REAL SKIN Lora - make your AI Images look like real Photos!

Video Statistics and Information

Video
Captions Word Cloud
Reddit Comments
Captions
create more realistic skins with this Laura hello my friends how are you doing today I want to show your lord restraint by one of my followers which is pretty amazing and this helps you create more realistic skin details so here let's look at an example so this is what you usually get from an AI image and as you can see here the skin is beautiful but it is too perfect it looks like a digital painting it has no blemishes no nothing in there it's really nice porcelain skin and just perfect now here I have used the Laura the rest of the prompt is exactly the same the seed is the same all of it is the same but now if I zoom in you will see that here now we have a little bit of blemishes here we have these smiling wrinkles of course I would say there's even some popped blood vessels in here things you would normally have in a face and that also create the character of the purse person I also feel with this Laura even the hair is more realistic also when you look down here for example at the neck again we have here these blemishes on the skin we have some wrinkles here that are from just living a good life and smiling and being active and going out for parties and so on so this actually gives you a nicer result now one of the downsides here is because you can't use face restore the eyes don't look so good but personally I would say you can fix that within painting afterwards here's the very same image and I have replaced the eyes now to be honest they look a little bit digital I could have done a better job but I was more focused on putting this video out to you but let's look at how this actually works and what settings I'm using so in this video you're going to see me use three different loras now this is for the skin and hand now I've tested this mainly for skin this is the low Aura this video is about but in the description text by the way there's also here linked the video about Photon from me but uh yeah never mind that there is also another Laura here that is specifically for eyes the reason for that is because this Laura is best used without restore face because restore face is going to take away all of the advances that this Laura is adding to the image so when you click on this link you come here to hugging face and there you have this safe tensor file that you can use I did not use it in my testing however you will see me use another Laura this is called Low era and this is to create these kind of low light situations I use it with a very soft strength just to get a bit more interesting light in there you don't need it but it might help you get better results and of course in case you're wondering when you download these files into your automatic 1111 folder you want to load it into the model fallen in there into the Laura folder then when we are inside of automatic 1111 on the right side you have this pink button this opens up the section down here with your textual inversions and so on you have here a tab for Laura so you want to click on this and then search through your Laura files and then down here you can see you have the pulley hedron skinny all Laura this is what we are going to use so next let's have a look here at the setup that I'm using to get better results for my main prompt I'm writing raw photo Rim lighting two-tone light professional photo portrait close up of a beautiful young French woman upper body curly hair summer dress centered Park in background highly detailed skin full sharp with a weight of 1.2 in round brackets and then we have the Laura for pull heater on skinny all with a weight of only 0.3 and then we have the Laura for the lower light with a weight of only 0.2 now here's the reason why I'm using only 0.3 because by doing that in my main prompt up here I'm preparing the image to give me a better skin quality but at the same time I'm don't interfere with the output so that I still can generate the images I want to have in my negative prompt I'm using bad dream digital painting painting 3D so it is a very reduced negative prompt now bad dream is a negative embedding you can find it here I will link all of that below the video now this of course goes into the embeddings folder inside of your automatic 1111 folder now here comes the interesting part when we are going to look at the settings I have here my sample method and I will show you an overview of all the sample methods and how they compare for the sampling steps here I'm setting 40 sampling steps a little bit higher in this case but you can experiment with lower settings too I have restore face turned off I have high-res fix turned on I'm using the 4X Ultra sharp upscaler again I'm going to link this in my video description now of course the upscaler goes into the models folder into the ESR again folder I'm setting my high res steps to 20 but you can set this to zero so it's going to use the same amount of steps that you have used to render the images or you set it to any kind of other value I'm using a denoise strength of 0.3 so we are staying true to the image but at the same time allow the upscaler to make changes to improve the image I'm using an upscale of 2 for the next high-res settings I'm showing you you have to go to settings in automatic 1111 on the left side click on user interface and then scroll down to see these two adjustments take the two boxes and restart the UI to see these special adjustments for the high res fix and then I use the same sampler as I have used for the image you can use a different sampler here if you feel like a different sampler gives you better results now down here I have a field where I can choose to set my prompt again and here I'm writing the same prompt as a buff but what I'm doing here differently is that I'm setting my skin Laura to a weight of one and I'm going to remove the low ra Lora for the lower light now what this does is now in the upscaling this is using the full power of the Laura to create the better skin so let's look at some examples on how this method actually turns out with the different Samplers here we have Euler a and feel free to pause the video to look more detailed on this picture next we have the Euler sampler also a very nice result here is LMS also nice here we have Hoin also good but here we have dpm2 and now I think we get more into nicer details of the skin so here we have DPM 2A and I feel like here you get a little bit more detail so it looks nicer DPM plus plus 2sa also has very nice details and here you can see they're also slightly different all of them here we have DPM plus plus 2m this is one of my favorites here we have DPM plus plus sde also very nice from the details on the skin DPM plus plus 2m sde is my favorite here the only thing I don't like is this peaking highlights you get on the left side in the hair but that might just be the seat when you roll a different seat or have a different scene you might not get these problems here we have DPM fast which is okay but it creates an error in between the eyes here on the nose red is some kind of strange fragment there DPM adaptive is also very nice LMS Keras has some good results there dbm2 Keras is nice but a little bit soft I feel like and also as you can see up here it created some error with the hair again this might be only this kind of seat DPM a Keras also creates a nice result and also nice lighting and a specific image DPM plus plus 2sa Keras again is a very nice result and one of the assemblers I use a lot for images DPM plus plus 2m Keras is also nice but again you can see up here it has the same kind of error that we have seen before with the hair here we have DPM plus plus sde Keras which is very nice again and DPM plus plus 2m sde Keras also something I'm using often especially for upscaling here we have ddim which I am surprised it looks very nice it has a very consistent result also so I really enjoy that and here we have plms which I don't really use much but I kind of like the result and uni PC which is often good for photorealistic results I also like the result here even though the lips look a little bit too digital I feel like now let's go back into automatic 1111 here we are in in painting so I sent the image over to in painting as you can see here I've masked out the eyes specifically I'm reducing the prompt here to Raw photo Rim light two-tone light professional portrait beautiful eyes and then down here in the settings I leave most of the settings I change the in paint area to only mask for the sampler you can choose whatever works best for you or the same sampler that you have used to generate the image and then down here for the size I'm setting 768 by 768 because the area is a little bit bit bigger because we are rendering both eyes at the same time but you can also get away with 512 by 512. for denoise in this case I'm setting 0.5 so that the AI has a chance to actually create new eyes but at the same time have them fit to the face and the perspective and here we have the result from that it looks pretty good even though the eyes could be more realistic so you might want to fiddle a little bit more around with the prompt you're getting here let me know in the comments what you think about this Laura or if you know a better method to get even more realistic skin thanks for watching subscribe to my channel if you want to see more videos like that and see you soon bye
Info
Channel: Olivio Sarikas
Views: 65,515
Rating: undefined out of 5
Keywords: oliviosarikas, olivio sarikas, olivio tutorials, lora, skin lora, photo lora, photorealisitc skin, a111 lora, realisitc lora, realsitic skin, a1111 stable diffusion, a1111 tutorial, a1111 inpainting, a1111 upscale, a1111 lora, lora portrait, ai realistic photo, ai realistic art, ai realistic girl, girl, woman, ai art, image generation, ai generator free, free ai tools, best ai tools, ai news, new lora, lora 1.5
Id: sCOsZFH7xlU
Channel Id: undefined
Length: 10min 39sec (639 seconds)
Published: Mon Jul 17 2023
Related Videos
Note
Please note that this website is currently a work in progress! Lots of interesting data and statistics to come.