Offset Noise for Stable Diffusion: Model, LORA and Embeddings

Key Takeaways

is offset noise the mid-journey killer hello my friends how are you doing today I have something really amazing for you to look at not just my face also my beard but let's talk about the actual news here so offset noise tackles a problem with stable effusion here you see an image that is created with a normal model in that case deliberate two but here we have an image that is trained on offset noise and you can see how this other image has more Darkness more brightness a nicer range in between these values and this is something that you can't cover with a classic stable effusion model because it tries to stay somewhere in the medium values so the scene is always kind of bright now of course light is a very important medium to tell your story bring artistic expression bring drama into the scene so this is actually a huge step forward so let let's look into the solutions here in this video I'm going to show you a model for stable diffusion 2.1 negative embeddings not even sure you knew these existed and then also a Lora model you can use with any 1.5 model which gives you a ton more flexibility so of course here we have the first model Illuminati diffusion versus 1.1 let's have a look at what this model can do so here we have an example by kakoi who is actually the creator of the model and this image uses a very nice light is often used by Studio photographers for the reason that it nicely brings out the features of the subject but also the structure and texture of the surface so it brings a lot of Dimension to the image now here we have a second example by kamenduru and this is another classic method in photography here we have a rim light around the edges of the dock and also the cloud and this is balanced out with the Dark Cloud and the dark fur of the dark very nice expression beautiful realistic sunlight and then we have a fashion Portrait by Hashim from my community thanks for sending that really amazing and again you can see how the dynamic between the bright parts and the Shadows really brings out the qualities of this image this is trained on stable diffusion 2.1 that means inside of automatic 1111 you need to have a minimum resolution of 768 pixels which means you need to have a strong GPU to render this or the render time is going to be longer because the resolution is going to be higher now there's another thing that is really important to look into and this is down here you can see you have to download the yaml file and this goes also in automatic 1111 in the same folder with the model now when you scroll down here you will see that this is best used with these three embeddings these are textual versions and they are negative embeddings now what does it actually mean a negative embedding goes into your negative prompt not into the normal prompt you can see here it says in real fixer and end fixer these are textual inversions that are trained on the things you don't want to have in an image this is why they go into the negative prompt so that then when it's rendered these things can be avoided which is actually a genius solution so here I rendered you examples that show using the three different embeddings now the first one is using just the prompt and you can see that the body of the robot doesn't fit and also the rest of the image is not so great now the end fixer is fixing the coherence the connection between the different parts so in that case the body of the robot turned out a lot better the N art fixer fixes the artistic style in the image to give you a better overall expression and make everything look more beautiful the unreal fixer gives you more realistic light and makes things look more like a movie scene or a photo because they are embeddings you can mix and match them any way you want so down here we have examples of the end fixer with the unreal fixer so this is fixing the shape of the robot and then also the Light then we have end fixer and an art fixer fixing the robot and the artistic style then we have the unreal fixer and the an art fixer fixing the light and the artistic style and then in the last one we have all three together now I would suggest for you to play around and test these embeddings with the model to see you what gives you the best results because this really depends on what you want to render and the Artistic Styles you want to have with that to get these three embeddings you can follow these links and simply download them into your automatic 11 11 4 folder in the models folder and in there in the embeddings folder after you've done that and started automatic 1111 you can simply call them by the name in the negative prompt now here's something else I want to point out about this model I myself and also other members of my community have found automatic 1111 can be a bit buggy and can crash with that model also it is buggy to use with opennet so you're probably going to have some errors so always check your command window to see what is going on here and if nothing else helps just restart automatic 1111 another thing to point out here with Illuminati diffusion is that I could not make it load in invoke AI for some reason of course all of this can be avoided by my second suggestion for today and this is called EP noise offset this is a Laura mod and it is trained on SD 1.5 now the huge benefit here is you can use this with any 1.5 model and you can see from the images here you also get this nice dynamic range between the brighter and the darker areas you get very nice beautiful light and you can also create these dark and dystopian light scenes with very expressive dramatic lights but because this is rendering in 1.5 you can render all of that in a lower resolution and like I said with any model you want to now the way to use this model is very simple download the file into your Laura folder in the models folder in automatic 1111 and then as you can see here with this prompt you simply call this Laura in your positive prompts this also doesn't require the negative embeddings from the other model to work personally I would suggest to you to use both of them but of course the Laura model is easier to use it's a smaller download and it works with all of your models so that might be a huge benefit on the other hand as we have seen the Illuminati model can give really stunning high resolution results also because it is trained on stable diffusion 2.1 let me know what you think in the comments thanks for watching and leave a like if you enjoyed this video 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 leave a like if you haven't yet and well um yeah