0:00 how using AI render anything 100 times faster and absolutely for free using just your computer hello I'm Stefan and today I want to show you how I use AI in my daily routine to speed up renders like 100 times so I show you how from this to this just in 10 seconds and all of this just using your computer so so let's take a look so I will be showing you how uh I set up the comy UR flow that works together with blender and how do I render everything like 100 times faster using my computer absolutely for
0:46 free so here it is uh this is my workflow using confi so you have to pre-install confi for all the things I used here I prepared the document in otion and uh there are all the links all the custom nodes all the models everything you basically need I only not explain how to install conf but I believe like you can find it out so here it is uh there is a flow that where we can load an image uh which is absolutely flat there is no Shadows Etc and here's the scene where I take it like we'll be able to play around so you can see the scene in blender quite flat and uh no shadows and I actually intentionally disabled all the Shadows the only thing
1:33 here and you will later know why is a contrast I just added a very high contrast so we can actually distinguish the separate objects and that's the main core core core thing so and from this image doing some stuff let's say I will explain all the things later we're able to achieve this absolutely amazing result I mean and all this depends on your computer and your GPU it will take you around like 10 seconds but again like if you guys have really good GPU it can take like for three 4 seconds I mean that's absolutely insane and yeah and then you have this result and I able even to upscale it right away in my
2:19 workflow this is an amazing tool worthwhile investment of your time so let's get it started so in my file there's a few blocks uh in my workflow there's like few blocks uh and let's actually start from the very beginning so there's two nodes here basically What's Happening Here is that we take an image connected to the image input and then we ask set a task to do more detailed Caption This sort of pretty pretty standard uh settings we connect download and load floor uh which is a part of this custom nodes yeah so for Florence to model so you can see I'm already using it yeah it will have
3:08 all of this and then you can see this will basically give me the description but uh the direct description is not really what I need because the scene looks quite flat right uh there's no Shadows I mean and what I want it to do is to edit the description so so the next step will be prompt editing and what it actually does is it takes a text or in my case it's a prompt and then with some instruction it to the modification I mean like think of it as CH GPT or like CLA basically the same mechanism right
3:55 so this is more instruction I always say like okay you'll receive a prompt which is a prompt is the description of my image make it sound like it is something really cool like real nice lightning I mean something nice uh landscape uh behind the windows Etc but keep the important information about the colors because again like I want to be able to set up color in blender and uh render them like here keeping our information so yeah you will can take a look here then uh there's one more step is sdxl promp Styler also a custom node like you can
4:42 easily download it like I will explain everything here uh about installation I give all the tips just right after the prompt Styler is basically another step of editing your prompt here you know like instead of adding like like Photo realistic all blah blah blah blah blah we will just choose something from these presets in my case CU I'm doing like realistic PR this I choose R style HP realis so here we basically editing a prompt uh for our for diffusion model right so I will be here just in a moment uh but before I show you one more step that we do some imagy touch so what's
5:27 happening here is that we take an image we do a little bit resizing is or the sdxl model we're using is quite strict about the resolution you put inside right so uh it will be and we work like as image and image to image which means like we need to care about input so I just do set resolution I do a little bit of uh brightness and contrast just just slightly but again you can adjust it it depends like what images I mean maybe you do screenshot from SketchUp right so you can just do some settings here and you don't need to do uh any Photoshop Etc so a simple procedure here then I do
6:14 upscaling it is optional but I was playing around and I found out that if we do upscaling well basically what I do is not just upscaling is I'm upscale and then resize back just to do this and trying to keep as much details as possible so all these black uh notes is for preview uh they basically just show some intermediate result so this is a control net and this is is the Excel basic image generation so uh this is the main block uh let me break it down for you so first thing here is to load the model oh I believe you heard of
7:02 them and here I use Juggernaut Xcel which is basically stable diffusion Excel model Excel 1.0 model I actually attached the link where you can download it C ai.com is absolutely amazing website where we can find a lot of things related to confi I mean to all your needs the only thing you need to know that uh there is some compatibility uh issues you need to take into account like for example if you use sdxl there are some additional things like low and control net which we will cover just in a moment they need to be actually developed for that main model so in in our case sdxl is our main model
7:49 the main model has three outputs the model which is connected uh to low and then connect it to a model clip which is basically prompting thing and clip uh here uh is split into prompts the positive prompt and negative prompt and we actually get all this thing from our prompt here and there is a CR apply multi control net which is basically applying control Nets and K sampler is where we do all the magic we take all the inputs and using sdxl Create an image after that we just need to de coded and here we go here's the
8:34 result right so few things what I just mentioned is Laura and control net let's start with a control net here we go control net actually designed to give us more control well I mean obviously right uh give us more control what what's is going to be generated right what's going to be generated I create a deps image from my image using uh config Pi control net node here basically you can create a lot of them but not all of them are supported you need to check which one is supported by sdxl I create deps and canny which show
9:20 outlined version of my image which later is used as a prompt so here um is a config r no that helped me to combine multiple uh control Nets and you can see there's a switch one and this one and switch two this one and basically on on you can connect one more and they are off by the third one off not used that you just need to adjust the control net straight like how how strong this will take effects on the final image right so uh this are more or less nice vales but you can play around with them more if you need model to follow all the lines more accurate
10:06 it's not not enough accurate as you think you can add more value here and we'll you do it precisely right using like one which is 100% right not really gives us very good result and then it's connected to General throw before we do uh sampling right some basic uh setup here but what is uh really interesting is uh sfg which is how we actually Follow The Prompt uh this value is going like from the recommended settings from The Jug model so we don't really need to adjust it you can play around but the result will be not not that good uh but what you can
10:52 really play with is steps so playing like around this is how many iterations is done before you get the final image so you can play around with like 20 to 4 and find what's best for you the seed actually can randomize you can just I mean inter manual value well basically what this means seed is that like whenever you have the same value the result will be exactly the same if the input is not changed right uh all these settings also comes from the model recommended setting but the one interesting thing here is the noise when we send image here as a prompt as a latent image right so this basically it takes an image and cover with noise and the noise level is how much noise it
11:39 would actually add so the more noise the less it will recognize the original image and take some details from there but from the opposite side right less noise the more it will look like the original image or maybe it will actually not change anything because it will look one by one there just too much of effect you can play around with this failure in range of of 1.8 and 1.95 I think and see how it results in my case because I have all the information about lines and deps here uh from the control net I need image to image to transfer the to additionally
12:26 transfer uh the color information right I do that with the prompt as well but I mean Never Enough we're working with AI I found out that like value something like 92 give more precise result uh it's like it's just enough to do the job and one last thing here uh which you might be not familiar is Laura which stands for low rank adaptation so Laura is something you can use along with main model that will improve your results by teaching it on top right to do some specific task like in my case it might be Interior right you can see this is Laura and this is Laura for base model sdxl
13:14 1.0 and uh you can read uh the description that if I use it like I can expect my interior to be quite nice like because it's additionally trained to do nice Interiors but it actually adds additional prompting and teaching the model to do something maybe she doesn't do good before giving you an example if you need to do clouds and sdxl model doesn't do clouds really nice by default you can find a Lura that actually is trained to do nice clouds use it here and I mean you will have the result and the very last step uh of the whole process is upscale quite simple we load
14:01 the upscale model uh this one can be loaded directly from the manager then we do upscale image uh using that model and 0.5 means it will be increased like two times in a size in a in a resolution and yeah you can see like it really works it's it so see this original image and here's the output so let's try it out uh let's do another image and let's give it a try uh actually for this case let me do something crazy I will just just do it like this cuz again we're talking about rendering anything
14:47 right not only interior so the only thing I don't want maybe to see a black screen right so I did the screenshot of flower and let's go back and choose the file here all right and let's see what may happen but I think I will need to edit a little bit of positive prompt because it's it is not real estate photography anymore right uh let's just remove this and here that mentions the furniture specifically right describe and uh
15:32 photo realistic awesome render I mean good enough there is no windows uh oh that's enough right let's run it and's see how it goes as you can see the process here we get the description over we a view of clear glass vase on a great table blah blah blah amazing it's stunning modern living room with this pristine with walls and I guess it's still processing my bad so here we are it is you can see like it's this part is
16:20 done and you can see now it describes a render in details as actually as asked we can see here simple image plus processing like so just adding a little bit contrast yes okay so let's see how depths works right so like here's the depths there is the Cy well I guess the floor here makes some noises but let's see how it figured out so oh my God I mean this is actually more beautiful than I actually imagined the cotton
17:05 looks so real I don't like it missed information about the floor but probably need to a little bit play around with the D noise uh it has too much freedom to play with with colors but again like still and let's take a look on the upscale I mean like this just more beautiful wow yeah just adding like this small well that's it I mean I really hope this will be very useful for you guys I use it in my daily routine it saves me a lot of time you can play around with
17:51 different models you might heard of like flux uh St defusion tree or other other models you can take a look on the whole range here like quite a lot of them 3.5 quite interesting flux very promising I use sdxl here because this it has a lot of control net stuff relatively old so a lot of lures and control net exist for it for flux there's not much of them but still already have some stuff so guys also a few tips about how to install it so there is a workflow itself so the very first thing you need to do iSell conf manager when you have this done it is extremely easy just uh load the workflow
18:39 uh from here just open the workflow I I I provided then go to the manager install missing custom nodes and it will St install 80% of all nodes used in this project you will then need to close and restart the confy UI and it will install all the dependencies ETC the few things might need draw attention here is Florence and search LM If you experience any difficulties to install this model automatically just go to the GitHub and see uh the instruction how to install it and where to download the models they use for the Florence for example you don't need to download anything uh I
19:25 added the link but it should be totally fine here on the first run it will download it automatically for the third llm there is mol 7B model that you will be able to download either by my link here or you can go to the GitHub and download it from there thanks a lot guys I hope that was helpful and you learned something new I encourage you to play around again guys this is really powerful tool use it play around let me know if you have any questions and let me know like how what cool things you can do that speak later
