your computer it's been a long week and I need a break thank you all for your

support see you on Discord or in the Facebook group leave a like or a comment if you found something useful have a

great day [Music]